MDF-Det: Motion-Aware Decoupling and Scene Filtering for Wide Area Small Moving Target Detection Authors: Kangqiushi Li, Xiaoran Zhang, Zheng Zhang, Huaxin Xiao, Yu Liu Published: 2026 ABSTRACT Highlights What are the main findings? MDF-Det establishes a unified coarse-to-fine framework that jointly addresses weak-target preservation, dense-target decoupling, and scene-induced false-alarm suppression in wide-area motion imagery.The spatial attention-guided decoupling and learned scene-prior filtering mechanisms separate merged target responses and suppress contextually implausible detections, enabling MDF-Det to achieve an average F1 score of 0.878 across six WPAFB AOIs. What are the implications of the main findings? The results demonstrate that reliable WAMI detection requires coordinated control of candidate recall, dense-target separation, and scene-level contextual filtering rather than relying solely on local motion cues.The proposed modular framework provides an interpretable and extensible solution for detecting tiny and densely distributed moving targets under complex wide-area aerial imaging conditions. AbstractObject detection in Wide Area Motion Imagery (WAMI) is crucial for large-scale intelligent surveillance and monitoring systems. However, detecting extremely small moving targets in low-frame-rate grayscale WAMI remains highly challenging. In particular, three key problems limit the performance of existing methods: weak motion responses caused by low target contrast can lead to missed detections; densely distributed targets often produce merged responses that are difficult to separate; and registration artifacts, parallax, and dynamic background clutter can generate numerous false alarms. To address these problems, we propose MDF-Det, a coarse-to-fine spatiotemporal framework for WAMI small moving target detection. The framework consists of three complementary components. First, a Motion and Appearance Feature Fusion (MAFF) strategy integrates dense optical-flow-derived motion saliency with grayscale frame differencing to enhance weak target responses and improve candidate preservation. Second, a Spatial Attention-Guided Target Decoupling (SA-TD) module employs fine-grained heatmap decoding and step-threshold degradation to separate merged responses in densely populated scenes. Finally, a Scene-Prior Guided Filtering (SPGF) mechanism learns complementary vehicle-accessibility and motion-activity priors from scene context to suppress contextually implausible false alarms caused by complex background interference. Extensive experiments on six evaluation AOIs of the WPAFB 2009 dataset demonstrate that MDF-Det achieves an average F1 score of 0.878, corresponding to a relative improvement of 5.5% over the strongest baseline. 1. Introduction Wide-Area Motion Imagery (WAMI) is an advanced aerial imaging technology that enables persistent coverage of large geographical areas, often spanning entire cities or tens of square kilometers [1]. The detection and tracking of moving objects such as vehicles and pedestrians in WAMI constitute important tasks for a range of emerging applications, including urban monitoring, critical infrastructure protection, disaster response, border surveillance, and defense [2,3,4]. Recent studies have also explored neuromorphic event cameras for WAMI, demonstrating their potential for effective aerial sensing under low-frame-rate and challenging illumination conditions [5]. With continued advancements in sensor and platform technology, WAMI data collection has become increasingly accessible, driving a growing demand for object detection and tracking algorithms to support effective wide-area surveillance. However, object detection in WAMI presents unique and formidable challenges when compared to the task of object detection in ordinary videos [1,6,7]. First, the individual frames have extremely high resolutions. For example, the WPAFB dataset [8] contains frames exceeding 315 megapixels, whereas moving vehicles typically occupy only about 50 pixels on average [9], as illustrated in Figure 1a. Second, WAMI data are often acquired at a low frame rate (typically about 1∼2 frames per second). This slow temporal sampling causes fast-moving vehicles to exhibit large interframe displacements that often exceed the target’s own size, which severely undermines conventional motion-based methods that rely on small, smooth motion assumptions [6,10]. Third, these tiny targets are densely distributed across complex, heterogeneous aerial backgrounds, making efficient and accurate localization highly challenging. To address these challenges, WAMI object detection has evolved from traditional motion-based techniques [4,10] to deep spatiotemporal frameworks [6,11,12]. Yet, despite these advances, the extreme imaging conditions of WAMI still expose several critical limitations in existing approaches, as illustrated in Figure 1. First, the extremely small target size, weak appearance contrast, and complex aerial backgrounds often obscure subtle motion responses, making faint targets difficult to preserve during candidate generation (Figure 1b) [6,10]. Second, in high-density traffic scenes, neighboring vehicles may produce highly overlapping responses, which severely hinders conventional heatmap regression from separating densely packed targets (Figure 1c) [3,4,12]. Third, without scene-level contextual constraints, existing detectors may misclassify registration artifacts, building parallax, swaying vegetation, and other deceptive background responses as genuine moving targets, resulting in numerous contextually implausible false positives (Figure 1d) [3,10]. To deal with these problems, we propose MDF-Det, a coarse-to-fine framework for vehicle detection in WAMI. By integrating temporal motion priors with grayscale appearance-change features, the proposed framework adaptively enhances low-contrast targets and subtle motion signals at the front end. It then exploits spatial attention mechanisms to separate densely clustered targets in the deep feature space, while further incorporating scene-level contextual priors to suppress background artifacts. Together, these components form an effective cascaded detection pipeline that addresses the inherent challenges of wide-area small moving target detection. The main contributions of this work are summarized as follows:We propose MDF-Det, a unified coarse-to-fine framework for small moving target detection in WAMI. The proposed cascaded pipeline sequentially improves candidate quality through motion–appearance fusion, separates densely clustered targets, and suppresses false positives via scene-level filtering.We develop a Motion and Appearance Feature Fusion (MAFF) strategy. By integrating temporal motion saliency derived from optical flow with grayscale difference features, MAFF adaptively enhances low-contrast signals, thereby improving the detection capability for faint targets.We design a Spatial Attention-Guided Target Decoupling (SA-TD) module. By combining fine-grained heatmap decoding with a step-threshold degradation algorithm, the proposed module effectively separates overlapping targets in highly crowded scenarios.We introduce a Scene-Prior Guided Filtering (SPGF) mechanism. By learning complementary vehicle-accessibility and motion-activity priors from trajectory-supervised scene context, SPGF suppresses false positives caused by background clutter in contextually implausible regions.Extensive experiments on the WPAFB 2009 dataset demonstrate that MDF-Det achieves the highest F1 scores across all six evaluation AOIs, with an average F1 score of 0.878, corresponding to a relative improvement of 5.5% over the strongest baseline. We propose MDF-Det, a unified coarse-to-fine framework for