TriPath3DNet: an efficient real-time model for multi-class classification in real-life surveillance videos of fixed duration Authors: K Mohanarangan, P Palanisamy Published: 2026 ABSTRACT AbstractThis paper presents TriPath3DNet, a novel, efficient, and interpretable 3D CNN architecture designed for real-time, multi-class classification of short, motion-triggered surveillance video clips under challenging real-world conditions–including occlusion, variable lighting, adverse weather, and subtle low-motion anomalies such as loitering or zone intrusion. TriPath3DNet integrates three complementary temporal pathways–short-term motion, long-term context, and Temporal Difference Encoding (TDE)–into a lightweight ResNet3D (R3D)-18 backbone to jointly model transient dynamics and sustained activities. Evaluated on four datasets–including the newly curated Virat1-RC, Virat2-RC, UCF-Crime, and our proprietary In-House Dataset (IHD)–TriPath3DNet achieves state-of-the-art or near state-of-the-art performance, with up to 95.37% accuracy, 99.42% AUC, and an inference latency of 129 to 137 ms per 50-frame clip (approx 2.6 ms per frame) on an 11 GB GPU, using only 33.46 M parameters. Notably, it outperforms both CNN- and transformer-based baselines–including MViTv1, MViTv2, and VideoSwin–by significant margins, especially on anomaly-dense benchmarks like UCF-Crime, where most vision transformers struggle. While MViTv2 achieves slightly higher accuracy on IHD (91.87% vs. 89.46%), TriPath3DNet delivers substantially better AUC (98.07% vs. 90.96%), indicating superior calibration for critical anomaly detection. Ablation studies confirm that each temporal branch contributes meaningfully to performance, and Grad-CAM visualizations demonstrate spatially precise and temporally coherent attention maps. By aligning architectural design with edge–cloud deployment constraints and the operational realities of industrial surveillance, our work bridges the gap between academic research and real-world video analytics.Subject terms: Engineering, Mathematics and computing Introduction Modern closed-circuit television (CCTV) surveillance demands minimal human intervention, making manual monitoring of hundreds of camera feeds impractical. Relying solely on human operators leads to fatigue, staffing challenges, and high operational costs. In contrast, AI assisted analytics enable real-time, scalable monitoring, allowing proactive incident prevention rather than post-mortem analysis1,2. AI-driven video analytics can simultaneously process multiple feeds, enhancing precision while reducing manpower and costs. High resolution and bitrate are crucial for detecting long-range objects and individuals up to 150 feet away. However, continuous analysis of high-frame-rate (5–30 fps) video streams using deep-learning models is computationally expensive and inefficient, as most frames contain no significant activity. Processing every frame can overload AI systems, introducing latency and degrading performance, particularly in time-sensitive events like intrusion, theft, or vandalism. To overcome this, edge–cloud surveillance frameworks3 integrate lightweight motion detection at the edge, using frame differencing4. When notable motion is detected, the system triggers an event, prompting the AI model to analyze a short 50-frame events. Two common approaches in AI-based video analytics are object detection and video-level activity classification. Object detection operates frame-by-frame, focusing on spatial features such as color and shape–typified by the YOLO family of models5. However, these models often neglect temporal context, limiting their effectiveness for motion-centric surveillance tasks. Even with motion masks fused to YOLO detectors6, issues persist: static intruders or low-motion behaviors may go undetected, and low-contrast or partially occluded subjects can be missed in dim or cluttered environments. Conversely, small pixel variations (e.g., light reflections on stationary vehicles) can cause false alarms, reducing reliability. Surveillance anomalies typically occur as short, isolated events during off-hours, such as unauthorized entry or unexpected vehicle movement. Leveraging this, our system employs a single-activity, multi-class video classification framework7,8 to detect threats precisely and efficiently. At its core lies the proposed TriPath3DNet, a novel 3D CNN that fuses spatial and temporal information to enhance classification accuracy and contextual understanding of real-world surveillance. This holistic approach enables intelligent, human-like interpretation of activities for robust, real-time monitoring. Fig. 1 illustrates the edge–cloud surveillance pipeline designed for scalable, real-time analytics across distributed CCTV installations. High-resolution cameras at key locations transmit streams to an edge node running lightweight motion detection. Only motion-triggered event clips (50 frames) are forwarded to the cloud AI engine hosting the TriPath3DNet classifier. The results are then sent to a centralized alert dashboard for operators or integrated security systems. This hybrid setup minimizes bandwidth, reduces inference cost, and enables low-latency analytics9–11, making it ideal for large-scale deployments. AI-assisted surveillance system with edge–cloud video classification architecture. Contributions: (i)Propose TriPath3DNet, a lightweight multi-branch 3D CNN that explicitly fuses short-term, long-term, and temporal-difference (TDE) motion cues within a ResNet3D-18 backbone for real-time video