ARUDet: active retrieval and uncertainty-aware detection for sports video object detection Authors: Lijing Yu, Jin Gao, Xinqi Ji, Bo Liu Published: 2026 ABSTRACT AbstractVideo object detection in sports scenarios faces severe challenges posed by high speed motion, heavy occlusion, and complex deformation. Existing methods employ passive feature aggregation that applies a uniform fusion strategy across all regions, neglecting the type specific differences of local degradation and failing to match targeted complementary information for different defects. Meanwhile, visual degradation causes object boundaries to exhibit probabilistic distribution characteristics, yet conventional deterministic regression ignores such geometric ambiguity and forces fitting a single coordinate, making localization reliability difficult to guarantee. To address these issues, we propose ARUDet (Active Retrieval Uncertainty-aware Detector), comprising an Active Temporal Retrieval Module (ATRM) and an Uncertainty Rectified Regression Head (URH). Specifically, ATRM first identifies the degradation type of each region in the current frame through a Defect Responsive Assessor and encodes it into an explicit query vector. Subsequently, under dual constraints of semantic consistency and quality complementarity, it actively retrieves the best matching historical feature patches from a temporal memory bank and selectively replaces only low quality regions, achieving on demand restoration. URH collaboratively suppresses geometric uncertainty through a Probabilistic Boundary Projector (PBP) and a Lower Bound Optimizer (LBO): the former explicitly models the geometric distribution to improve localization quality, while the latter constrains the worst case error to tighten the prediction lower bound. Experiments on multiple sports datasets demonstrate that ARUDet achieves significant performance improvements, further validating the effectiveness of the proposed method.Keywords: Video object detection, Sports video analysis, Temporal feature retrieval, Uncertainty-aware regressionSubject terms: Engineering, Mathematics and computing Introduction Sports video analysis is an important research direction in computer vision, with broad application value in intelligent broadcasting, tactical analysis, and athlete training assistance. With the accelerating digital transformation of the sports industry, the demand for automated athlete detection and tracking is growing rapidly, covering team sports such as football, basketball, and volleyball, as well as individual competitive events such as skiing and badminton. However, sports scenarios possess unique visual characteristics: athletes exhibit high speed and large amplitude motion patterns, frequent physical contact and mutual occlusion, and drastic pose deformation. Taking badminton as an example, racket swing speeds can reach hundreds of kilometers per hour, causing severe motion blur; in skiing scenarios, the contrast between athletes and snowy backgrounds is low while pose changes are drastic; in football and basketball games, dense player gatherings cause extensive occlusion. These characteristics impose extremely high demands on the robustness of detection algorithms, making traditional static image detection methods difficult to apply directly. Current video object detection methods mainly face two core challenges, as illustrated in Fig. 1. First, passive feature aggregation leads to restoration failure. Existing temporal methods typically adopt fixed window feature aggregation strategies: extracting features from K adjacent frames and then fusing them through attention mechanisms or convolution operations. This indiscriminate aggregation approach has a fundamental flaw: different regions of the current frame suffer from different types of degradation (occlusion, blur, illumination changes, etc.), and they require complementary information from different sources. For example, an occluded player needs to retrieve complete appearance from unoccluded historical frames, while motion blurred regions need texture details from clear moments. However, adjacent frames within a fixed window may not contain the required high quality information, and a unified fusion strategy cannot match the best restoration source for each degraded region. As a result, the model passively accepts a large amount of ineffective or even harmful features, which instead dilutes the contribution of useful information. Second, deterministic regression ignores geometric uncertainty. Under high speed motion and severe occlusion scenarios, the physical boundaries of objects appear as ambiguous probability distributions due to residual images and deformation, yet traditional bounding box regression methods universally assume deterministic point estimation, forcing the fitting of blurry distributions into single values. The problem with this approach is that the model outputs a bounding box in an “average” sense, but on difficult samples, this average may deviate far from the true boundary, and the model cannot quantify the reliability of its own predictions. When the visibility of objects fluctuates drastically across consecutive frames, deterministic predictions jump between different “average positions,” producing temporal localization jitter; in extremely degraded frames, the model may even output incorrect boundaries with high confidence, leading to tracking drift. Motivation of ARUDet. Top left: Existing methods passively aggregate features from neighboring frames within a fixed window, regardless of region specific degradation. Top right: Our ATRM actively retrieves complementary features from a temporal memory bank, selectively restoring only degraded regions (e.g., obstructed players) while preserving unobstructed ones. Bottom left: Conventional certainty regression produces unstable bounding boxes that jitter across consecutive frames. Bottom right: Our uncertainty correction yields temporally consistent and precise localization. To address these challenges, we propose ARUDet (Active Retrieval Uncertainty aware Detector), a robust detection framework that integrates active retrieval with worst case optimization strategies. ARUDet comprises two core modules: the Active Temporal Retrieval Module (ATRM) and the Uncertainty Rectified Regression Head (URH). For the first challenge, ATRM abandons the passive aggregation paradigm and instead adopts a pull based active retrieval strategy. The module first evaluates the degradation degree and type at each spatial location of the current frame through a Defect Responsive Assessor, and encodes the features of degraded regions into explicit query vectors. Subsequently, under dual constraints of semantic consistency and quality complementarity, it actively retrieves the best matching historical feature patches from a dynamically maintained temporal memory bank, and selectively replaces only low quality regions, achieving on demand restoration. This mechanism breaks the limitation of fixed temporal windows, ensuring that each degraded region can obtain targeted complementary information across longer time spans, rather than passively accepting potentially irrelevant neighboring frame features. For the second challenge, URH collaboratively suppresses geometric uncertainty from two aspects: the Probabilistic Boundary Projector (PBP) extends boundary coordinate prediction from single values to probability distributions, explicitly characterizing localization uncertainty; the Lower Bound Optimizer (LBO) estimates the confidence lower bound of the prediction distribution and minimizes the localization error at the lower bound, indirectly constraining the uncertainty range of predictions. Through this dual mechanism, the model further reduces the worst case localization error while maintaining baseline localization accuracy, thereby adaptively suppressing geometric uncertainty. Experiments on multiple sports video datasets demonstrate that ARUDet achieves significant improvements in both detection accuracy and localization robustness, with particularly notable advantages in motion blur and occlusion scenarios. Beyond sports VOD, the defect-aware retrieval and worst-case probabilistic regression in ARUDet align with the broader paradigm of uncertainty-aware, reliability-driven perception, with conceptual parallels in robust multimodal recognition1, uncertainty quantification for reliable decision making2, and look-ahead Bayesian design3. The main contributions of this paper are as follows: We propose the Active Temporal Retrieval Module (ATRM), whose primary novelty is twofold: a defect-aware query generation scheme that encodes the local degradation type into the query vector itself (driving retrieval by what is missing rather than what looks similar), and a dual-constraint retrieval rule that jointly enforces semantic consistency and quality complementarity with selective region-level replacement. We propose the Active Temporal Retrieval Module (ATRM), whose primary novelty is twofold: a defect-aware query generation scheme that encodes the local degradation type into the query vector itself (driving retrieval by what is missing rather than what looks similar), and a dual-constraint retrieval rule that jointly enforces semantic consistency and quality complementarity with selective region-level replacement. We propose the Uncertainty Rectified Regression Head (URH), which directly improves localization accuracy and indirectly constrains the uncertainty range of predictions through probabilistic boundary modeling and confidence lower bound optimization. We propose the Uncertainty Rectified Regression Head (URH), which directly improves localization accuracy