small moving target detection in WAMI. The proposed cascaded pipeline sequentially improves candidate quality through motion–appearance fusion, separates densely clustered targets, and suppresses false positives via scene-level filtering. We propose MDF-Det, a unified coarse-to-fine framework for small moving target detection in WAMI. The proposed cascaded pipeline sequentially improves candidate quality through motion–appearance fusion, separates densely clustered targets, and suppresses false positives via scene-level filtering. We develop a Motion and Appearance Feature Fusion (MAFF) strategy. By integrating temporal motion saliency derived from optical flow with grayscale difference features, MAFF adaptively enhances low-contrast signals, thereby improving the detection capability for faint targets. We develop a Motion and Appearance Feature Fusion (MAFF) strategy. By integrating temporal motion saliency derived from optical flow with grayscale difference features, MAFF adaptively enhances low-contrast signals, thereby improving the detection capability for faint targets. We design a Spatial Attention-Guided Target Decoupling (SA-TD) module. By combining fine-grained heatmap decoding with a step-threshold degradation algorithm, the proposed module effectively separates overlapping targets in highly crowded scenarios. We design a Spatial Attention-Guided Target Decoupling (SA-TD) module. By combining fine-grained heatmap decoding with a step-threshold degradation algorithm, the proposed module effectively separates overlapping targets in highly crowded scenarios. We introduce a Scene-Prior Guided Filtering (SPGF) mechanism. By learning complementary vehicle-accessibility and motion-activity priors from trajectory-supervised scene context, SPGF suppresses false positives caused by background clutter in contextually implausible regions. We introduce a Scene-Prior Guided Filtering (SPGF) mechanism. By learning complementary vehicle-accessibility and motion-activity priors from trajectory-supervised scene context, SPGF suppresses false positives caused by background clutter in contextually implausible regions. Extensive experiments on the WPAFB 2009 dataset demonstrate that MDF-Det achieves the highest F1 scores across all six evaluation AOIs, with an average F1 score of 0.878, corresponding to a relative improvement of 5.5% over the strongest baseline. Extensive experiments on the WPAFB 2009 dataset demonstrate that MDF-Det achieves the highest F1 scores across all six evaluation AOIs, with an average F1 score of 0.878, corresponding to a relative improvement of 5.5% over the strongest baseline. The remainder of this paper is organized as follows. Section 2 reviews the literature on moving object detection in WAMI. Section 3 presents the proposed MDF-Det framework and its core components. Section 4 describes the dataset, evaluation metrics, baseline methods, implementation details, and network training settings. Section 5 presents the experimental results. Finally, Section 6 discusses the main findings and limitations, and Section 7 concludes the paper. 2. Related Work The primary objective of object detection in WAMI is to efficiently localize and track high-value dynamic targets, such as moving vehicles, across gigapixel-scale aerial backgrounds. Based on methodological evolution, existing detection approaches can be broadly categorized into three distinct classes: traditional techniques relying on temporal motion cues, two-stage spatiotemporal architectures, and end-to-end single-stage networks driven by heatmap regression. 2.1. Traditional Methods Utilizing Temporal Motion Cues Early research in WAMI primarily relies on temporal motion cues to compensate for the scarcity of spatial appearance details. Background subtraction and multi-frame differencing constitute the foundational techniques for extracting motion pixels from quasi-static backgrounds. Although initial approaches frequently employ two- or three-frame differencing to capture dynamic responses, they suffer from the aperture problem for slow-moving targets and ghosting artifacts induced by the displacement of objects [12]. To adapt to gradual environmental variations, subsequent research introduces more statistical techniques for background modeling—such as adaptive Gaussian Mixture Models [13] and median or mean filtering algorithms [4,14]—to separate moving vehicles from static scenes. Despite the computational efficiency of these methods, they remain highly susceptible to errors in image registration. Specifically, parallax-induced artifacts frequently generate false positives within these traditional frameworks. Furthermore, they inherently struggle to detect slow-moving targets and objects with low visual contrast against the background. 2.2. Two-Stage Architectures with Spatiotemporal Context To balance the extensive background search space against computational constraints, two-stage coarse-to-fine detection architectures emerge as a pivotal strategy in this domain. Early hybrid frameworks typically employ traditional techniques, such as background subtraction, to generate candidate regions for initial coarse screening. Subsequently, they cascade lightweight Convolutional Neural Networks (CNNs) as backend classifiers to filter out false positives and separate tightly clustered targets [15]. Although these pipelines prioritize efficiency, the overall performance remains highly susceptible to the quality of the initial physical preprocessing. To circumvent these limitations, subsequent research transitions the two-stage approach toward fully deep convolutional architectures. Pioneering this shift, ClusterNet [11] introduces an anchor-free, fully deep hierarchical framework. In this architecture, the first stage extracts large-scale motion clusters, whereas the second stage performs refined localization via heatmap regression. This design improves detection sensitivity and represents an early deep-learning approach in WAMI for detecting tiny vehicles. Although two-stage architectures effectively balance search efficiency and detection accuracy, they remain heavily dependent on the quality of proposals generated in the first stage. If a missed detection occurs during this initial phase, the downstream deep networks lack the inherent capacity to recover the target, which ultimately leads to irrecoverable target loss. 2.3. Heatmap-Based Single-Stage Regression Networks Inspired by anchor-free methods such as CenterNet [16], regression based on heatmaps gradually emerges as a leading approach in the field of WAMI. By representing targets as two-dimensional Gaussian keypoints and directly regressing the center positions, these methods effectively bypass the complex design and matching heuristics associated with traditional anchor boxes. Representative networks, such as HM-Net [6] and HMRN [17], extend this concept by incorporating the image and the predicted heatmap of the previous frame as temporal feedback priors alongside the input of the current frame. This architecture enables the joint detection and prediction of target displacement in a single forward pass, thereby efficiently facilitating the tracking of multiple targets. To capture complex long-range spatiotemporal dynamics, Negin et al. [12] introduce a self-attention mechanism that leverages VQ-VAE to extract unsupervised, discrete spatiotemporal embeddings. These latent embeddings are subsequently integrated into an encoder-decoder architecture based on the Transformer to generate refined heatmaps of keypoints. Furthermore, by incorporating the regularization of shape priors into the loss function, the framework encourages the network to produce responses that align with the physical morphology of small targets. However, lacking decoupling mechanisms for complex dense scenarios, these methods frequently suffer from overlapping responses in heatmaps when targets are clustered, which leads to a significant degradation in the accuracy of localization. Limitations and Motivation: In summary, although existing approaches have advanced vehicle detection in WAMI, they still struggle to jointly preserve weak motion responses and separate densely clustered targets during subsequent refinement. Motivated by these limitations, we reformulate the conventional coarse-to-fine detection pipeline and propose MDF-Det. By enhancing faint motion signals through motion priors, separating overlapping targets via fine-grained heatmap decoding, and suppressing contextually implausible false positives using scene-level contextual constraints, MDF-Det provides a unified framework for wide-area moving target detection. 