classification under challenging surveillance conditions. Propose TriPath3DNet, a lightweight multi-branch 3D CNN that explicitly fuses short-term, long-term, and temporal-difference (TDE) motion cues within a ResNet3D-18 backbone for real-time video classification under challenging surveillance conditions. (ii)Curate and restructure public datasets–Virat12 and UCF-Crime13 into temporally coherent, fixed-length benchmarks (Virat1-RC, Virat2-RC, and UCF-Crime), and compile a proprietary In-House Dataset (IHD) captured across diverse real-world industrial sites with varying lighting, occlusion, and weather. Curate and restructure public datasets–Virat12 and UCF-Crime13 into temporally coherent, fixed-length benchmarks (Virat1-RC, Virat2-RC, and UCF-Crime), and compile a proprietary In-House Dataset (IHD) captured across diverse real-world industrial sites with varying lighting, occlusion, and weather. (iii)Demonstrate near state-of-the-art performance across four datasets in accuracy, AUC, and inference speed (129–137 ms/clip), significantly outperforming both CNN- and transformer-based baselines while maintaining a low computational footprint (33.46M parameters). Demonstrate near state-of-the-art performance across four datasets in accuracy, AUC, and inference speed (129–137 ms/clip), significantly outperforming both CNN- and transformer-based baselines while maintaining a low computational footprint (33.46M parameters). (iv)Validate the architectural design through extensive ablation studies, Grad-CAM interpretability analysis, and efficiency benchmarks, confirming the necessity of multi-scale temporal modeling for robust anomaly-sensitive surveillance. Validate the architectural design through extensive ablation studies, Grad-CAM interpretability analysis, and efficiency benchmarks, confirming the necessity of multi-scale temporal modeling for robust anomaly-sensitive surveillance. The paper is organized as follows: Section 2 reviews related works; Section 3 details the proposed model; Section 4 outlines the experimental setup; Section 5 presents results and discussion; and Section 6 concludes. Supplementary Materials include extended results, illustrations, and for reproducibility.1 Related works Recent developments in Video Anomaly Detection (VAD) span weak supervision, self-supervised learning, CLIP/prompt-guided designs, transformer-based architectures, federated learning, and hybrid CNN–RNN pipelines. Since most real-world surveillance scenarios fall under weakly supervised settings, we focus on works published from 2021 onward that address this regime. (i)Weakly Supervised MIL and Self-Supervised FrameworksLv et al14. proposed UMIL, which partitions snippets into confident and ambiguous sets and applies pairwise supervision to reduce context bias, improving temporal localization and yielding strong results on UCF-Crime and TAD. Ullah et al15. combined BoTNet-152 with a dilated TCN and multi-head temporal attention. Their method improves performance on five datasets, including +1.9% AUC on LAD-2000 and +1.5% on Crowd Violence. Karim et al16. introduced REWARD, a real-time weakly supervised framework using self-supervised kNN learning with uniform frame sampling, achieving 63 FPS and outperforming MIL baselines. Zhou et al17. proposed BN-WVAD, which defines a Divergence-from-Mean anomaly score and a Mean-based Pull–Push loss. Their Sample–Batch Selection strategy improves robustness to label noise. Weakly Supervised MIL and Self-Supervised FrameworksLv et al14. proposed UMIL, which partitions snippets into confident and ambiguous sets and applies pairwise supervision to reduce context bias, improving temporal localization and yielding strong results on UCF-Crime and TAD. Ullah et al15. combined BoTNet-152 with a dilated TCN and multi-head temporal attention. Their method improves performance on five datasets, including +1.9% AUC on LAD-2000 and +1.5% on Crowd Violence. Karim et al16. introduced REWARD, a real-time weakly supervised framework using self-supervised kNN learning with uniform frame sampling, achieving 63 FPS and outperforming MIL baselines. Zhou et al17. proposed BN-WVAD, which defines a Divergence-from-Mean anomaly score and a Mean-based Pull–Push loss. Their Sample–Batch Selection strategy improves robustness to label noise. (ii)CLIP-Based and Prompt-Guided DesignsZhang et al18. enhanced pseudo-labeling via completeness and uncertainty modeling, yielding strong results across UCF-Crime, TAD, and XD-Violence. Wu et al19. proposed STPrompt, a CLIP-based two-stream model with training-free spatial localization using vision–language prompts. Yang et al20. introduced TPWNG, combining dual ranking losses, learnable text prompts, and adaptive temporal modeling, achieving SOTA performance. Ye et al21. presented VERA, an explainable VAD framework adapting frozen vision–language models via verbalized learning, achieving 86.55% AUC on UCF-Crime. Wang et al22. proposed a federated weakly supervised framework integrating multimodal prompts for privacy-preserving anomaly detection. Xu et al23. introduced PLOVAD, an open-vocabulary weakly supervised framework using prompt tuning with a GAT-based temporal module, enabling detection of unseen anomalies. CLIP-Based and Prompt-Guided DesignsZhang et al18. enhanced pseudo-labeling via completeness and uncertainty modeling, yielding strong results across UCF-Crime, TAD, and XD-Violence. Wu et al19. proposed STPrompt, a CLIP-based two-stream model with training-free spatial localization using vision–language prompts. Yang et al20. introduced