and indirectly constrains the uncertainty range of predictions through probabilistic boundary modeling and confidence lower bound optimization. We validate the effectiveness and generalization capability of ARUDet on multiple sports video datasets including football, basketball, volleyball, skiing, and badminton. We validate the effectiveness and generalization capability of ARUDet on multiple sports video datasets including football, basketball, volleyball, skiing, and badminton. Related work Object detection The development of modern object detection has generally undergone three paradigm shifts: from convolutional detectors to end to end Transformers, and then to generative detection. Early two stage methods, represented by Faster R-CNN, established high accuracy detection baselines through joint optimization of region proposal generation and classification regression4; subsequently, one stage methods (such as RetinaNet) effectively alleviated class imbalance through Focal Loss, pushing the upper bound of real time detection performance5. Building upon this, DETR reformulated detection as a set prediction problem, replacing NMS with Hungarian matching to achieve a unified end to end training paradigm6. Follow up works including Deformable DETR, Conditional DETR, and DAB-DETR addressed slow convergence and insufficient small object performance from the perspectives of sparse attention, conditional queries, and dynamic anchor queries, respectively7–9. DINO further significantly improved accuracy and stability through denoising training and query initialization strategies10, while RT-DETR validated the engineering feasibility of the DETR pipeline in real time scenarios11. Meanwhile, diffusion models began entering the detection task. DDPM provided a unified probabilistic framework for step by step denoising modeling12, and DiffusionDet explicitly modeled detection as a reverse diffusion process from noisy boxes to target boxes, demonstrating the potential of diffusion based detection in both performance and inference flexibility13. Furthermore, large scale vision language pre-training methods such as GLIP demonstrated open vocabulary and cross domain generalization capabilities, providing stronger priors for detection models14. Overall, detection research has shifted from “structural engineering” to “unified modeling and large model prior driven,” laying the methodological foundation for temporal modeling and robust localization in video scenarios. Temporal feature aggregation in video object detection Building upon the above developments, the core problem of video object detection has transformed into cross frame information utilization. Early methods primarily focused on feature propagation and neighboring frame alignment: Deep Feature Flow reduced redundant computation through keyframe feature propagation15, FGFA improved moving object detection through optical flow guided aggregation16, and STSN enhanced alignment capability using deformable spatiotemporal sampling17. Subsequently, research shifted from local alignment to long range relation modeling: SELSA performed aggregation in sequence level semantic space18, RDN strengthened cross frame discriminative relations through relation distillation19, and MEGA combined global and local memory to alleviate information insufficiency in short windows20. Entering the Transformer era, TransVOD incorporated spatiotemporal interaction into a unified framework, further enhancing cross frame modeling expressiveness21. In recent work, MaskVD and Beyond Boxes advanced the balance between efficiency and accuracy in video detection from the perspectives of region mask efficient computation and mask guided spatiotemporal fusion, respectively22,23. In the sports domain, SoccerNet-Tracking and SportsMOT provided more challenging multi-object benchmarks24,25, while SportMamba and FieldMOT emphasized the importance of long range modeling efficiency, field priors, and cross camera consistency26,27. Although the aforementioned methods have significantly improved temporal utilization capabilities, most still adopt the passive paradigm of “unified fusion rules,” reading historical features within predefined windows and executing similar aggregation strategies. This design is effective in general scenarios, but in heterogeneous degradation regions of sports videos (where occlusion, blur, and low contrast coexist), it is prone to information mismatch and noise amplification. Therefore, the more critical question at the current stage has shifted from “whether to aggregate” to “how to perform controllable aggregation according to degradation type.” At the same time, the stability of localization regression on difficult samples has also become another determining factor for the performance upper bound. Distinction from Prior Memory-based VOD and Tracking. Unlike MEGA, SELSA, and Beyond Boxes that apply uniform aggregation at every location, our ATRM uses defect-aware queries under joint semantic and quality constraints and replaces only degraded regions, avoiding the dilution of high-quality features. Recent tracking studies on motion-aware temporal modeling28,29 and shared-query design30,31 share the same selective-temporal-modeling spirit. Bounding box regression and uncertainty modeling Bounding box prediction research has gradually evolved from “deterministic point estimation” to “quality estimation and distribution modeling.” IoU-Net pointed out that localization quality and classification confidence are not equivalent, requiring explicit learning of localization reliability32; GIoU improved traditional regression loss through geometrically consistent objectives33. In the uncertainty direction, Bounding Box Regression with Uncertainty introduced variance into regression learning, advancing probabilistic localization representation34; Probabilistic Object Detection (PDQ) further systematized the definition and evaluation protocol for uncertainty aware detection35. Subsequently, GFL and GFLv2 alleviated the inconsistency between classification scores and localization quality through distributed boundary representation and joint quality estimation learning36,37. Recent research has paid more attention to the interpretability and training utility of uncertainty, such as aleatoric uncertainty analysis based on foundation models38, DiL’s explainable evaluation of abnormal uncertainty39, and UGS’s direct incorporation of uncertainty aware mechanisms into gradient stabilization optimization40. DualAdaptNet uses an auxiliary regression head to estimate the worst case and minimize its error, embodying min max style lower bound optimization41; DIG-FACE improves generalization on unknown distributions through bias upper bound constraints42. However, existing methods are mostly based on the image domain or static distribution assumptions, and still lack targeted risk control under severe degradation and visibility abrupt change conditions: on one hand, merely optimizing expected error is insufficient to suppress tail localization distortion; on the other hand, there is a lack of unified design that jointly constrains probabilistic boundaries, temporal consistency, and worst case error. Based on this, we adopt “probabilistic boundary modeling + confidence lower bound optimization” to enhance localization stability on difficult samples. Method Overall framework overview Figure 2 illustrates the overall framework of ARUDet and the detailed structure of its core module ATRM. ARUDet adopts RT-DETR as its base detection framework. Given the t-th frame image , a backbone network first extracts multi-scale features , which are then processed by a Hybrid Encoder for intra-scale enhancement and cross-scale fusion, yielding the fused feature . On this basis, we introduce two core modules to strengthen the detection pipeline. Overall framework overview of ARUDet. Top: the detection pipeline from input frames through backbone, Hybrid Encoder, ATRM, Transformer decoder, to URH, which outputs classification, bounding box means, and uncertainty estimates. Bottom: detailed view of ATRM, consisting of the Defect Responsive Assessor (DRA), dual constraint retrieval with semantic consistency and quality complementarity, and selective replacement fusion. Active Temporal Retrieval Module (ATRM) is positioned between the Hybrid Encoder and the Transformer decoder. This module first evaluates the degradation degree and degradation type of at each spatial location through a Defect Responsive Assessor (DRA), generating explicit query vectors for each low-quality position. It then retrieves the best matching historical feature patches from a dynamically maintained temporal memory bank under dual constraints of semantic consistency and quality complementarity, and selectively replaces only degraded regions through a gating mechanism, outputting the repaired feature . The repaired features are fed into the Transformer decoder, where object queries perform matching on high-quality features via cross-attention, producing instance embeddings for each query. Uncertainty Rectified Regression Head (URH) replaces the regression FFN at the decoder output and operates on instance embeddings . URH comprises two synergistic components, the Probabilistic Boundary Projector (PBP) and the Lower Bound Optimizer (LBO): PBP extends each boundary coordinate prediction from a single value to a Gaussian distribution , explicitly modeling localization uncertainty; LBO further estimates the interval boundaries of this distribution at a given confidence level and minimizes the worst case localization error, thereby tightening the uncertainty range. The overall forward process can be summarized as: where is the classification prediction, denotes the expected bounding box coordinates, and represents the uncertainty estimate for each boundary. The specific design of each component is detailed below. Backbone network and hybrid encoder Backbone network. We use ResNet-50 as the backbone network to extract multi-scale features: where correspond to the outputs of stages 3, 4, and 5, respectively. Hybrid Encoder. The efficient Hybrid Encoder of RT-DETR processes multi-scale features in two steps: Attention-based Intra-scale Feature Interaction (AIFI) enhances high-level semantic features through self-attention, and CNN-based Cross-scale Feature-fusion Module (CCFM) achieves cross-layer information interaction through multi-path fusion, outputting the fused multi-scale feature sequence: However, in sports videos, certain spatial locations of may be degraded due to occlusion, motion blur, and other factors. ATRM (Section 3.3) operates at this stage, repairing into before feeding it into the decoder. Active Temporal Retrieval Module (ATRM) As discussed above, the fused feature output by the Hybrid Encoder often suffers from local degradation in sports scenarios due to occlusion, blur, and low contrast. The core idea of ATRM is that different degraded regions require different complementary information; therefore, retrieval should be performed actively on a per-region and per-defect-type basis, rather than passively applying uniform fusion across all positions. ATRM consists of three sub-modules: the Defect Responsive Assessor, the dual constraint retrieval mechanism, and selective replacement fusion. Defect Responsive Assessor (DRA) Intuition: DRA decides where in the current feature map needs repair and what kind of defect is present. The goal of the Defect Responsive Assessor is to produce two outputs for each spatial position of the current frame feature map : a degradation degree score and a degradation type embedding . Degradation degree assessment. We design a lightweight quality assessment sub-network. Two layers of 1 × 1 convolutions followed by Sigmoid activation are applied to the feature map to obtain a per-pixel degradation score map: where denotes the degradation degree at position (i, j). A larger indicates more severe degradation at that position, implying a greater need for supplementary information from historical frames. Degradation type encoding. Different types of degradation (occlusion, motion blur, low contrast) require different types of complementary features. We define K learnable degradation prototype vectors , each implicitly corresponding to a different degradation pattern. For each spatial position, the degradation type embedding is computed through projection and Softmax attention: where is the projection matrix. The degradation type embedding encodes the degradation characteristics of that position and serves as the query vector in the subsequent retrieval stage, guiding the matching of semantically consistent and quality complementary historical features from the memory bank. Query vector generation. Finally, queries are generated only for positions whose degradation degree exceeds the threshold δ. Let the degraded region set be , and the corresponding query set is: Each query triplet contains the degradation type embedding (for matching), the original feature (for semantic consistency verification), and the degradation score (for controlling replacement intensity). With the query set , the next step is to retrieve matching high-quality features from historical frames. Temporal memory bank and dual constraint retrieval Intuition: among all historical patches, we look for one that both refers to the same semantic region and is itself of high quality. Temporal memory bank. The system maintains a first-in-first-out (FIFO) temporal memory bank with a capacity of frames. For each historical frame τ, the memory bank stores its feature map and the corresponding degradation score map . To control storage overhead, the memory bank samples at fixed intervals (storing one frame every s frames) and evicts the oldest frame when full. Additionally, a global quality score is computed for each frame before insertion; only frames with exceeding the minimum quality threshold are admitted, preventing low-quality frames from polluting the memory bank. Dual constraint retrieval. Given a query from the query set , the retrieval objective is to find historical features from that are semantically consistent and quality complementary. We define two retrieval constraints: Semantic consistency constraint. The retrieved historical features should describe the same object or the same semantic region as the query position. For position (i', j') of frame τ in the memory bank, the semantic similarity is computed as: This score ensures that the retrieved features are aligned with the query position in semantic space, avoiding the introduction of features from irrelevant objects. Quality complementarity constraint. The retrieved historical features should exhibit high quality (low degradation degree) at the corresponding position. The quality complementarity score is defined as: This score is the product of two terms: measures the quality of the candidate position itself, and represents the repair demand intensity of the current position. The complementarity score is maximized when the candidate position has high quality and the current position is severely degraded. Joint retrieval score. The two constraints are linearly combined into the final retrieval score: where is the balance coefficient. For each query position (i, j), the Top-1 match with the highest retrieval score is selected across all spatial positions of all frames in the memory bank: The corresponding retrieved feature is . Efficient retrieval strategy. To avoid the high computational cost of pixel-wise full-bank search, we adopt a two-level acceleration: (1) coarse-grained frame selection, which first selects the Top- candidate frames based on global semantic similarity; (2) fine-grained position matching, which computes per-position retrieval scores only within the candidate frames. This reduces the search complexity from to , where . After retrieving the matched features, the next step is to fuse them into the current frame to complete feature repair. Selective replacement fusion Intuition: we replace only the degraded positions, so good features in the current frame are never diluted by historical noise. After retrieval, the historical features need to be fused into the current frame. Unlike global fusion, ATRM performs replacement only on degraded regions, preserving the original features of high-quality regions without interference. Degradation aware gated fusion. For each position (i, j) in the degraded region set , the retrieved feature has been obtained. Direct replacement may introduce semantic discontinuity; therefore, we adopt soft gated fusion based on degradation degree: where the gating coefficient is obtained from the degradation score through a learnable transformation: are learnable scalar parameters, and σ is the Sigmoid function. The more severely degraded a position is, the larger the gating coefficient, and the stronger the contribution of the historical feature. Adaptation transform. The retrieved historical features may have domain discrepancies with the current frame (e.g., illumination changes, viewpoint differences), and direct fusion may produce artifacts. Therefore, a lightweight adaptation transform is applied to the retrieved features before fusion: where and are learnable parameters. LayerNorm further eliminates magnitude discrepancies. Non-degraded region preservation. For non-degraded regions , the original features are directly preserved: This ensures that high-quality regions are free from any interference, avoiding the problem of “good features being diluted” in conventional global fusion. ATRM loss. To supervise the degradation degree prediction of the Defect Responsive Assessor, we introduce an auxiliary loss. We use the IoU between the current frame detection results and ground truth annotations as pseudo labels to supervise the degradation score map: where Ω is the foreground region, is the IoU quality score based on the detection box to which position (i, j) belongs, and BCE denotes binary cross entropy loss. This loss guides the assessor to mark low-quality regions as highly degraded, thereby driving subsequent retrieval and replacement. ATRM forward pass for frame t Transformer decoder The repaired feature is fed into the Transformer decoder. The decoder receives a set of object queries initialized by an IoU-aware query selection mechanism, iteratively updates them through multiple layers of self-attention and cross-attention, and outputs instance embeddings: Note that the cross-attention of the decoder operates on the ATRM-repaired feature rather than the original , allowing query matching to benefit from high-quality features. The final label assignment employs Hungarian matching, requiring no NMS post-processing. Although feature-level degradation has been mitigated, bounding box localization still faces geometric uncertainty issues, as high-speed motion and residual occlusion render the object boundaries inherently ambiguous. To address this challenge, we introduce URH at the decoder output. Uncertainty Rectified Regression Head (URH) URH replaces the standard regression FFN at the RT-DETR decoder output, addressing the boundary geometric uncertainty caused by high-speed motion and severe occlusion in sports scenarios. As shown in Fig. 3, it comprises two synergistic components: the Probabilistic Boundary Projector (PBP, yellow path) extends boundary prediction