3. Proposed Method 3.1. Overview As illustrated in Figure 2, MDF-Det adopts a coarse-to-fine architecture for wide-area moving target detection. The pipeline begins with the Motion and Appearance Feature Fusion (MAFF) strategy, which incorporates temporal motion priors to enhance weak target responses and generate coarse candidates. These candidates are subsequently processed by the Spatial Attention-Guided Target Decoupling (SA-TD) module, where fine-grained heatmap decoding separates merged target responses in densely populated scenes. Finally, the Scene-Prior Guided Filtering (SPGF) mechanism exploits learned scene-level contextual priors to suppress false responses caused by parallax and dynamic background clutter, yielding the final detection results. 3.2. Motion and Appearance Feature Fusion Strategy In WAMI, traditional background subtraction algorithms predominantly rely on pixel-level grayscale differences. Consequently, when dealing with low-contrast or slow-moving targets, rigid difference thresholds may lead to missed detections [18]. To improve candidate region extraction, we propose the MAFF strategy, which combines two complementary cues. Specifically, dense optical flow provides explicit displacement-based motion information by estimating inter-frame pixel motion, whereas grayscale differencing captures photometric appearance changes through local intensity variations between the current frame and the temporally aligned reference background. MAFF integrates these motion and appearance-change cues to enhance weak target responses while preserving the efficiency of the conventional grayscale difference framework. 3.2.1. Spatial Registration and Photometric Compensation Given a current frame It and a sequence of historical frames {It−k}k=1K, our initial objective is to eliminate the global background displacement induced by the ego-motion of the aerial platform. To achieve this, we first employ the ORB feature extractor [19] to compute image correspondences and estimate the global homography matrix Tk∈R3×3, which maps each historical frame to the current frame. Utilizing this matrix, we execute a perspective transformation [20] to project the historical frames into the unified spatial coordinate system of It. This spatial registration process yields the aligned template set, denoted as {I˜t−k}k=1K:(1)I˜t−k(x)=It−kW(x;Tk−1) where x=[x,y]T denotes the two-dimensional pixel coordinates in the current frame, and W(·) represents the spatial warping and bilinear resampling operator parameterized by the inverse homography matrix. During the registration process, out-of-bounds regions and void areas lacking pixel values—induced by the perspective transformation—are systematically extracted and morphologically dilated to construct a global invalid mask, Ωinvalid, for subsequent processing. Despite achieving precise spatial alignment, dynamic factors such as cloud occlusion and viewpoint variations frequently introduce significant photometric inconsistencies. Consequently, prior to constructing the reference background, we apply a large-scale Gaussian low-pass filter, denoted as Gillum, to perform adaptive photometric compensation on the aligned templates [15]. This operation effectively isolates the low-frequency illumination drift components:(2)I^t−k=ClipI˜t−k+GillumIt−I˜t−k,0,255 where Clip(·,0,255) is a clipping function that bounds the pixel values within the standard grayscale interval of [0,255]; and I^t−k represents the resulting photometrically aligned template. Utilizing the photometrically aligned template set {I^t−k}k=1K, we construct a local reference background, Bt, via temporal median filtering Median(·). Subsequently, we compute the initial absolute grayscale difference map Dt:(3)Bt(x,y)=Median1≤k≤KI^t−k(x,y),Dt=|It−Bt| 3.2.2. Enhanced Detection and Candidate Generation To overcome the inherent limitations of traditional grayscale differencing—which frequently fails to detect low-contrast, slow-moving targets—we adopt a dense optical flow algorithm based on local polynomial expansion [21,22]. This approach explicitly captures subtle motion information, thereby enhancing weak motion responses during candidate generation. Specifically, we extract the pixel-level motion field d=(u,v)T between the current frame It and the preceding frame I^t−1 following spatial registration and photometric compensation. For a given spatial coordinate x=[x,y]T, the algorithm first approximates the image intensity signal within its local neighborhood as a quadratic polynomial surface:(4)It(x)≈xTAtx+btTx+ct where At denotes a symmetric matrix encoding the local spatial second-order derivatives; bt is a column vector representing the first-order spatial gradients; and ct is a scalar corresponding to the locally smoothed central grayscale intensity. Assuming the target satisfies the brightness constancy constraint over an infinitesimal time interval—formally expressed as It(x)=I^t−1(x−d)—equating the polynomial coefficients of adjacent frames yields the ideal theoretical constraint: Atd=−12(bt−b^t−1). However, to ensure effective motion estimation in the presence of image noise, we formulate a weighted least squares optimization objective with respect to the displacement vector d over a local spatial neighborhood Q:(5)d*(x)=argmind∑Δx∈Qw(Δx)∥A(x+Δx)d−Δb(x+Δx)∥2 where A(·)=12(At(·)+A^t−1(·)) denotes the arithmetic mean of the second-order coefficient matrices across the adjacent frames; Δb=−12(bt−b^t−1) represents the scaled difference vector of the first-order gradients; Δx indicates the spatial coordinate offset within the local neighborhood Q; and w(·) is a two-dimensional Gaussian weighting function that governs the neighborhood contributions. Following the estimation of the optical flow field, we derive a normalized two-dimensional temporal motion saliency map, Mt∈[0,1]W×H, by computing the pixel-wise motion magnitude and subsequently applying a small-kernel Gaussian smoothing filter, Gflow:(6)Mt=NGflowu2+v2 where W and H denote the spatial dimensions (width and height) of the input image, respectively; and N(·) represents a linear normalization operator that scales the local motion magnitude into the interval [0,1]. Subsequently, we introduce an independent additive motion offset term to explicitly integrate the motion saliency as a physical prior, thereby yielding the enhanced difference map Dt′:(7)Dt′=Dt,if(x,y)∈ΩinvalidDt+α(Mt−Tm),otherwise where Mt denotes the estimated motion magnitude; Tm represents a static noise threshold; and α is a scaling factor that modulates the motion offset. This additive formulation selectively amplifies the signal in dynamic regions (Mt>Tm) while penalizing static background responses. Subsequently, we apply a binarization threshold, τdiff, to extract the potential target regions:(8)Bcand(x)=IDt′(x)≥τdiff where Bcand denotes the resulting binary candidate