TPWNG, combining dual ranking losses, learnable text prompts, and adaptive temporal modeling, achieving SOTA performance. Ye et al21. presented VERA, an explainable VAD framework adapting frozen vision–language models via verbalized learning, achieving 86.55% AUC on UCF-Crime. Wang et al22. proposed a federated weakly supervised framework integrating multimodal prompts for privacy-preserving anomaly detection. Xu et al23. introduced PLOVAD, an open-vocabulary weakly supervised framework using prompt tuning with a GAT-based temporal module, enabling detection of unseen anomalies. (iii)Transformer-Based Temporal AggregationHuang et al24. proposed WSTD, combining transformer-style temporal aggregation with discriminative encoding. Shao et al25. presented EventVAD, a training-free event-aware framework combining spatio-temporal graphs, event boundary detection, and hierarchical prompting with multimodal LLMs. Zhou et al26. introduced MA-PDM, a motion- and appearance-guided patch diffusion framework for fine-grained anomaly detection. Yang et al27. proposed a self-supervised framework generating pseudo-anomalous samples via geometric transformations and temporal scrambling. Zahid et al28. developed a few-shot scene-adaptive pipeline integrating diffusion and GAN-based prediction under meta-learning. Su et al29. proposed a federated weakly supervised framework with a mixture of local-to-global experts, achieving strong results under domain shift. Transformer-Based Temporal AggregationHuang et al24. proposed WSTD, combining transformer-style temporal aggregation with discriminative encoding. Shao et al25. presented EventVAD, a training-free event-aware framework combining spatio-temporal graphs, event boundary detection, and hierarchical prompting with multimodal LLMs. Zhou et al26. introduced MA-PDM, a motion- and appearance-guided patch diffusion framework for fine-grained anomaly detection. Yang et al27. proposed a self-supervised framework generating pseudo-anomalous samples via geometric transformations and temporal scrambling. Zahid et al28. developed a few-shot scene-adaptive pipeline integrating diffusion and GAN-based prediction under meta-learning. Su et al29. proposed a federated weakly supervised framework with a mixture of local-to-global experts, achieving strong results under domain shift. (iv)CNN–RNN and Hybrid Architectures Feichtenhofer30introduced X3D, an efficient scalable 3D CNN architecture. Ayhan et al31. proposed early VideoGraph-based temporal relation modeling for anomaly detection. Maqsood et al32. used CNN-based weak supervision with limited scalability. Zhou et al33. developed a trajectory-based detector for stealing events. Joemon et al34. combined CNN and LSTM for loitering detection. Hao et al35. integrated C3D features with GCN and R3D classifiers. Rezaee et al36. applied CNN–RNN hybrids for crowd anomaly detection. Ullah et al37. presented an edge–cloud collaborative framework balancing efficiency and accuracy. Hendi et al38. combined CNNs and GRUs for multi-class events. Li et al39. proposed STDRN-HA with hierarchical spatiotemporal attention. Ahmed et al40. employed a C3D–autoencoder–GAN cascade. Jeon et al41. developed a CNN-based system for loitering and abandoned-object detection. Kumar et al42. combined InceptionResNetV3 with Bi-LSTM and attention. Mumtaz et al43. designed a two-stage CNN pipeline with high computational cost. Yousaf et al44. proposed a CNN–BiLSTM model for inappropriate-content detection. Kaur et al45. proposed a bagging ensemble with stacked LSTMs for violence detection. Majhi et al46. introduced GS-MoE, a Gaussian-splatter guided mixture-of-experts for adaptive weakly supervised detection. Elmetwally et al47. presented a real-time CNN reconstruction-based VAD system emphasizing deployment feasibility. Dabouei et al48. proposed a lightweight CLIP+Transformer framework tailored for laboratory automation anomaly detection. CNN–RNN and Hybrid Architectures Feichtenhofer30introduced X3D, an efficient scalable 3D CNN architecture. Ayhan et al31. proposed early VideoGraph-based temporal relation modeling for anomaly detection. Maqsood et al32. used CNN-based weak supervision with limited scalability. Zhou et al33. developed a trajectory-based detector for stealing events. Joemon et al34. combined CNN and LSTM for loitering detection. Hao et al35. integrated C3D features with GCN and R3D classifiers. Rezaee et al36. applied CNN–RNN hybrids for crowd anomaly detection. Ullah et al37. presented an edge–cloud collaborative framework balancing efficiency and accuracy. Hendi et al38. combined CNNs and GRUs for multi-class events. Li et al39. proposed STDRN-HA with hierarchical spatiotemporal attention. Ahmed et al40. employed a C3D–autoencoder–GAN cascade. Jeon et al41. developed a CNN-based system for loitering and abandoned-object detection. Kumar et al42. combined InceptionResNetV3 with Bi-LSTM and attention. Mumtaz et al43. designed a two-stage CNN pipeline with high computational cost. Yousaf et al44. proposed a CNN–BiLSTM model for inappropriate-content detection. Kaur et al45. proposed a bagging ensemble with stacked LSTMs for violence detection. Majhi et al46. introduced GS-MoE, a Gaussian-splatter guided mixture-of-experts for adaptive weakly supervised detection. Elmetwally et al47. presented a real-time CNN reconstruction-based VAD system emphasizing deployment feasibility. Dabouei et al48. proposed a lightweight CLIP+Transformer framework tailored for laboratory automation anomaly detection. Despite these advances, several