from point estimation to distribution estimation, and the Lower Bound Optimizer (LBO, purple path) constrains the worst-case error of the prediction distribution. Detailed structure of the Uncertainty Rectified Regression Head (URH). Instance embeddings are processed through three parallel FC branches. PBP (yellow path): the upper branch predicts bounding box means , the middle branch predicts log-variance ; after regularization, both jointly model the uncertainty distribution and output probabilistic bounding boxes, supervised by and . LBO (purple path): the lower branch predicts the adaptive confidence level α, which is combined with the mean and variance to estimate confidence interval boundaries; constrains the worst-case localization error and regularizes the confidence level, outputting refined probabilistic bounding boxes. The right-side example demonstrates the visualization of boundary uncertainty produced by URH in actual detection. Assumptions and Theoretical Positioning. The per-coordinate Gaussian rests on three explicit assumptions: (A1) Hungarian one-to-one matching in RT-DETR makes the boundary deviation approximately unimodal, justifying a single Gaussian rather than a mixture; (A2) the four boundary coordinates are conditionally independent given the instance embedding , a standard simplification adopted in probabilistic regression literature34,35; (A3) the symmetric Gaussian is intentionally paired with an asymmetric worst-case loss (the \max operator in , Section 3.5) so that biased predictions are still corrected directionally. The Gaussian is therefore not used to perfectly fit every form of geometric ambiguity; it is used because it admits a closed-form confidence interval that LBO can directly minimize. Unlike KL-Loss34 which lacks a worst-case constraint and GFL/GFLv236,37 which use discrete coordinate grids with quantization bias, URH unifies a continuous probabilistic boundary with a confidence-interval worst-case optimizer in one regression head. Probabilistic Boundary Projector (PBP) The standard regression FFN of RT-DETR predicts bounding boxes as four deterministic coordinates (x, y, w, h), implicitly assuming a unique correct position for each boundary. However, under motion blur and occlusion scenarios, object boundaries possess inherent ambiguity. PBP extends each boundary prediction from a single value to a Gaussian distribution, explicitly characterizing localization uncertainty. Distribution parameterization. Given the instance embedding output by the decoder, PBP predicts the mean and variance through two parallel fully-connected branches: where is the expected prediction of bounding box coordinates, and is the corresponding variance. Taking the logarithm of the variance ensures numerical stability, and the actual variance is recovered via . The prediction distribution for each boundary is modeled as an independent Gaussian: Probabilistic regression loss. Negative log-likelihood (NLL) is adopted as the regression loss for PBP, replacing standard L1 regression: where is the corresponding coordinate of the ground-truth bounding box. The advantage of this loss lies in the following: when the prediction error is large, the model can reduce the penalty of the first term by increasing , but the second term will simultaneously increase, forming a natural regularization that prevents the model from indefinitely increasing the variance to “escape” the penalty. This encourages the model to output small variance (high confidence) for easy-to-localize samples and large variance (low confidence) for difficult samples, achieving adaptive uncertainty estimation. Variance regularization. To prevent the variance from collapsing to zero (degenerating into deterministic regression) or exploding to infinity (completely abandoning prediction), a variance range constraint is introduced: where and are the preset lower and upper bounds of the variance. At this point, PBP has achieved the extension from point estimation to distribution estimation, but optimizing only the expected error is insufficient to control localization deviation in extreme cases. LBO further provides constraints from the worst-case perspective. Lower Bound Optimizer (LBO) PBP provides probabilistic boundary predictions, but optimizing only the expected error (NLL loss) may still lead to tail localization distortion: in low-probability regions of the distribution, predictions may deviate far from the true boundary, while the model cannot perceive this worst-case risk. LBO suppresses worst-case errors by explicitly constraining the confidence lower bound of the prediction distribution. Confidence lower bound estimation. For the prediction distribution of the k-th boundary, at a given confidence level α (e.g., α = 0.9), the predicted value falls within the interval with probability α, where is the quantile of the standard normal distribution. We define the confidence lower and upper bounds as: It should be noted that although the confidence interval of the Gaussian distribution is symmetric about in form, this does not mean that the model penalizes both sides symmetrically. In the subsequent LBO loss, the operation always selects the side with the greater distance from the ground truth to the confidence interval boundaries for optimization. When is biased relative to , the distances from the ground truth to the two ends are naturally unequal, and the max operation automatically focuses on the direction with larger error, achieving asymmetric correction at the gradient level. In other words, the symmetric interval form and the asymmetric optimization behavior are not contradictory: the former ensures modeling simplicity, while the latter achieves directional constraints on the worst-case direction through the loss function design. Lower bound localization loss. The core idea of LBO is to minimize the localization error at the confidence interval boundaries, thereby indirectly constraining the uncertainty range of predictions. The lower bound loss is defined as: This loss takes the maximum distance from the ground truth to both ends of the confidence interval, measuring the worst-case localization error. Minimizing this loss has two effects: 1. Tightening the confidence interval width: When is too large, and are far from , and one end must be far from the ground truth, producing a large loss. This forces the model not to simply increase the variance to reduce the NLL loss. 2. Correcting mean bias: When deviates from , the distances to both ends increase. This loss cooperates with the NLL loss to optimize both expected accuracy and tail risk. Adaptive confidence level. To adapt to samples of different difficulty, we further extend the confidence level α from a fixed constant to a sample-adaptive value. It is predicted from the instance embedding through a lightweight branch: where σ is the Sigmoid function, constraining within (0, 1). For difficult samples (high degradation, heavy occlusion), the model tends to predict a lower , corresponding to a wider confidence interval that provides greater tolerance; for easy samples, it predicts a higher , demanding more precise localization. To prevent from being too low (losing its constraining effect), a lower bound regularization is applied: where is the minimum confidence level threshold (set to 0.7 in experiments). Combining all the modules above, ATRM completes degradation repair at the feature level, while URH achieves probabilistic localization and risk constraint at the instance level. The two modules synergistically enhance detection robustness from both ends of input quality and output reliability. The overall training objective is presented below. Overall loss function The overall training loss of ARUDet is composed of the base detection loss and the dedicated losses of each module: where is the base detection loss of RT-DETR, comprising classification Focal Loss and GIoU loss, with label assignment performed through Hungarian matching. is the degradation prediction supervision loss (Section 3.3), guiding the assessor to accurately identify degraded regions. is the probabilistic regression NLL loss (Section 3.5), replacing the original L1 regression term to achieve distributional boundary prediction. is the worst-case localization loss (Section 3.5), constraining the localization error at the confidence interval boundaries. and are the variance range regularization loss and confidence level lower bound regularization loss, respectively, preventing the prediction distribution from degenerating. The balance coefficients for each loss term are set to , , , , . Training strategy. Training is divided into two stages: (1) the warm-up stage (first epochs), during which ATRM is frozen and only the Backbone, Hybrid Encoder, Decoder, and URH are trained; (2) the joint training stage, during which ATRM is unfrozen and all modules are optimized end-to-end. The warm-up stage provides reliable detection quality pseudo-labels for ATRM’s degradation assessment. Experiments Datasets We evaluate ARUDet on three datasets: SportsMOT, Badminton-Players-Detection, and SkiTB, covering team sports scenarios, badminton close-range scenarios, and skiing long-sequence scenarios, respectively. To unify the task definition, we convert the tracking annotations of SportsMOT and SkiTB into detection annotations: per-frame bounding boxes are retained while track IDs are removed, and the data is used solely for object detection training and evaluation. SportsMOT. This dataset contains three types of scenes: basketball, volleyball, and football, comprising 240 video clips, 150,379 frames, 1,629,490 annotated bounding boxes, and 3,401 tracks. The data is split into train/val/test = 45/45/150. The video resolution is 1280×720 at 25 FPS. We train and validate on its official train/val split and report results under the detection setting. Badminton-Players-Detection. This badminton dataset is collected and annotated by us, containing 1,022 images with two categories: player and shuttlecock, split into train/valid/test = 819/139/64. Based on local annotation file statistics, the dataset contains a total of 3,197 annotated bounding boxes, of which 2,370 are player and 827 are shuttlecock; all images are resized to 1280×1280. The main experiments focus on player detection, with shuttlecock results reported as supplementary. SkiTB. This dataset contains 300 multi-camera videos and 2,019 single-camera videos, totaling 352,978 frames, covering 196 athletes and 161 venues, spanning three major categories of alpine skiing, ski jumping, and freestyle skiing (11 sub-events in total). We adopt its date-based split and use per-frame athlete bounding boxes for single-class detection training and evaluation. Implementation details ARUDet is implemented based on RT-DETR: ATRM is inserted between the Hybrid Encoder and the Decoder, and URH replaces the regression branch of the decoder. Training employs one-to-one label assignment via Hungarian matching, and inference maintains an NMS-free pipeline. For video data (SportsMOT, SkiTB), input samples are organized in temporal order and the FIFO memory bank of ATRM is maintained online; for the badminton data, adjacent frames are used to construct short temporal context. ATRM retrieval is driven by the joint scoring of semantic consistency and quality complementarity, and selective replacement is applied only to degraded regions. Optimization follows a two-stage strategy: the first stage freezes ATRM and trains only the Backbone, Hybrid Encoder, Decoder, and URH; the second stage unfreezes ATRM and performs end-to-end joint training of all modules. The total loss comprises the detection loss, ATRM loss, PBP loss, LBO loss, and variance/confidence level regularization terms, with weights set to , and the minimum confidence level of URH set to . Evaluation adopts COCO-style detection metrics (mAP, AP50, AP75); to ensure cross-dataset comparability, all three sets of experiments are unified under the main task of “athlete detection”, where the track IDs of SportsMOT and SkiTB are not used during training or evaluation, and only per-frame bounding box annotations are utilized. Main experimental results and analysis To comprehensively evaluate the effectiveness of ARUDet, we conduct systematic comparisons with two categories of representative baselines across three datasets. The first category consists of image-level detectors, including Faster R-CNN, DETR, Deformable DETR, DINO, and RT-DETR (i.e., the base framework of ARUDet), to measure the gain from temporal information. The second category consists of video object detection methods, including FGFA, SELSA, RDN, MEGA, TransVOD, MaskVD, and Beyond Boxes, covering mainstream paradigms from optical flow alignment to Transformer aggregation. All methods uniformly adopt the ResNet-50 backbone for fair comparison. Image-level methods are independently trained on each dataset and perform per-frame inference; video methods use temporal context according to their original settings. For the Badminton image dataset, video methods are adapted by constructing short temporal context from adjacent frames (marked as †). Results on SportsMOT dataset Table 1 reports the overall detection performance of each method on the SportsMOT dataset. ARUDet achieves the best results across all metrics, with an mAP of 62.1%, surpassing the base framework RT-DETR by 4.6 percentage points and the best video baseline Beyond Boxes by 2.9 percentage points. Table 1Main experimental results on the SportsMOT dataset. All methods use the ResNet-50 backbone. Bold indicates the best result.MethodTypemAPAP50AP75Faster R-CNNImage51.677.855.9DETRImage49.775.353.4Deformable DETRImage54.380.558.7DINOImage56.882.961.4RT-DETRImage57.583.462.3FGFAVideo53.979.758.1SELSAVideo55.281.059.6RDNVideo55.881.660.3MEGAVideo56.482.161.0TransVODVideo58.183.963.2MaskVDVideo58.884.563.9Beyond BoxesVideo59.284.864.5ARUDet (Ours)Video62.187.267.8 Main experimental results on the SportsMOT dataset. All methods use the ResNet-50 backbone. Bold indicates the best result. Three key observations can be drawn from the table. First, the improvement in AP75 is significantly larger than that in AP50. Compared to RT-DETR, ARUDet improves AP75 by 5.5 percentage points, while AP50 improves by only 3.8 percentage points. This asymmetric gain reveals the core mechanism of URH: at loose IoU thresholds, most methods can already correctly localize the target with minor performance differences; at strict thresholds, the localization jitter of deterministic regression methods is amplified, whereas the probabilistic modeling and lower bound optimization of URH effectively tighten the confidence interval of boundary predictions, enabling high-precision localization. Second, some passive video methods perform worse than strong image detectors. The mAP of FGFA (53.9%), SELSA (55.2%), and RDN (55.8%) are all lower than those of DINO (56.8%) and RT-DETR (57.5%). This seemingly counterintuitive phenomenon actually corroborates the core observation of this paper: under the heterogeneous degradation conditions of sports scenarios, neighboring frames within a fixed window do not necessarily contain the required high-quality complementary information, and indiscriminate uniform fusion instead introduces noisy features that dilute the effective information in the original frame. Third, Transformer-based video methods show progress but with diminishing gains. TransVOD (58.1%), MaskVD (58.8%), and Beyond Boxes (59.2%) progressively narrow the gap with image detectors through stronger spatiotemporal modeling, yet remain limited by the passive aggregation paradigm; during prolonged occlusion, consecutive frames within the fixed window may all contain occlusion and thus fail to provide effective repair sources. The ATRM of ARUDet breaks this limitation by actively retrieving clean features from the memory bank across longer temporal spans. To further analyze the performance differences across different sports scenarios, Table 2 presents the per-category mAP results for the three subcategories of SportsMOT. Table 2Per-sport mAP comparison on SportsMOT.MethodBasketballVolleyballFootballRT-DETR54.860.257.5TransVOD55.661.057.8Beyond Boxes56.462.358.9ARUDet (Ours)59.964.561.9 Per-sport mAP comparison on SportsMOT. The improvement is most pronounced in basketball (+3.5 over Beyond Boxes), which is directly related to the dense player confrontations and frequent body contact in basketball games: ten players interact at high frequency within the small half-court area, resulting in high occlusion rates with diverse patterns, where the defect-aware retrieval of ATRM has the greatest room to operate. Football shows the second largest improvement (+3.0), benefiting from two aspects: first, athletes appear as small targets under distant viewpoints, and motion blur causes more severe boundary ambiguity, where URH’s uncertainty rectification is particularly effective; second, the large football field produces alternating sparse and clustered player distributions, allowing ATRM to match differentiated repair strategies for different regions. The improvement in volleyball is relatively modest (+2.2); we attribute this to the relatively regular occlusion pattern caused by the net in volleyball (players are horizontally truncated by the net), resulting in relatively homogeneous degradation types that even passive fusion methods can partially handle. Results on badminton-players-detection dataset Table 3 reports the detection performance on the Badminton dataset. Since this dataset is in image form, video methods are adapted by constructing short temporal context from adjacent frames. The main experiment focuses on the player category, with shuttlecock results reported as a supplement. Table 3Detection results on the Badminton-Players-Detection dataset. † indicates adaptation through constructing pseudo-temporal context from adjacent frames.MethodTypeplayer mAPplayer AP50player AP75shuttlecock mAPFaster R-CNNImage72.493.578.628.3DETRImage70.191.876.226.1Deformable DETRImage74.894.780.930.7DINOImage76.595.382.832.1RT-DETRImage77.295.683.532.8FGFA†Video74.194.280.329.5SELSA†Video75.695.081.831.2TransVOD†Video77.895.984.133.5ARUDet (Ours)Video80.596.887.437.6 Detection results on the Badminton-Players-Detection dataset. † indicates adaptation through constructing pseudo-temporal context from adjacent frames. For player detection, ARUDet achieves an mAP of 80.5%, surpassing RT-DETR by 3.3 percentage points. The improvement in AP75 (+3.9) is significantly larger than that in AP50 (+1.2), consistent with the trend observed on SportsMOT, further validating the contribution of URH to fine-grained localization accuracy. Notably, the improvement in shuttlecock detection is even more significant (+4.8 mAP over RT-DETR), which has a clear physical explanation: the shuttlecock travels at extremely high speeds (exceeding 400 km/h at the moment of impact), appearing in images as severe motion blur and deformation. Its physical boundary is inherently a probability distribution rather than a deterministic contour. Conventional deterministic regression that forces fitting a single coordinate incurs significant errors in such extreme scenarios, whereas the Probabilistic Boundary Projector of URH explicitly models this inherent ambiguity, and the Lower Bound Optimizer further constrains the worst-case