mask, wherein pixels with a value of 1 designate potential moving targets; I(·) is the standard indicator function. Following the application of morphological denoising to Bcand, we partition the binary mask into 3×3 non-overlapping spatial cells to facilitate lightweight, grid-based non-maximum suppression and eliminate redundant, locally clustered responses [23]. As illustrated in Figure 3, by jointly executing local max pooling and center-anchored mapping, the extraction of the initial candidate center set, Ccand, is mathematically formulated as:(9)Ccand={(3xd+1,3yd+1)∉Ωinvalid| maxi,j∈{0,1,2}Bcand(3xd+i,3yd+j)=1} where (xd,yd) denotes the downsampled grid coordinates; the constraint ∉Ωinvalid ensures that the remapped candidate coordinates strictly avoid the predefined global invalid regions; and the constant offset of +1 serves to anchor these coordinates to the geometric center of each activated cell. Subsequently, this sparsified candidate set is forwarded for classification and target decoupling. The pseudocode for the proposed MAFF strategy is summarized in Algorithm 1. Algorithm 1: Motion and Appearance Feature Fusion StrategyInput : Current frame It and sequence of historical frames {It−k}k=1K; Hyperparameters α,Tm,τdiff.Output: Coarse candidate center set Ccand. 1 {I˜t−k},Ωinvalid←SpatialRegistration({It−k}k=1K,It) 2 {I^t−k}←PhotometricCompensation({I˜t−k},It) 3 Bt←Median1≤k≤K({I^t−k}) 4 Dt←|It−Bt| 5 (u,v)←DenseOpticalFlow(It,I^t−1) 6 Mt←N(Gflow(u2+v2)) 7 Dt′←EnhanceDifference(Dt,Mt,Tm,α,Ωinvalid) 8 Bcand←I(Dt′≥τdiff) 9 Ccand←GridNMS(Bcand,Ωinvalid)10 return Ccand Input : Current frame It and sequence of historical frames {It−k}k=1K; Hyperparameters α,Tm,τdiff. Output: Coarse candidate center set Ccand. 1 {I˜t−k},Ωinvalid←SpatialRegistration({It−k}k=1K,It) 2 {I^t−k}←PhotometricCompensation({I˜t−k},It) 3 Bt←Median1≤k≤K({I^t−k}) 4 Dt←|It−Bt| 5 (u,v)←DenseOpticalFlow(It,I^t−1) 6 Mt←N(Gflow(u2+v2)) 7 Dt′←EnhanceDifference(Dt,Mt,Tm,α,Ωinvalid) 8 Bcand←I(Dt′≥τdiff) 9 Ccand←GridNMS(Bcand,Ωinvalid)10 return Ccand 1 {I˜t−k},Ωinvalid←SpatialRegistration({It−k}k=1K,It) {I˜t−k},Ωinvalid←SpatialRegistration({It−k}k=1K,It) 2 {I^t−k}←PhotometricCompensation({I˜t−k},It) {I^t−k}←PhotometricCompensation({I˜t−k},It) 3 Bt←Median1≤k≤K({I^t−k}) Bt←Median1≤k≤K({I^t−k}) 4 Dt←|It−Bt| Dt←|It−Bt| 5 (u,v)←DenseOpticalFlow(It,I^t−1) (u,v)←DenseOpticalFlow(It,I^t−1) 6 Mt←N(Gflow(u2+v2)) Mt←N(Gflow(u2+v2)) 7 Dt′←EnhanceDifference(Dt,Mt,Tm,α,Ωinvalid) Dt′←EnhanceDifference(Dt,Mt,Tm,α,Ωinvalid) 8 Bcand←I(Dt′≥τdiff) Bcand←I(Dt′≥τdiff) 9 Ccand←GridNMS(Bcand,Ωinvalid) Ccand←GridNMS(Bcand,Ωinvalid) 10 return Ccand return Ccand 3.3. Spatial Attention-Guided Target Decoupling Module Although the front-end perception stage captures moving pixels, the resulting candidate regions frequently exhibit severe spatial overlap among targets [24]. To resolve this challenge, we propose the SA-TD module. This module transitions the detection task from conventional coarse-grained bounding box regression to fine-grained heatmap decoding, thereby enabling the decoupling of densely clustered targets within the deep feature space. 3.3.1. Candidate Screening via Binary Classification To eliminate false positives induced by parallax effects and dynamic background clutter, we initially employ a lightweight binary classifier to serve as a spatiotemporal filter [15]. Let Ccand={ci}i=1Nc denote the set of coarse candidate centers derived from the preceding background subtraction stage. For each candidate center ci, we crop localized patches from the current frame t and three consecutive historical templates (i.e., t−1, t−2, and t−3) to construct a spatiotemporal input tensor, Xibin∈R4×21×21. Ultimately, the probability of genuine target presence, pi, is predicted by the binary classification network Fbinary:(10)pi=Fbinary(Xibin;θbin) where θbin denotes the learnable parameters of the network. To suppress evident background clutter while preserving potential challenging targets, we establish a confidence threshold, τbin, to conduct a preliminary screening of the initial candidate set:(11)Cpending=ci∈Ccand|pi≥τbin Following this lightweight filtering stage, the retained candidate set, Cpending, is forwarded to the subsequent multi-target heatmap regression network. 3.3.2. Spatial Attention-Guided Heatmap Generation To establish a unified mathematical formulation for the network architecture, we initially define a generalized dual-convolutional base operator, denoted as Φm(·). For any given input feature tensor F, its strictly cascaded expansion is mathematically defined as:(12)Φm(F)=ReLU(BN(Convm3×3(ReLU(BN(Convm3×3(F)))))) where m denotes the output channel dimension of the module; Convm3×3 represents a 3×3 convolutional operator for feature mapping that yields m output channels; BN signifies the batch normalization operation; and ReLU corresponds to the Rectified Linear Unit non-linear activation function. With the base operator established, for each candidate point ci∈Cpending, we extract a localized spatiotemporal patch of 45×45 pixels to formulate the regression input tensor, Xreg∈R4×45×45. This tensor is subsequently processed through two cascaded feature extraction stages—each equipped with a 2×2 spatial max pooling operation (M↓)—and is ultimately projected to yield the mid-level feature map Fmid∈R64×11×11:(13)Fmid=M↓Φ64M↓(Φ32(Xreg)) To accentuate the subtle motion signatures of tiny targets amidst complex background clutter, we integrate a spatial attention module within the mid-level feature space [25]. Specifically, we initially derive two-dimensional spatial descriptors by applying average pooling (Pavg) and max pooling (Pmax) operations across the channel axis:(14)Favgs=Pavg(Fmid),Fmaxs=Pmax(Fmid) where Favgs∈R1×11×11 denotes the average-pooled feature descriptor, which aggregates global statistical information; and Fmaxs∈R1×11×11 represents the max-pooled feature descriptor, which captures the most salient local motion responses. By collapsing the channel dimension, these paired descriptors effectively steer the network’s attention toward spatial regions exhibiting prominent motion signatures. These descriptors are concatenated along the channel axis and subsequently processed through a 7×7 convolutional layer, Conv17×7, coupled with a Sigmoid activation function to derive the spatial attention weight map As∈R1×11×11:(15)As=SigmoidConv17×7([Favgs;Fmaxs]) where Conv17×7 denotes a spatial convolutional operator featuring a 7×7 kernel and yielding a single output channel; and [·;·] represents the channel-wise feature concatenation operation. Subsequently, the attention mask As is element-wise multiplied with the mid-level feature map Fmid, and the resulting tensor is directly propagated to the high-order feature extraction stage. By employing the previously defined generalized operator Φm(·), these saliency-enhanced features are further projected into high-order semantic representations, yielding Fhigh∈R128×11×11:(16)Fhigh=Φ128Fmid⊗As where ⊗ denotes the spatial element-wise multiplication operation. In contrast to conventional networks that regress a single coordinate offset, the SA-TD module decodes the high-order semantic representations, Fhigh, into a spatial probability distribution to effectively decouple multiple densely clustered targets within the localized patch. Specifically, the flattened high-order feature tensor is mapped to a probability vector via a Multi-Layer Perceptron and subsequently reshaped to form a coarse local heatmap, H^coarse∈[0,1]15×15:(17)H^coarse=Rreshape15×15(Sigmoid(W225ReLUBN(W512Flatten(Fhigh)))) where Flatten(·) denotes the operation that flattens a three-dimensional feature tensor into a one-dimensional vector; W512 and W225 represent the weight matrices of the fully