critical gaps remain in current Video Anomaly Detection (VAD) research: (i) Most frameworks, including recent MIL, CLIP-based, and transformer designs, assume access to full or variable-length videos, which is unsuitable for real-time surveillance where anomalies often occur in short, motion-triggered clips; few methods explicitly address fixed-duration formats (e.g., 50 frames at 5 fps) required to trigger preventive action within minutes of incident onset. (ii) Stationary or subtle anomalies such as loitering, zone intrusion, or perimeter breach remain difficult to detect, and even advanced transformer-based and federated approaches struggle when subjects occupy small spatial footprints or appear under poor lighting, occlusion, or cluttered backgrounds. (iii) While recent works such as MA-PDM, self-supervised geometric scrambling, and GS-MoE explore fine-grained cues, lightweight multi-scale modeling across both spatial and temporal dimensions is rarely achieved, leaving variations in object scale (near vs. far) and activity duration (sudden trespass vs. prolonged lingering) underexplored in efficient architectures. (iv) No existing work leverages event-specific manual annotation that reflects real surveillance scenarios, where anomalies such as theft or intrusion occur as isolated activities during off-hours; current datasets like UCF-Crime, Virat, and ShanghaiTech contain overlapping, untrimmed, or multi-activity events, which poorly represent the clean, single-anomaly structure of industrial footage. (v) Although CLIP-based methods such as VERA and PLOVAD introduce verbalized learning and prompt tuning for unseen anomalies, most frameworks still lack robust open-vocabulary detection and interpretable outputs, with explainability limited to detection scores rather than anomaly rationales. (vi) Emerging federated approaches such as those by Wang and Su demonstrate promise in collaborative anomaly detection under domain shift, but scalability, communication efficiency, and robustness to heterogeneous client data remain open challenges for deployment in distributed surveillance networks. (vii) Few methods explicitly target industrial security anomalies such as restricted-zone intrusion, perimeter breach, or equipment tampering, with most focusing on generic violence or crowd datasets, leaving critical asset-protection scenarios underrepresented. (viii) Many recent works report performance only on high-end GPUs; while real-time CNN reconstruction (Elmetwally) and lightweight CLIP+Transformer designs (Dabouei) highlight deployment feasibility, overall inference efficiency and model complexity are frequently overlooked, hindering adoption on edge devices with limited compute and power budgets. In addition, although transformer-based architectures have recently gained attention in video understanding, their reliance on dense temporal attention and quadratic complexity limits their practicality in our experiments for short, low-motion, motion-triggered surveillance clips that require rapid decision-making. Our work directly addresses these limitations by proposing a lightweight, multi-branch architecture that enables efficient, multi-scale spatio-temporal analysis tailored for short, fixed-duration surveillance clips, while emphasizing explainability, edge deployment, and alignment with real-world asset-protection tasks. Materials and methods The proposed TriPath3DNet model builds upon the base R3D-18 architecture. As shown in Fig. 2, it enhances the standard model by introducing three parallel temporal branches to capture motion at multiple scales. These include a short-term branch for rapid movements, a long-term branch for extended temporal patterns, and a TDE branch to emphasize subtle frame-to-frame variations. The outputs of these branches are fused via a convolution and subsequently processed through the R3D-18 layers for spatiotemporal feature extraction and final classification. This integration of temporal branches significantly improves performance. The heatmaps are generated by averaging feature maps across channels for each frame and overlaying them on the input frames using perceptual colormaps. Regions with high activation (typically appearing in yellow or red) indicate areas the model attends to for motion understanding, while low-activation regions (blue or dark) are considered less informative. The short-term branch focuses on fine-grained, rapid movements, the long-term branch captures slow and consistent motion patterns, and the TDE branch highlights abrupt changes across frames. The fused heatmaps demonstrate how the model integrates multi-scale temporal cues into a coherent representation. TriPath3DNet classifier architecture. Input and base network: R3D-18 Given an input video clip, the model accepts a 5D tensor input represented as: where B is the batch size, C is the number of channels (typically 3 for RGB input), T denotes the temporal length (i.e., number of frames), and H × W are the spatial dimensions of each frame. The initial stage of the model involves a stem layer, adapted from the pretrained R3D-18 architecture, which is designed to perform early fusion of spatial and temporal information. The stem comprises a 3D convolutional layer with a kernel size of , stride of 2 in all dimensions, and padding of 3, followed by batch normalization, a ReLU activation function, and a max pooling layer with stride 2 and padding 1. This sequence of operations reduces the input resolution across spatial and temporal axes while capturing low-level motion cues and spatial