error. Their synergy enables the model to maintain reliable localization even for highly uncertain targets. Another noteworthy phenomenon is that some passive video methods also underperform the base image detector on this dataset. FGFA† (player mAP 74.1%) is not only lower than RT-DETR (77.2%) but even lower than DINO (76.5%), indicating that under limited temporal context (pseudo-temporal sequences constructed from image data), optical flow-guided feature propagation may introduce alignment errors that actually degrade the original feature quality. In contrast, the ATRM of ARUDet uses the Defect Responsive Assessor to determine which regions genuinely require repair and leaves high-quality regions untouched, thereby avoiding the risk of “over-repair.” Results on SkiTB dataset Table 4 reports the detection performance of each method on the SkiTB dataset. SkiTB is the most challenging among the three datasets, characterized by low contrast between athletes and the snowy background, dramatic pose variations (especially in ski jumping and freestyle skiing), and longer video sequences. Table 4Main experimental results on the SkiTB dataset.MethodTypemAPAP50AP75Faster R-CNNImage44.771.247.9DETRImage42.869.045.6Deformable DETRImage47.574.651.2DINOImage49.876.953.8RT-DETRImage50.677.554.7FGFAVideo46.873.350.4SELSAVideo48.275.152.0RDNVideo48.975.852.7MEGAVideo49.776.553.5TransVODVideo51.378.155.4MaskVDVideo51.978.756.1Beyond BoxesVideo52.379.156.6ARUDet (Ours)Video55.482.060.3 Main experimental results on the SkiTB dataset. On this dataset, the advantage of ARUDet is particularly prominent: the mAP reaches 55.4%, surpassing RT-DETR by 4.8 percentage points and Beyond Boxes by 3.1 percentage points. The improvement in AP75 reaches 5.6 percentage points (relative to RT-DETR), the largest among the three datasets. We analyze the reasons from the following three perspectives. Long sequences fully activate the memory bank advantage. The video sequences in SkiTB average approximately 175 frames per segment, substantially longer than typical segments in SportsMOT. This allows the FIFO memory bank of ATRM to accumulate richer historical features. When athletes are occluded by snow spray or undergo dramatic pose transitions, ATRM can retrieve clear and complete reference features from distant historical frames, whereas fixed-window methods in such scenarios can only access equally degraded neighboring frames. Low-contrast degradation amplifies localization ambiguity. In skiing scenarios, athletes’ equipment colors are often similar to the snowy background (white ski suits, gray sky), causing target contours to become blurred in the feature space. This low-contrast degradation makes the inherent ambiguity of boundary predictions far greater than in other sports scenarios. The Probabilistic Boundary Projector of URH explicitly captures this ambiguity, and the Lower Bound Optimizer further constrains the worst-case prediction deviation, effectively suppressing localization jitter caused by unclear contours. The largest AP75 improvement (+5.6) directly reflects this point. The failure of passive fusion is more thorough. The mAP of FGFA (46.8%) and SELSA (48.2%) are not only lower than RT-DETR (50.6%) but also lower than DINO (49.8%), with the performance inversion being the largest among the three datasets. We attribute this to the highly mixed degradation types in skiing scenarios (motion blur, low contrast, and pose deformation often co-occur), making the noise introduction problem of passive fusion particularly severe under these conditions. Computational efficiency analysis To evaluate the additional computational overhead introduced by ARUDet, we compare the model parameters and inference speed of each method on a single NVIDIA A100 GPU using the SportsMOT dataset (input resolution 1280×720) as the test condition. Since the core overhead of ATRM comes from random access and feature retrieval in the memory bank (memory-intensive operations) rather than dense matrix computations, the conventional GFLOPs metric cannot accurately reflect its actual cost. Therefore, we adopt FPS and latency as the primary efficiency measures. Table 5 reports the overall efficiency comparison, and Table 6 provides the overhead breakdown of each module in ARUDet. Table 5Computational efficiency comparison. All methods use the ResNet-50 backbone and are tested on a single NVIDIA A100 GPU with 1280×720 input.MethodTypeParams (M)FPSLatency (ms)RT-DETRImage42.07213.9FGFAVideo50.32050.0SELSAVideo43.81662.5RDNVideo45.21566.7MEGAVideo48.61471.4TransVODVideo46.52835.7MaskVDVideo43.23429.4Beyond BoxesVideo44.73033.3ARUDet (Ours)Video43.54820.8 Computational efficiency comparison. All methods use the ResNet-50 backbone and are tested on a single NVIDIA A100 GPU with 1280×720 input. Table 6Per-module computational overhead breakdown of ARUDet (incremental over the RT-DETR baseline).ModuleAdded Params (M)Added Latency (ms)ATRM (incl. memory bank retrieval)1.26.6URH (replacing the original regression FFN)0.30.3Total Increment1.56.9 Per-module computational overhead breakdown of ARUDet (incremental over the RT-DETR baseline). As shown in Table 5, ARUDet achieves an inference speed of 48 FPS, meeting the real-time application requirement (\ge 30 FPS) while being significantly faster than most video detection methods. Compared to earlier video methods (MEGA 14 FPS, RDN 15 FPS, SELSA 16 FPS), ARUDet achieves approximately 3× speedup, primarily attributable to two factors: first, the RT-DETR base framework itself employs an efficient end-to-end architecture with inference efficiency far superior to traditional video methods built upon Faster R-CNN; second, the two-stage accelerated retrieval strategy of ATRM (coarse-grained frame selection followed by fine-grained location matching) avoids exhaustive pixel-level search across the entire memory bank. Compared to recent methods, ARUDet is also faster than Beyond Boxes (30 FPS) and TransVOD (28 FPS). From the per-module breakdown in Table 6, ATRM is the primary source of overhead (+6.6 ms), dominated by memory bank maintenance and feature retrieval; URH adds only 0.3 ms of inference latency, with negligible computational cost. In terms of parameters, ARUDet adds only 1.5M parameters (+3.6%) over RT-DETR, and the total parameter count (43.5M) is comparable to or lower than most video methods (FGFA 50.3M, MEGA 48.6M, TransVOD 46.5M), because the core operation of ATRM is cosine similarity-based feature retrieval rather than stacking additional parameterized network layers. Overall, ARUDet achieves a reasonable balance between accuracy and efficiency. Scalability and Cost-benefit. The fixed 6.6 ms ATRM overhead pays off non-uniformly: occlusion stratification (Section 4.7) shows mAP gains of +2.2, +6.9, and +10.5 from mild to severe occlusion, so ATRM trades a constant latency for a near-linear payoff with respect to degradation severity. Retrieval cost is and is therefore independent of memory bank size once coarse selection has chosen candidates; memory grows linearly as but is bounded by the FIFO capacity, so per-frame compute stays constant regardless of sequence length. For resource-constrained deployment, reducing from 50 to 20 cuts memory by about 60% with mAP dropping only about 0.8 points, and product quantization of stored features further compresses memory roughly 4×. Ablation studies To verify the contribution of each module in ARUDet, we conduct systematic ablation studies on the SportsMOT dataset, examining module-level ablation, ATRM component ablation, and URH component ablation in turn. Module-level ablation Table 7 reports the module-level ablation results of ATRM and URH. Starting from the RT-DETR baseline, each module is progressively added to isolate its independent contribution and synergistic effect. Table 7Module-level ablation study on the SportsMOT dataset. \checkmark indicates that the module is enabled.ConfigurationATRMURHmAPAP50AP75(a) RT-DETR baseline✗✗57.583.462.3(b) + ATRM\checkmark✗60.085.865.1(c) + URH✗\checkmark59.484.765.0(d) ARUDet (Full)\checkmark\checkmark62.187.267.8 Module-level ablation study on the SportsMOT dataset. \checkmark indicates that the module is enabled. The following key observations can be drawn from the table. First, both modules make independent and significant contributions. ATRM alone brings a +2.5 mAP improvement (configuration b), and URH alone brings a +1.9 mAP improvement (configuration c), with their independent contributions summing to 4.4. Second, their joint use produces a positive synergistic effect. The full model achieves a +4.6 mAP gain, exceeding the sum of the two modules’ independent gains (4.4), indicating that the high-quality features repaired by ATRM provide more reliable input for URH’s probabilistic modeling, forming a mutually beneficial relationship. Third, the two modules exhibit complementary gain patterns. The AP50 gain of ATRM (+2.4) is comparable to its AP75 gain (+2.8), suggesting that feature repair simultaneously improves both detection capability and localization quality; in contrast, URH’s AP75 gain (+2.7) is significantly larger than its AP50 gain (+1.3), confirming that probabilistic boundary modeling primarily benefits high-precision localization. This complementarity also explains why the joint model achieves the largest improvement at AP75 (+5.5). ATRM component ablation To thoroughly analyze the necessity of each design within ATRM, we fix the URH module and sequentially remove individual ATRM components. Figure 4(a) presents the performance of each variant on mAP and AP75 using grouped bar charts. Component ablation studies on SportsMOT. (a) ATRM component ablation: blue bars represent