connected layers, projecting the features into 512 and 225 dimensions, respectively; and Rreshape15×15(·) signifies the spatial reshaping operator that rearranges the resulting 225-dimensional probability vector into a 15×15 two-dimensional grid. 3.3.3. Optimization Objective and Step-Degradation Inference During the training phase, given a localized patch encompassing Np genuine moving targets, let {μj}j=1Np denote their corresponding relative center coordinates within the local patch coordinate frame. The ground-truth heatmap, H, for this specific patch is subsequently synthesized by rendering two-dimensional Gaussian kernels:(18)H(x)=max1≤j≤Npexp−∥x−μj∥222σp2 where x=[x,y]T denotes the two-dimensional spatial coordinate vector within the localized patch; μj=[μxj,μyj]T represents the relative center coordinates of the j-th target; and σp signifies the standard deviation parameter governing the spatial extent of the Gaussian kernel. To model the spatial distribution of targets, the network is optimized via the Mean Squared Error loss function:(19)Lreg=1Nr∑x∥H^coarse(x)−H(x)∥22 where Nr denotes the total number of spatial pixels comprising the coarse heatmap. During the inference phase, the predicted coarse heatmap is bilinearly upsampled to the original localized patch resolution of 45×45, yielding the refined heatmap H^fine. To decouple and extract tiny targets exhibiting severe spatial overlap within densely clustered scenes—and to effectively mitigate the missed detection issue induced by the mutual suppression of adjacent genuine targets—we introduce a step-threshold degradation algorithm founded on topological connected components. Specifically, we define a monotonically decreasing discrete threshold sequence, T={τ0,τ1,…,τS}, where S represents the total number of degradation steps. The initial threshold is initialized as τ0=max(H^fine) and is iteratively decayed by a constant step size of Δτ until reaching a predefined lower bound, τmin. During the s-th iteration step (s∈{0,1,…,S}) associated with the specific threshold τs, we extract the set of independent connected components, denoted as O(s)={oj(s)}, wherein the refined heatmap satisfies the condition H^fine≥τs. The spatial coordinates of the target candidates are subsequently determined by identifying the local maxima within each respective connected component. This yields the candidate set for the current iteration, Vs, formulated as:(20)Vs={argmaxx∈oj(s)H^fine(x)|oj(s)∈O(s)} To preclude the redundant extraction of identical salient targets across varying threshold levels, we introduce an inter-step exclusion mechanism designed to dynamically maintain a global set of valid coordinates, D. Initializing with D0=V0, the subsequent updates to this global set strictly adhere to a set difference operational rule:(21)Ds=Ds−1∪Vs∖Ds−1 This exclusion operation (∖) prevents previously detected coordinates from being repeatedly added across different threshold levels. Upon reaching the terminal threshold, τS, the resulting set, DS, provides the spatial coordinates of the separated densely clustered targets within the localized patch. To illustrate this procedure, Figure 4 shows the progressive target decoupling process. Starting from the refined heatmap H^fine (Figure 4a), a high initial threshold (τ1=0.97, Figure 4b) extracts the most salient targets as independent connected components. As τ gradually decreases (τ7=0.90, Figure 4c), additional weaker target responses progressively emerge. At the terminal threshold (τmin=0.17, Figure 4d), further target responses are separated, while the inter-step exclusion mechanism reduces redundant detections across different threshold levels. The pseudocode for the proposed SA-TD module is summarized in Algorithm 2. Algorithm 2: Spatial Attention-Guided Target Decoupling Module 3.4. Scene-Prior Guided Filtering Mechanism Although SA-TD effectively separates densely clustered target responses, candidate-level features alone remain insufficient to suppress false alarms caused by building parallax, registration residuals, vegetation motion, and other structured background interference [26]. We therefore introduce the SPGF mechanism to provide a soft scene-level constraint for candidate verification. SPGF estimates vehicle-compatible scene responses directly from visual context and generates a continuous probability prior for suppressing contextually implausible detections. Unlike rigid binary masks constructed from external geographic data such as OpenStreetMap (OSM) [27], the learned scene prior does not require precise map-to-image registration and provides soft contextual constraints that reduce the risk of discarding valid targets near uncertain or transitional regions. 3.4.1. Formulation of the Network Topology SPGF reuses the temporally aggregated reference background Bt constructed in MAFF as its scene-context input. Since Bt is obtained from spatially and photometrically aligned historical frames through temporal median filtering, transient foreground responses are attenuated while relatively stable scene structures are retained. The network adopts a fully convolutional residual encoder–decoder architecture to extract contextual features relevant to vehicle accessibility and motion activity. The encoder progressively captures multi-scale representations, while an atrous spatial pyramid enlarges the effective receptive field at the bottleneck [28]. The decoder then progressively restores spatial resolution through bilinear upsampling and skip connections. For compact notation, let CGSnk,r(·) denote a k×k convolution with n output channels and dilation rate r, followed by Group Normalization (GN) [29] and Swish activation [30]. Moreover, Ψnδ(·) denotes a residual block with n output channels and stride δ, where a 1×1 projection is applied to the shortcut when the feature dimensions change. The number of output channels at the l-th encoder stage is defined as nl=24×2l−1, l=1,…,4, with δ1=1 and δl=2 for l=2,…,4. The complete network is formulated recursively as(22)E0=CGSn13,1(Bt),El=Ψnlδl(El−1),l=1,…,4,Z=CGSn41,1[CGSn31,1(E4),CGSn33,2(E4),CGSn33,4(E4),CGSn33,8(E4)],U4=Z,Ul=CGSnl3,1([Up2(Ul+1)El]),l=3,2,1,[Pacc,Pmot]=Sigmoid(Conv21×1(U1)). where El and Ul denote the encoder and decoder features at the l-th scale, respectively, Z denotes the bottleneck feature obtained through multi-scale atrous aggregation, and Up2(·) denotes bilinear upsampling by a factor of two. At the bottleneck, parallel convolutions with dilation rates {2,4,8} aggregate broader scene context while preserving local structural information. The decoder progressively fuses the bottleneck representation with encoder features and produces two complementary probability maps: Pacc characterizes vehicle-accessibility compatibility, whereas Pmot characterizes vehicle-motion compatibility. 3.4.2. Network Optimization Objective The supervision maps are constructed from temporally aggregated vehicle trajectories rather than from a coarse binary scene mask. For each training anchor, trajectory observations from the preceding 20 frames are aligned to the current reference frame. The accessibility branch uses all valid trajectory points, whereas the motion branch uses only trajectory points associated with displacements of at least 0.8 m. Each trajectory point is represented by a localized Gaussian response with a radius of 10 pixels and a standard deviation of 5 pixels at the native image resolution, yielding two soft supervision maps for vehicle accessibility and motion activity, respectively. The two prediction branches are jointly optimized using spatially weighted variants of focal loss [31] and Dice loss [32]. Specifically, Lfocalacc and Lfocalmot emphasize hard-to-classify pixels and reduce the dominance of easy background samples in the vehicle-accessibility and vehicle-motion branches, respectively. Meanwhile, Ldiceacc and Ldicemot encourage spatial consistency between the predicted and ground-truth prior maps. The overall objective is formulated as:(23)LSPGF=Lfocalacc+0.5Ldiceacc+0.7Lfocalmot+0.5Ldicemot. The coefficient 0.5 balances the region-level Dice constraint with the pixel-wise focal objective, while the coefficient 0.7 balances the contribution of motion-activity supervision relative to vehicle-accessibility supervision. 