patterns necessary for downstream processing. Formally, the output of the stem layer is defined as: where T', H', and W' are reduced temporal and spatial resolutions determined by the convolution and pooling operations. The output feature map X' encodes the fused spatiotemporal characteristics and serves as a shared feature backbone for subsequent branches in the proposed TriPath3DNet architecture. This early-stage processing is crucial for efficiently extracting motion patterns and appearance information from the video input, while also reducing the computational burden of later network stages. Multi-temporal motion-aware feature extractor To capture temporal patterns at multiple scales, the proposed TriPath3DNet architecture employs three parallel temporal streams: a short-term branch, a long-term branch, and a TDE Branch. Each stream operates on the shared feature map X' extracted by the R3D-18 stem. Short-term branch The short-term branch as shown in Fig. 3 captures fine-grained, transient motion by applying a temporal-only 3D convolution with kernel size (3, 1, 1) to the stem features : The kernel spans three consecutive frames in time but only a single pixel spatially, ensuring the operation responds exclusively to local temporal dynamics without spatial smoothing. Temporal padding preserves the original sequence length T. Short-term branch. Explicitly, for learnable weights and bias , the output at time t is: where denotes the temporal slice of the kernel at offset Δ, and X'[t] is zero-padded outside [1, T]. This design isolates rapid inter-frame changes–such as quick gestures or micro-movements–enhancing sensitivity to brief yet critical events while maintaining full spatial resolution and temporal fidelity, a key requirement for real-time anomaly detection in surveillance. Long-term branch The long-term branch as shown in Fig. 4 models extended temporal dynamics–such as slow walking, loitering, or vehicle parking–by applying a temporal-only 3D convolution with a wide kernel of size (15, 1, 1) to the stem features : The kernel spans 15 consecutive frames in time while covering only a single spatial pixel, ensuring the operation aggregates information across a broad temporal window without spatial smoothing. Temporal padding of (7, 0, 0) preserves the original sequence length T. Long-term branch. Let and denote the learnable weights and bias. With zero-padding outside [1, T], for all , , , , : This branch provides critical context for activities that evolve gradually over time, complementing the short-term branch’s sensitivity to rapid motion. By preserving spatial structure and temporal length, it enables seamless fusion with other pathways while maintaining computational efficiency. TDE branch To explicitly encode motion, the Temporal Difference Encoding (TDE) branch as shown in Fig. 5 computes the discrete temporal derivative of the stem features. Let denote the stem output. We define the frame-wise temporal difference as: where represents the change in feature values between consecutive frames. This signal–akin to a first-order approximation of optical flow–strongly responds to motion boundaries (e.g., object displacement, smoke, flicker) while attenuating static content. TDE branch. To restore the original temporal length T, we prepend a zero tensor along the time axis: where padding follows the PyTorch convention , adding one zero frame at t=1. A lightweight 3D convolution then refines this motion prior into a learnable representation: with kernel size (3, 1, 1) and padding (1, 0, 0) to preserve temporal dimensionality. By injecting an explicit Δ X'-based motion signal, the TDE branch complements the implicitly learned dynamics of the short- and long-term branches. This enhances robustness to subtle, rapid, or low-frame-rate anomalies–critical for real-world surveillance where motion cues are often sparse. Fusion and backbone propagation The outputs of the three temporal branches–short-term (), long-term (), and TDE ()–are fused via element-wise summation followed by a 3D convolution to align feature dimensions: This early fusion integrates complementary motion cues into a unified spatiotemporal representation, which is then processed by the R3D-18 backbone for hierarchical feature extraction. R3D-18 backbone with skip connections The fused tensor passes through four residual stages , with channel widths . Each stage consists of basic 3D residual blocks that apply two convolutions (each followed by batch normalization and ReLU) along the residual path F(· ) and add a skip connection : Skip connection. The skip branch is an identity mapping when input and output shapes match. Otherwise, a projection with batch normalization aligns dimensions: Downsampling schedule. Spatial resolution is halved at the first block of stages – using a stride of (1, 2, 2) (temporal stride = 1 to preserve clip length), while channel depth doubles: This design preserves fine-grained motion in early layers and builds high-level semantics in deeper stages. Residual connections stabilize training and enable efficient feature reuse. Pooling and classification After the final residual block, the feature tensor undergoes global average pooling (GAP) to produce a compact video-level descriptor: This aggregates spatiotemporal statistics without flattening, preserving global context. A fully connected (FC) layer then maps Z to class logits: where C is the number of activity classes. The output Y is used for softmax-based classification. Cross-entropy loss for multi-class classification Given model