mAP and orange bars represent AP75; values in parentheses indicate the drop relative to the full ATRM. (b) URH component ablation (built upon ATRM): showing the progressive construction from ATRM + L1 regression to the complete URH. The six ATRM variant experiments reveal the following findings. The Defect Responsive Assessor (DRA) is the most critical component, with its removal causing a 2.3 percentage point drop in mAP (62.1→59.8), the largest decline among all ablations. This indicates that the prerequisite for “on-demand repair” is accurately locating the regions that need repair; uniform retrieval without defect awareness cannot match the correct complementary information for degraded regions. The selective replacement mechanism ranks second in contribution after DRA (−2.1 mAP), validating the core observation of this paper: global fusion strategies dilute the original features of high-quality regions, and preserving high-quality regions from interference is crucial. The semantic consistency constraint (−1.9) and quality complementarity constraint (−1.6) have comparable impact, indicating that both constraints in the retrieval process are indispensable: relying solely on semantic similarity may retrieve historical features of equally low quality, while relying solely on quality scores may introduce semantically mismatched noise. Removing degradation type encoding leads to a 1.1 mAP drop, indicating that distinguishing different degradation patterns (occlusion vs. blur vs. low contrast) helps match complementary information more precisely. The adaptation transform has the smallest impact (−0.7) but still contributes, suggesting that although the domain gap between historical and current frames is not the primary concern, appropriate feature alignment can further reduce fusion artifacts. URH component ablation To analyze the progressive contribution of each component within URH, we incrementally build each URH component on top of ATRM. Figure 4(b) illustrates the progression from ATRM + L1 regression to the complete URH. The experimental results exhibit a clear progressive trend. The independent contribution of PBP (+1.0 mAP, +1.4 AP75) is significantly larger than that of LBO (+0.4 mAP, +0.7 AP75), which meets expectations: probabilistic boundary modeling provides the foundational capability of distributional representation, while LBO serves as a further constraint built upon it. Notably, LBO has limited effect without PBP (+0.4 mAP), because the lower bound constraint loses its target in the absence of probabilistic distributions. The joint use of PBP + LBO produces super-linear gains: the fixed α version already achieves +1.7 mAP, exceeding the sum of their independent gains (+1.4). Adaptive confidence level further brings an additional +0.4 mAP gain (61.7→62.1), indicating that samples of different difficulty levels indeed require differentiated confidence constraints. Removing the variance regularization causes a 0.6 mAP drop (62.1→61.5), showing that the lack of variance range constraints leads to unstable variance estimates for some samples, affecting the collaborative optimization of the NLL loss and LBO loss. Hyperparameter sensitivity analysis To examine the sensitivity of ARUDet to key hyperparameters, we conduct experiments on SportsMOT by varying each of the 6 core hyperparameters individually while keeping the remaining parameters at their default values. The following figure presents the effect of each hyperparameter on mAP (solid line) and AP75 (dashed line) in a 2×3 subplot grid. The experimental results reveal the following patterns. Memory bank capacity (Fig. 5, subplot 1,1) exhibits a “rise-then-saturate” trend: from to , mAP steadily improves (59.3→62.1), indicating that a larger memory bank provides richer candidate features for retrieval; from to , there is virtually no gain (62.1→62.2), suggesting that the temporal coverage of 50 frames is sufficient to provide adequately diverse high-quality complementary information, and further enlargement only increases storage and computational overhead without tangible benefit. Sensitivity curves of key hyperparameters on SportsMOT. Each subplot shows the mAP and AP75 trends for one hyperparameter, with circles marking the optimal values. Degradation threshold δ (Fig. 5, subplot 1,2) exhibits an inverted U-shaped curve, peaking at δ = 0.5. An excessively low threshold (δ = 0.3) causes over-repair, where a large number of non-degraded regions are incorrectly flagged for retrieval and replacement, introducing unnecessary feature interference; an excessively high threshold (δ = 0.7) leads to insufficient repair, where some degraded regions are missed and fail to receive effective feature supplementation. Retrieval balance coefficient λ (Fig. 5, subplot 1,3) also exhibits an inverted U-shape, peaking at λ = 0.5, indicating that the semantic consistency and quality complementarity constraints are equally important. Biasing toward the semantic constraint (λ = 0.9) may retrieve semantically matched frames that are equally degraded, while biasing toward the quality constraint (λ = 0.3) may introduce high-quality features that are semantically irrelevant. Coarse selection frame count (Fig. 5, subplot 2,1) rapidly improves from 1 to 5 (60.2→62.1) and then plateaus, indicating that 5 coarse-grained candidates already cover a sufficient retrieval space. Number of degradation prototypes K (Fig. 5, subplot 2,2) reaches the optimum at K = 8 (62.1), with a slight decline at K = 16 (61.8), possibly because too many prototypes lead to overly fine-grained type partitioning and overfitting. LBO loss weight (Fig. 5, subplot 2,3) peaks at 0.5; an excessively large weight () causes the lower bound constraint to be too strong, interfering with the normal optimization of the NLL loss and inducing training instability. Overall, ARUDet maintains stable performance across a wide range of most hyperparameters (mAP fluctuation within 1–2 percentage points), validating the robustness of the method. Robustness and calibration analysis To further understand the behavioral characteristics of ARUDet under different conditions, we conduct analysis from three dimensions: occlusion severity stratification, temporal retrieval span, and uncertainty calibration. Occlusion stratification analysis. We divide the detection samples of SportsMOT into three levels based on the occlusion area ratio: mild (<20%), moderate (20%–50%), and severe (>50%) (Fig. 6(a)). The results clearly demonstrate a trend of increasing gains with heavier occlusion: the gain under mild occlusion is +2.2 (74.5 vs. 72.3), expanding to +6.9 under moderate occlusion (62.7 vs. 55.8), and further reaching +10.5 under severe occlusion (48.9 vs. 38.4). In mild occlusion scenarios, the degree of degradation is limited and baseline methods already perform reasonably well; in severe occlusion scenarios, most of the target’s visual information is lost, and ATRM retrieves complete appearance features from distant historical frames for repair, while URH explicitly models the localization ambiguity caused by occlusion through probabilistic boundary modeling and constrains the worst-case error, with both modules synergistically achieving maximum effect. Beyond Boxes achieves 58.2/41.6 under moderate/severe occlusion, outperforming RT-DETR but still significantly below ARUDet, indicating that passive aggregation strategies within fixed windows still struggle to find sufficient repair information. Robustness and calibration analysis. (a) Detection performance comparison under different occlusion levels on SportsMOT. Red values indicate the absolute gain of ARUDet over RT-DETR. (b),(c) ATRM temporal retrieval span distributions on SportsMOT and SkiTB, respectively. Blue represents non-occluded frames and red represents occluded frames, with dashed lines marking the mean values. (d) Uncertainty calibration scatter plot. Gray denotes KL Loss and red denotes ARUDet, with the dashed line representing the perfect calibration line. Temporal retrieval span. We collect the distribution of temporal distances between ATRM retrieval hits and the current frame on SportsMOT and SkiTB (Fig. 6(b), (c)). On SportsMOT, the retrieval span for non-occluded frames concentrates on nearby frames (mean approximately 9 frames), while the retrieval span for occluded frames is significantly larger (mean approximately 20 frames), with a long tail extending to 40+ frames, indicating that ATRM automatically skips degraded neighboring frames during occlusion and retrieves clean features from more distant history. On SkiTB, the overall retrieval span is larger (non-occluded frame mean approximately 14 frames, occluded frame mean approximately 30 frames), which is consistent with the long-sequence characteristics of SkiTB: sustained low contrast and pose variations require ATRM to look back across longer temporal spans. A common pattern across both datasets is that the average retrieval span for occluded frames is approximately twice that for non-occluded frames, validating the adaptive retrieval capability of ATRM. Uncertainty calibration. To verify whether the σ predicted by URH faithfully reflects the actual localization error, we perform scatter plot analysis between σ and on the SportsMOT test set, comparing with KL Loss (Fig. 6(d)). The scatter distribution of ARUDet exhibits a strong positive correlation (Pearson r ≈ 0.92), with a fitted line slope of approximately 0.85, close to the ideal value of 1.0; in contrast, KL Loss achieves only r ≈ 0.72, with more dispersed scatter points. This difference can be attributed to the NLL loss of PBP providing a direct probabilistic supervision signal, and the confidence