3.4.3. Prior Refinement and Filtering During inference, the two complementary predictions are first fused through their geometric mean:(24)Pscene=Pacc⊙Pmot, where ⊙ denotes element-wise multiplication. This fusion assigns high prior confidence to regions jointly supported by vehicle-accessibility and vehicle-motion cues. Although Pscene provides a continuous scene-level constraint, isolated responses and short discontinuities may remain in the predicted map. We therefore apply a lightweight topology-preserving refinement Rtopo(·), consisting of Gaussian smoothing, weak- and strong-confidence support extraction, multi-directional morphological closing, connected-component filtering according to the presence of strong-confidence support, area, and elongation, and soft attenuation of unreliable responses. The overall refinement process is compactly formulated as:(25)Pprior=RtopoPscene. This refinement bridges short discontinuities and suppresses isolated or topologically inconsistent responses. Unlike binary scene masks, the refinement retains continuous confidence values, allowing uncertain contextual evidence near structural boundaries to be preserved rather than converted into hard binary decisions. The resulting prior Pprior is finally applied to the target set DS produced by SA-TD: (26)Dfinal=p∈DS|Pprior(p)≥τprior. where τprior denotes the scene-prior confidence threshold. Thus, SPGF acts as a soft contextual filter rather than an independent target detector, suppressing detections inconsistent with the inferred scene context while preserving the fine-grained localization produced by SA-TD. The pseudocode of the proposed SPGF mechanism is summarized in Algorithm 3. Algorithm 3: Scene-Prior Guided Filtering Mechanism 4. Experimental Setup 4.1. Datasets and Evaluation Metrics We evaluate MDF-Det on the WPAFB 2009 dataset [8], a standard WAMI benchmark acquired at 1.25 Hz using a camera system composed of six optical sensors. The individual sensor images are spatially stitched into gigapixel-scale mosaics covering an area of approximately 35 km2. The dataset contains 1025 frames and is divided into a 512-frame training sequence and a 513-frame test sequence. The binary classification and heatmap regression networks in SA-TD, together with the scene-prior network in SPGF, are trained exclusively using samples constructed from the training sequence, while the test sequence is reserved solely for final performance evaluation. The dataset presents substantial challenges for moving-target detection, including frame resolutions exceeding 315 megapixels, dense traffic with approximately 2000 vehicles per frame, and extremely small targets with an average size of approximately 9×18 pixels. Following the commonly adopted evaluation protocol [11], quantitative evaluation is conducted on six representative AOIs, namely AOIs 01, 02, 03, 34, 40, and 41. For qualitative cross-scene evaluation of SPGF, representative scenes from the Greene 2007 dataset [33] are additionally used. Greene 2007 is used exclusively for zero-shot inference and is not involved in training, validation, model selection, or hyperparameter tuning. Detection performance is quantitatively assessed using Recall, Precision, and the F1 score [11]. A predicted target location is considered a true positive if it is within 20 pixels of the corresponding ground-truth location. Computational efficiency is further evaluated in terms of the number of trainable parameters, model size, FLOPs per frame, peak GPU memory, and inference time in seconds per frame. Runtime contribution is additionally reported to characterize the relative computational cost of each module. 4.2. Baseline Methods and Implementation Details To evaluate the proposed approach, MDF-Det is compared with five representative baseline methods, including ClusterNet [11], the CNN-based method of Zhou et al. [15], the heatmap-regression-based HMRN [17] and HM-Net [6], and CATLoss [34]. Since official implementations directly applicable to our experimental setting are unavailable for these methods, all baselines are independently reimplemented according to the architectures, training procedures, and parameter settings described in their original publications. The principal implementation and training settings of the reproduced baselines are summarized in Table 1. ClusterNet, HMRN, HM-Net, and CATLoss are trained from random initialization, whereas CNNs are initialized from the publicly released weights of the reference implementation and subsequently retrained on the adopted training partition. To ensure a fair comparison, all reproduced baselines and MDF-Det use the same training, validation, and test data partitions, as well as the same evaluation protocol. Model selection and hyperparameter tuning are performed exclusively on the validation set, after which the selected configurations are fixed and evaluated on the test set without further adjustment. Therefore, all baseline results reported in this study are obtained from our reproduced implementations rather than directly adopted from the original publications. The MDF-Det pipeline is implemented using PyTorch and TensorFlow/Keras. All experiments are conducted on a workstation equipped with an Intel Core i9-14900K CPU and an NVIDIA RTX 3090 GPU. All methods are evaluated on the same six AOIs under an identical detection matching criterion and evaluation procedure. The hyperparameters of MDF-Det are set as follows. For MAFF, we use K=5, Tm=0.1, α=15, and τdiff=1. For SA-TD, we set τbin=0.7, Δτ=0.01, and τmin=0.17. For SPGF, the scene-prior confidence threshold is set to τprior=0.2. These hyperparameters are selected using the temporally separated validation set and are subsequently fixed for all test-set evaluations unless otherwise specified. 4.3. Network Training Details The binary classification network used for candidate screening in SA-TD follows the network architecture of Zhou et al. [15], but all model parameters are retrained in this study rather than using the pretrained weights from the original method. The binary classification, heatmap regression, and SPGF networks are all trained and validated using temporally separated and non-overlapping intervals constructed exclusively from the 512-frame training sequence: frames 100–379 are used for training, whereas frames 400–499 are reserved for validation. No random frame-level splitting is performed, thereby reducing temporal correlation between adjacent WAMI frames across the training and validation subsets. The 513-frame test sequence is reserved exclusively for final evaluation and is not involved in model training, validation, model selection, or hyperparameter tuning. For the binary classifier, 3,600,000 training samples and 36,000 validation samples are constructed as four-frame spatiotemporal patches of size 4×21×21 from candidate locations in the corresponding training and validation intervals. For the heatmap regression network in SA-TD, 2,824,620 training samples and 696,090 validation samples are constructed as four-frame spatiotemporal patches of size 4×45×45, with each target label represented by a 15×15 heatmap. A mean spatiotemporal patch, estimated from at most 50,000 training samples, is