output logits , where is the logit score for class i, the predicted probabilities are computed using the softmax function: where C is the number of classes. Let denote the one-hot encoded ground truth vector, where for the true class k, and for all j ≠ k. The cross-entropy loss is defined as: Since only one component is non-zero, the expression simplifies to: where is the predicted probability for the ground truth class. The gradient of the loss with respect to each logit is given by: The model parameters W are then updated using gradient descent as: where η is the learning rate and is computed via backpropagation from . This formulation penalizes incorrect and overconfident predictions, encouraging the model to assign higher probability to the correct class. Over successive iterations, it enhances class separation by reinforcing correct predictions while suppressing incorrect ones. The architectural comparison of Tripath3DNet with baseline is summarized in the Table. 1. Table 1Architectural comparison: R3D-18 vs. TriPath3DNet.ComponentR3D-18TriPath3DNet (Ours)AdvantageTemporal ModelingSingle 3D convolution pathThree parallel branches: short-term (3, 1, 1), long-term (15, 1, 1), and TDECaptures multi-scale dynamics: fine gestures, sustained actions, and explicit motionMotion EncodingImplicit (via 3D conv)Explicit frame differencing (TDE) + learnable refinementEnhances sensitivity to subtle anomalies (e.g., loitering, tampering)Backbone UsageEnd-to-end fine-tuned (Kinetics)Frozen stem + task-specific branchesRetains general priors while adapting to surveillance-specific motionFeature FusionN/A (single stream)Early fusion: sum + conv before backboneEnables cross-branch interaction while preserving spatiotemporal resolution Architectural comparison: R3D-18 vs. TriPath3DNet. Code availability To ensure transparency, reproducibility, and long-term version control, the implementation of TriPath3DNet, along with the comparative models used in this study, has been archived in a DOI-minting repository (Zenodo). The specific version corresponding to the results reported in this manuscript is publicly available at: https://doi.org/10.5281/zenodo.18718931 The repository contains the complete model architecture, implementations of baseline methods used for comparison, training and evaluation scripts, and configuration files (including requirements.txt) required to reproduce the experiments presented in this work. Experimentation Datasets In this work, we curate Virat, and UCF-Crime as a methodological adaptation that aligns academic evaluation with real-world, edge–cloud surveillance. Specifically, (i) we enforce fixed-duration, motion-triggered inputs (e.g., 50 frames of action oriented content) instead of untrimmed, variable-length streams; (ii) we adopt a single-event, multi-class assumption by segmenting videos into 50-frame clips to mirror real-world alert workflows–such as intrusion, loitering, or vehicle entry–and thereby reduce label ambiguity; (iii) we restore temporal coherence and clear activity boundaries via motion-energy–based centering, ensuring each clip is anchored on its most dynamic segment; (iv) we rebalance classes toward threat-centric tasks, placing emphasis on low-motion yet high-risk events like zone intrusion or static trespass to enhance operational relevance; (v) we ensure computational feasibility under strict GPU memory and latency budgets by standardizing all inputs to 50-frame clips, enabling fair comparisons even against transformer-based state-of-the-art models. Collectively, these design choices constitute a core contribution of our work: reengineering public benchmarks to reflect the actual deployment setting of modern surveillance systems. (i)Virat: The Virat dataset12 provides high-resolution outdoor surveillance videos with detailed activity annotations. However, the original clips are daylight-staged and lack contextual variation. To enhance real-world relevance, we curated two variants–Virat1-RC and Virat2-RC–by segmenting videos into 50-frame clips and redefining classes around intrusion and threat-related activities. The restructuring ensures temporal consistency and domain alignment (Fig. 6). Class-level statistics and descriptions are provided in Tables 2 and 12.2 Virat: The Virat dataset12 provides high-resolution outdoor surveillance videos with detailed activity annotations. However, the original clips are daylight-staged and lack contextual variation. To enhance real-world relevance, we curated two variants–Virat1-RC and Virat2-RC–by segmenting videos into 50-frame clips and redefining classes around intrusion and threat-related activities. The restructuring ensures temporal consistency and domain alignment (Fig. 6). Class-level statistics and descriptions are provided in Tables 2 and 12.2 (ii)UCF-Crime: UCF-Crime13consists of 1,900 untrimmed surveillance videos spanning 14 categories (13 anomalous + normal). To construct a balanced, GPU-efficient, and operationally representative subset through a multi-stage curation protocol. First, following the observation by Maqsood et al32. that the original dataset contains redundant, repeated and inappropriate clips that can degrade model performance, we removed such segments. Second, long videos were uniformly downsampled to 300-frame chunks at 5 fps to standardize temporal density. Third, within each chunk, we retained only the top-50 most dynamic frames using motion energy: 20 where is the pixel intensity at location (x, y) in frame t. Crucially, every resulting 50-frame clip was manually inspected to