lower bound constraint of LBO further refining the estimation range of σ. Well-calibrated uncertainty means that the model’s low σ output indeed corresponds to accurate localization, and high σ output indeed indicates localization deviation, which has important practical value for downstream tracking tasks. Visualization analysis To intuitively demonstrate the detection performance of ARUDet across different sports scenarios, we provide two groups of qualitative analysis: cross-dataset detection result visualization and SportsMOT per-sport detection result visualization. Cross-dataset detection results. Figure 7 presents representative detection results of ARUDet on the three datasets in a 3-row × 6-column grid. The first row shows basketball scenes from the SportsMOT dataset (person detection), the second row shows badminton scenes from the Badminton dataset (player and shuttlecock detection), and the third row shows skiing scenes from the SkiTB dataset (skier detection). Red bounding boxes indicate the predictions of ARUDet. Qualitative detection results of ARUDet across three datasets. Row 1: SportsMOT (basketball, person detection). Row 2: Badminton (player and shuttlecock detection). Row 3: SkiTB (skier detection). Red bounding boxes denote ARUDet predictions. In basketball scenes, ARUDet accurately detects each target under conditions of dense player aggregation and frequent body contact, maintaining robust detection performance even when severe mutual occlusion exists among players. In badminton scenes, ARUDet not only accurately localizes players in various poses (including jumping smashes and large-range movements) but also successfully detects the extremely small and fast-moving shuttlecock, demonstrating the model’s detection capability on multi-scale targets. In skiing scenes, ARUDet still provides accurate localization in the face of multiple challenges including low-contrast backgrounds (white snow and light-colored ski suits), dramatic pose variations (aerial flips in ski jumping, crouching postures in freestyle skiing), and distant small targets, validating the effectiveness of ATRM feature repair and URH uncertainty rectification under extreme degradation conditions. SportsMOT per-sport detection results. Figure 8 further presents the detection results of ARUDet on the three subcategories of SportsMOT in a 3-row × 6-column grid. The first row shows basketball, the second row shows football, and the third row shows volleyball. Qualitative detection results of ARUDet on the three SportsMOT subcategories. Row 1: Basketball. Row 2: Football. Row 3: Volleyball. Red bounding boxes denote ARUDet predictions. In basketball scenes, ARUDet excels in high-density interaction scenarios such as half-court transitions and paint area confrontations, distinguishing closely guarded players and providing accurate detection boxes for each individual. In football scenes, players appear as relatively small targets in the frame due to the distant camera viewpoint, with a wide distribution range from foreground to background; ARUDet maintains stable detection accuracy across different scales, with no significant missed detections even for small targets in distant views. In volleyball scenes, the net divides the frame into two regions, players are frequently partially occluded by the net, and large-range pose variations such as jumping spikes are common; ARUDet maintains reliable detection performance under these conditions as well. The visualization results across the three sports collectively demonstrate that ARUDet possesses cross-scenario generalization capability, adapting to the distinctive degradation patterns inherent in different sports types. Limitations and future work Limitations. Although ARUDet achieves significant improvements on multiple sports datasets, two limitations remain. First, the memory bank retrieval of ATRM introduces additional inference latency, reducing the inference speed from 72 FPS of RT-DETR to 48 FPS, which may become a bottleneck in multi-stream concurrent broadcasting or edge device deployment scenarios; although the two-stage acceleration strategy has significantly reduced the retrieval complexity, the random memory access pattern of the memory bank inherently conflicts with the continuous computation paradigm of GPUs, and further optimization requires joint advances at both the algorithmic and engineering levels. Second, ATRM relies on the availability of retrievable high-quality complementary information in the memory bank; when drastic shot transitions, abrupt viewpoint changes, or prolonged complete disappearance and reappearance of targets occur, the memory bank may not contain semantically consistent historical features, causing retrieval quality to degrade or even introduce mismatched noise; the retrieval span analysis indicates that the current ATRM can effectively look back in most occlusion scenarios, but for extreme long-duration degradation exceeding the memory bank capacity, a reliable fallback mechanism is still lacking. Future work. Based on the above analysis, we plan to advance subsequent research in the following directions. (1) Explore hierarchical memory banks and approximate nearest neighbor indexing (such as locality-sensitive hashing or product quantization), combined with dynamic eviction strategies and knowledge distillation schemes to further compress retrieval latency, promoting the deployment of ARUDet on edge devices. (2) Introduce scene change detection and automatic memory bank reset mechanisms, and explore cross-viewpoint feature alignment techniques to improve the robustness of ATRM in strongly non-stationary scenarios such as shot transitions and multi-camera broadcasting. (3) Directly inject the uncertainty estimates output by URH into the tracker’s association cost matrix and the observation noise model of the Kalman filter to achieve end-to-end joint optimization of detection and tracking, and validate the transferability and generalization capability of the method on general video detection benchmarks (such as ImageNet VID) and multi-category targets. Conclusion This paper proposes ARUDet, a robust video object detection framework designed for sports scenarios, addressing the detection challenges posed by high-speed motion, severe occlusion, and complex deformation from two complementary dimensions: feature repair and localization reliability. To tackle the information mismatch problem caused by the passive feature aggregation of existing methods, the Active Temporal Retrieval Module (ATRM) identifies the type and severity of degradation in each region through a Defect Responsive Assessor, and actively retrieves targeted historical features from a temporal memory bank under dual constraints of semantic consistency and quality complementarity, performing selective replacement only on low-quality regions to achieve on-demand repair. To address the problem of deterministic regression ignoring geometric ambiguity, the Uncertainty Rectified Regression Head (URH) extends boundary predictions to Gaussian distributions through a Probabilistic Boundary Projector, and constrains worst-case localization error via a Lower Bound Optimizer, synergistically achieving explicit modeling and effective suppression of uncertainty. Experiments on three datasets, SportsMOT, Badminton-Players-Detection, and SkiTB, demonstrate that ARUDet achieves the best results across all evaluation metrics, with the most significant advantages under strict IoU thresholds (AP75) and severe occlusion scenarios. Ablation studies validate the independent contributions and positive synergistic effects of the two modules, and robustness analysis further confirms the adaptive retrieval capability of ATRM and the uncertainty calibration quality of URH. Author contributions Lijing Yu was responsible for the overall conceptualization and methodology design, conducted the experiments, and drafted the initial manuscript. Jin Gao participated in data curation, prepared figures and tables, and contributed to the revision and editing of the manuscript. Xinqi Ji and Bo Liu supervised the entire study, provided critical revisions, and approved the final version of the manuscript as the corresponding authors. All authors reviewed and approved the final manuscript and take full responsibility for its content and accuracy. Funding This work was supported by the National Social Science Fund of China (Grant Nos. 24BJY086 and 23FGLB024) and the Key Project of the Professional Competence Committee, China Association of Plant Engineering (Grant No. ZYSYX00690). Data availability All datasets used in this study are publicly available. The SportsMOT dataset is accessible through its official GitHub repository (https://github.com/MCG-NJU/SportsMOT). The Badminton-Players-Detection dataset is collected and annotated by us and will be released upon acceptance. The SkiTB dataset can be obtained from the Machine Learning Group at the University of Udine (https://machinelearning.uniud.it/datasets/skitb). Code will be released upon acceptance. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Jin Gao, Email: chdjgxy@chd.edu.cn. Bo Liu, Email: 202113@cdmc.edu.cn. References Associated Data Data Availability Statement All datasets used in this study are publicly available. The SportsMOT dataset is accessible through its official GitHub repository (https://github.com/MCG-NJU/SportsMOT). The Badminton-Players-Detection dataset is collected and annotated by us and will be released upon acceptance. The SkiTB dataset can be obtained from the Machine Learning Group at the University of Udine (https://machinelearning.uniud.it/datasets/skitb). Code will be released upon acceptance.