subtracted from the inputs for normalization. The heatmap regression network is trained for 100 epochs using Adam with an initial learning rate of 1×10−4, a batch size of 256, and the mean squared error loss. For SPGF, 5133 training samples and 1766 validation samples are independently constructed from the same temporally separated intervals using a sliding-window strategy on the original-resolution temporal reference background Bt. Specifically, 384×384 patches and their corresponding trajectory-supervised probability maps are extracted without resizing the full image to a fixed low resolution. For each anchor frame, up to 96 patches are sampled using a 40:30:30 ratio of trajectory-centered positive samples, high-texture hard negatives, and random negatives, respectively. Negative samples are maintained at least 48 pixels from annotated trajectory points, while weak negative regions are assigned a reduced loss weight of 0.03 to account for the incompleteness of trajectory-based supervision. To prevent spatial information leakage, all trajectory annotations and image patches overlapping the six evaluation AOIs are excluded together with an additional 64-pixel safety margin. The resulting spatial partition of the training, validation, and evaluation regions is illustrated in Figure 5. During inference, the same 384×384 sliding-window strategy is applied to the full-resolution reference background, and predictions from overlapping windows are merged to reconstruct the scene-prior probability maps. For each fixed AOI, the full-resolution scene-prior map is constructed once before online frame-by-frame detection and is subsequently reused for candidate filtering. The SPGF network is trained for a maximum of 80 epochs using Adam with an initial learning rate of 2×10−4 and a batch size of 4, using the combined focal and Dice losses described above. 5. Results 5.1. Illustrative Operation To qualitatively illustrate the operation of MDF-Det, Figure 6 presents representative intermediate results and final detections across four WAMI scenarios. Figure 6a shows the overall detection result on a full-scale scene, demonstrating that MDF-Det can maintain reliable target responses under an extremely low target-to-background ratio. The subsequent examples further illustrate the roles of the three cascaded modules: (1) MAFF enhances weak target responses by jointly exploiting motion and appearance cues, particularly for targets affected by shadows or weak motion contrast (Figure 6b); (2) SA-TD separates merged responses in densely distributed target regions, improving the discrimination of closely spaced targets (Figure 6c); and (3) SPGF suppresses false responses caused by registration residuals and structured background interference using the inferred scene-prior probability map (Figure 6d). The close correspondence between the predicted target locations (red circles) and the ground-truth annotations (yellow dots) further illustrates the ability of the proposed framework to localize moving targets while suppressing background-induced false alarms. To evaluate the cross-scene generalization of SPGF, Figure 7 presents scene-prior probability maps on the six WPAFB 2009 evaluation AOIs and representative scenes from the Greene 2007 dataset. On WPAFB 2009, SPGF produces higher responses on vehicle-compatible structures such as roads, intersections, and parking areas, while assigning lower probabilities to most non-traffic regions. The model trained on WPAFB 2009 is further transferred directly to Greene 2007 without retraining, fine-tuning, or hyperparameter adjustment, where similar scene-aware response patterns are preserved across diverse environments. These results provide qualitative evidence that SPGF captures transferable scene-level compatibility rather than relying solely on scene-specific target distributions. Moreover, the inferred priors remain spatially selective: across the six WPAFB AOIs, the fraction of pixels with prior probability above 0.15 ranges from 23.40% to 48.41%, while only 12.02–24.90% of pixels exceed a probability of 0.50. 5.2. Comparison with Baseline Methods Table 2 compares MDF-Det with five representative baseline methods across the six evaluation AOIs. MDF-Det achieves the highest F1 score on all six AOIs, with an average performance of 0.878±0.043, compared with 0.832±0.071 for the strongest baseline, HM-Net. The corresponding mean F1 improvement is 0.046 in absolute terms and 5.5% in relative terms. When compared with the best competing method on each individual AOI, MDF-Det yields absolute F1 gains ranging from 0.013 to 0.048. The largest improvement is observed on AOI 03, where the F1 score increases from 0.840 to 0.888. Improvements are also observed on AOIs 01 and 40, where the best baseline scores of 0.805 and 0.792 increase to 0.848 and 0.835, respectively. Overall, the results indicate that the performance improvement of MDF-Det is maintained across all six evaluated regions, while its cross-AOI standard deviation remains relatively low. The reported mean and standard deviation characterize cross-region variability but do not directly quantify the uncertainty of the paired improvement over the strongest baseline. To further assess this uncertainty while accounting for temporal dependence between adjacent WAMI frames, we additionally perform a paired temporal-block bootstrap analysis between MDF-Det and HM-Net on the test sequence [35]. Under this additional temporal-block analysis, 10,000 paired bootstrap replicates are generated using non-overlapping 20-frame blocks, yielding an estimated paired F1 improvement of 0.0828 with a 95% percentile confidence interval of [0.0766,0.0892]. Sensitivity analyses using 10- and 30-frame blocks yield confidence intervals of [0.0775,0.0874] and [0.0763,0.0887], respectively, demonstrating that the estimated paired improvement remains consistent across reasonable temporal block lengths. 5.3. Ablation Study Table 3 evaluates the individual and complementary contributions of MAFF, SA-TD, and SPGF on AOI03. Among the individual modules, SA-TD provides the largest improvement, primarily by substantially increasing Recall, indicating its effectiveness in recovering and separating densely distributed target responses. In contrast, MAFF shifts the detection balance toward higher Precision, while SPGF improves Precision without reducing Recall, consistent with its role as a contextual false-alarm filter. More importantly, the modules exhibit clear complementarity: combining MAFF with SA-TD increases the F1 score from 0.801 to 0.883, and further introducing SPGF improves Precision from 0.838 to 0.847 while preserving Recall at 0.933, yielding the best overall F1 score of 0.888. These results indicate that MAFF improves candidate quality, SA-TD enhances target separability, and SPGF provides an additional scene-level constraint for false-alarm suppression. Table 4 further examines the internal contributions of spatial attention and threshold degradation (TD) in SA-TD. When TD is disabled, a simple greedy peak-selection strategy is used instead, which iteratively selects the maximum response from the heatmap, removes the selected response, and retains the top 20 candidate locations. The results show that spatial attention provides the dominant improvement by markedly increasing Precision while maintaining high Recall, whereas TD further suppresses ambiguous heatmap responses and improves the Recall–Precision balance. Their combination achieves the highest F1 score of 0.888, confirming the complementary roles of feature enhancement and adaptive target decoupling within SA-TD. 5.4. Hyperparameter Analysis The main hyperparameters of MAFF, SA-TD, and SPGF are analyzed on the temporally separated validation set to