ensure it contains a clear, single-activity instance that aligns with its assigned class label–eliminating clips with ambiguous, missing, or multi-event content. All manually verified clips were inspected by domain experts with prior experience in surveillance event annotation to ensure label correctness and temporal coherence. This human-in-the-loop validation guarantees label fidelity and temporal coherence, ensuring that our benchmark reflects real-world motion-triggered surveillance scenarios. The final standardized clips emphasize action-relevant regions (Fig. 7), and the class-level distribution is summarized in Table 11. UCF-Crime: UCF-Crime13consists of 1,900 untrimmed surveillance videos spanning 14 categories (13 anomalous + normal). To construct a balanced, GPU-efficient, and operationally representative subset through a multi-stage curation protocol. First, following the observation by Maqsood et al32. that the original dataset contains redundant, repeated and inappropriate clips that can degrade model performance, we removed such segments. Second, long videos were uniformly downsampled to 300-frame chunks at 5 fps to standardize temporal density. Third, within each chunk, we retained only the top-50 most dynamic frames using motion energy: 20 where is the pixel intensity at location (x, y) in frame t. Crucially, every resulting 50-frame clip was manually inspected to ensure it contains a clear, single-activity instance that aligns with its assigned class label–eliminating clips with ambiguous, missing, or multi-event content. All manually verified clips were inspected by domain experts with prior experience in surveillance event annotation to ensure label correctness and temporal coherence. This human-in-the-loop validation guarantees label fidelity and temporal coherence, ensuring that our benchmark reflects real-world motion-triggered surveillance scenarios. The final standardized clips emphasize action-relevant regions (Fig. 7), and the class-level distribution is summarized in Table 11. (iii)In-House Dataset (IHD): This consists of real-world surveillance footage collected from multiple commercial sites, captured using Axis, Hanwha, and Hikvision cameras under diverse operational conditions–including varying illumination, occlusion, and adverse weather. Cameras were mounted at heights of 8–20 ft with tilt angles between and , and equipped with standardized IR calibration to ensure uniform scene coverage with road and sky masking at configuration itself from the FOV. All video clips were preprocessed to a resolution of 1280×720 at 5 fps and segmented into 50-frame event clips triggered by edge-side motion detection. Specifically, motion was detected by observing consistent activity across three consecutive frames, with each clip including contextual pre- and post-event frames to preserve temporal semantics for real-time analytics (Fig. 8). The class taxonomy and dataset split statistics are detailed in Tables 2 and 12. In-House Dataset (IHD): This consists of real-world surveillance footage collected from multiple commercial sites, captured using Axis, Hanwha, and Hikvision cameras under diverse operational conditions–including varying illumination, occlusion, and adverse weather. Cameras were mounted at heights of 8–20 ft with tilt angles between and , and equipped with standardized IR calibration to ensure uniform scene coverage with road and sky masking at configuration itself from the FOV. All video clips were preprocessed to a resolution of 1280×720 at 5 fps and segmented into 50-frame event clips triggered by edge-side motion detection. Specifically, motion was detected by observing consistent activity across three consecutive frames, with each clip including contextual pre- and post-event frames to preserve temporal semantics for real-time analytics (Fig. 8). The class taxonomy and dataset split statistics are detailed in Tables 2 and 12. Virat2-RC curation workflow showing segmentation into 50-frame clips, class remapping, and balance checks. Table 2Class descriptions for Virat1-RC, Virat2-RC, and IHD datasets. Short Name Class Name Description NTMNon-Threat MovementMovements of flags, trees, or other benign objects without person or vehicle motion.PMVPerson MovementIndividuals entering/exiting vehicles, walking, or running.PSAPerson Stationary ActStationary human activities such as loading or object handling.PVMPerson and Vehicle MovementConcurrent movement of persons and vehicles within the scene.RAWRain and Weather InducedDrizzle, rain, or environmental effects reducing visibility.SUSSuspicious ActivityLight or shadow movements, or behavior suggesting potential intrusion.VMVVehicle MovementMovement of one or more vehicles without human presence. Class descriptions for Virat1-RC, Virat2-RC, and IHD datasets. Table 12Class-wise distribution for Virat1-RC, Virat2-RC, and IHD datasets (Train/Val/Test/Total).ClassVirat1-RCVirat2-RCIHDNTM111/25/69/20581/18/43/14291/20/52/163PMV111/25/24/16081/18/296/39591/20/77/188PSA111/25/68/20481/18/59/15891/20/21/132PVM111/25/448/58481/18/143/24291/20/81/192PVI–81/18/17/116–RAW––91/20/42/153SUS––91/20/26/137VMV111/25/178/31481/18/133/23291/20/33/144 Total 555/125/787/1467486/108/691/1285637/140/332/1109 Class-wise distribution for Virat1-RC, Virat2-RC, and IHD datasets (Train/Val/Test/Total). UCF-Crime preparation showing clip segmentation, filtering, and motion-energy based frame selection. Table 11UCF-Crime class-wise video distribution across Train/Validation/Test