characterize their effects on the Recall–Precision trade-off and overall detection behavior. The selected settings are then fixed for all subsequent test-set inference without further adjustment. (1) MAFF benefits from moderate motion enhancement, while excessive enhancement introduces additional background responses. As shown in Figure 8a, increasing α initially improves Recall, whereas larger values reduce Precision and F1. Increasing τdiff further shifts the operating point toward higher Precision and lower Recall. The F1 score peaks at 0.917 when α=15 and τdiff=1, and these values are therefore selected for subsequent test-set inference. (2) SA-TD becomes increasingly precision-oriented as the screening thresholds increase, with intermediate settings remaining the most stable. Figure 8b shows that increasing τbin or τmin generally reduces Recall while improving Precision, producing a unimodal F1 trend. A stable high-performance region appears around τbin=0.7 and τmin=0.15–0.17, with the maximum F1 of 0.918 obtained at τbin=0.7 and τmin=0.17; these values are therefore selected. (3) SPGF progressively favors Precision as scene-prior filtering becomes stricter, while excessive filtering reduces overall performance. As shown in Table 5, increasing τprior consistently lowers Recall and improves Precision, causing F1 to first increase and then decline. The highest F1 of 0.922 is obtained at τprior=0.2 for both τbin=0.7 and 0.9. Since τbin=0.7 retains higher Recall, τprior=0.2 and τbin=0.7 are selected for subsequent test-set inference. 5.5. Sensitivity to Localization Radius To examine the influence of the localization criterion, we further evaluate MDF-Det using matching radii ranging from 5 to 25 pixels. All results are obtained from the same fixed detection outputs, with only the localization radius used for one-to-one matching being changed. Each ground-truth target is matched to at most one detection; unmatched detections and ground-truth targets are counted as false positives and false negatives, respectively. As shown in Table 6, detection performance improves gradually as the localization radius increases, while the improvement becomes marginal at larger radii. Increasing the radius from 10 to 20 pixels improves the F1 score by only 0.005, whereas a further increase to 25 pixels yields an additional gain of only 0.003. This relatively small variation indicates that MDF-Det is not highly sensitive to the localization radius and that the reported performance is not primarily attributed to the adopted 20-pixel matching criterion. 5.6. Computational Complexity and Runtime Analysis Table 7 summarizes the computational complexity and runtime distribution of MDF-Det. The complete framework contains 10.794 M parameters, while its per-frame online neural-network inference requires 490.069 G FLOPs, with an overall runtime of 3.161 s/frame. FLOPs are reported for per-frame online neural-network inference: MAFF mainly consists of non-neural image-processing operations, whereas SPGF prior construction is performed once for each fixed AOI and is therefore excluded from the per-frame FLOP count. The binary classifier is the main computational bottleneck, accounting for 51.28% of the total runtime, mainly because it processes a large number of motion-generated candidate patches. In contrast, although SA-TD contains more parameters, it is applied only to a small subset of retained candidates, resulting in substantially lower FLOPs and runtime. MAFF introduces no trainable parameters, with its computational cost mainly arising from multi-frame registration, motion compensation, optical-flow fusion, and background subtraction. SPGF incurs negligible online overhead because its scene prior is constructed only once for each fixed AOI, while online processing involves only prior lookup and candidate filtering. With a runtime of 3.161 s/frame, the current implementation does not yet meet the real-time processing requirement of WAMI systems operating at approximately 1–2 fps. Dense optical flow is currently executed on the CPU because the adopted OpenCV build does not provide CUDA optical-flow support, while the binary classifier remains the dominant computational bottleneck. Further acceleration could therefore be achieved through GPU-based optical-flow computation and more efficient candidate screening. 6. Discussion The experimental results demonstrate that MDF-Det effectively addresses several key challenges in WAMI target detection. By integrating dense optical flow with grayscale differencing, MAFF enhances weak motion responses and improves the quality of foreground candidates. SA-TD further separates merged responses in densely populated regions through fine-grained heatmap decoding and step-threshold degradation, while SPGF introduces soft scene-level constraints to suppress false alarms caused by registration artifacts, vegetation motion, and other contextually implausible background responses. The ablation results confirm that the three components perform distinct yet complementary functions: MAFF shifts the detection balance toward higher Precision, SA-TD provides the largest gain in Recall and improves target separability, and SPGF further improves Precision through scene-level filtering. Despite these advantages, MAFF remains dependent on the quality of inter-frame registration and optical-flow estimation. Strong parallax, abrupt illumination changes, or rapid camera motion may introduce residual misalignment or unreliable motion responses, thereby degrading candidate generation. MDF-Det also inherits a common limitation of candidate-driven cascaded pipelines. Since SA-TD and SPGF operate only on candidates retained by MAFF, weak targets completely missed during the initial candidate-generation stage cannot be recovered in subsequent refinement. Although the zero-shot transfer results indicate cross-scene transferability, SPGF is learned from trajectory-derived supervision within the training sequence and may therefore still exhibit some dependence on scene-specific spatial statistics. A promising future direction is therefore to introduce a semantic feedback branch that complements the motion-based front end. By integrating motion confidence, foundation-model semantics, and soft road-network priors through uncertainty-aware gating, such a branch could reassess low-confidence regions and generate supplementary candidates in areas with weak motion responses but strong contextual evidence. This would provide a practical way to improve weak-target recovery while reducing the dependence of SPGF on scene-specific spatial statistics. 7. Conclusions In this paper, we propose MDF-Det, a coarse-to-fine framework for tiny moving target detection in WAMI. By sequentially integrating motion–appearance fusion, spatial attention-guided target decoupling, and scene-prior filtering, MDF-Det effectively mitigates missed detections caused by weak motion responses, merged detections in dense scenes, and scene-induced false alarms. Experimental results on six evaluation AOIs of the WPAFB 2009 dataset demonstrate that MDF-Det achieves an average F1 score of 0.878, corresponding to a relative improvement of 5.5% over the strongest baseline. These results demonstrate the effectiveness of MDF-Det across the six evaluated regions of the WPAFB benchmark. Future work will focus on improving the cross-scene generalization of scene-prior modeling and reducing the dependence of the cascaded framework on the initial candidate-generation stage. We also plan to investigate semantic feedback mechanisms that integrate motion confidence, foundation-model semantics, and soft road-network priors, and to further validate MDF-Det on additional WAMI datasets and more diverse aerial imaging scenarios.