splits.ClassTrainVal.TestTotalClassTrainVal.TestTotalAbuse3581760Arrest3582164Arson3581356Assault3582467Burglary3584184Explosion358851Fighting3582972Normal455104115674Road Accidents3582770Robbery3583174Shooting3585194Shoplifting358750Stealing3584992Vandalism358750 Total 9102084401558 UCF-Crime class-wise video distribution across Train/Validation/Test splits. IHD event creation workflow. Computing resources Experiments ran on an Intel i9-9900K (8C/16T, 3.6–5.0 GHz), NVIDIA RTX 2080 Ti (11 GB), 62 GB RAM, and SSD storage. Software: Ubuntu 20.04, Python 3.8.10, PyTorch 1.9.0/TorchVision, CUDA 11.1, cuDNN, NumPy, OpenCV 4.2.0, and Matplotlib. We used PyTorch DataLoader (multiprocessing), caching, and tensor-level optimizations for efficient GPU utilization. Training protocol and hyperparameter configuration We adopt a stratified split of 70:15:15 for training, validation, and testing. For IHD and Virat1/2-RC, the split is anchored to the minority class: the class with the fewest samples defines the per-split counts, and all other classes are downsampled proportionally to preserve balance across train/val/test. For UCF-Crime, we retain the original per-class proportions while enforcing the same 70:15:15 partition. All dataset splits were generated using a fixed random seed (42) and stratified sampling to ensure reproducibility and consistent class distributions. All models are trained for 50 epochs; the proposed architecture uses the hyperparameters in Table 3. The single checkpoint with the highest validation accuracy is selected for final test-set evaluation. Table 3Training hyperparameters used for the TriPath3DNet.HyperparameterValueHyperparameterValueBatch size2OptimizerAdamLearning rate0.001Weight decayTraining epochs50Input frame size224 × 224 Training hyperparameters used for the TriPath3DNet. Performance evaluation metrics We evaluate the proposed framework from two complementary perspectives: (1) classification performance using standard recognition metrics, and (2) spatio–temporal interpretability using Grad-CAM–based localization metrics. (a) Classification Performance. Model-level performance is quantified using Accuracy (Acc), Precision (Pre), Recall (Rec), F1-score, and Area Under the ROC Curve (AUC). For imbalanced datasets, Recall and F1-score provide more reliable insight into discriminative performance: The ROC curve plots True Positive Rate (TPR) against False Positive Rate (FPR): For multi-class classification, we report the macro-averaged One-vs-Rest AUC: where C is the total number of classes. (b) Spatio–Temporal Interpretability Let the Grad-CAM volume be with frame maps , and binary motion/GT masks . Foreground Energy Ratio (FER, ↑). Spatial alignment of saliency with true foreground: Temporal Coherence (TC, ↑). Frame-to-frame stability of salient regions. With per-frame threshold and , define : Spatial Concentration (Entropy, ↓). Sharpness of attention within a frame. Let and : Prediction Confidence. Decisiveness of the final prediction for logits with : Reading guide: FER ↑ (better spatial alignment), TC ↑ (smoother temporal tracking), Entropy ↓ (tighter focus), Conf ↑ (stronger decision). Results and discussion Spatio-temporal interpretability analysis We evaluated the proposed TriPath3DNet against the baseline R3D-18 using Grad-CAM on LPS dataset, with representative per-frame visualizations at shown in Fig. 9. We summarize five interpretability indicators–FER), BSR, TC, Entropy, and Confidence–to quantify spatial alignment, background rejection, temporal stability, map compactness, and decisiveness, respectively. Quantitatively, TriPath3DNet achieves higher FER (0.1953 vs. 0.1928) and BSR (0.1953 vs. 0.1928), indicating that its Grad-CAM energy concentrates more within motion/GT regions while leaking less into the background. It also yields lower Entropy (10.034 vs. 10.155), reflecting sharper and more compact attention maps, and higher Confidence (0.915 vs. 0.871), denoting more decisive predictions. Although R3D-18 shows a slightly higher TC (0.9739 vs. 0.9682), the gap is marginal and both models are highly stable (TC >\!0.96). Overall, TriPath3DNet exhibits superior spatial precision, cleaner background suppression, and stronger predictive certainty while maintaining comparable temporal smoothness–supporting its advantage for interpretable and reliable activity localization in surveillance video. Frame-wise qualitative comparison between R3D and TriPath3DNet using Grad-CAM at frames for two representative samples (top: VMV, bottom: PSA). Each column shows the Original frame, R3D Grad-CAM, and TriPath3DNet Grad-CAM. Classification performance analysis Table 4 summarizes the generalization capability of TriPath3DNet across four heterogeneous surveillance datasets, demonstrating strong consistency despite variations in environment, motion patterns, and event complexity. On IHD, a highly diverse industrial dataset with low-quality frames and challenging lighting conditions, the model achieves solid performance (Acc = 89.46%, F1 = 87.69%), indicating robustness to real-world noise and short-clip constraints. Performance improves significantly on Virat1-RC (Acc = 93.77%, F1 = 88.48%) due to more predictable action patterns and clearer spatial cues. The strongest results appear on Virat2-RC, where TriPath3DNet obtains exceptionally high discrimination (Acc = 95.37%, F1 = 95.09%, AUC = 99.42%), validating the model’s ability to leverage its multi-branch temporal design to capture both short-term motion bursts and long-duration behaviors. Although UCF-Crime is a weakly supervised, highly noisy dataset, the model still maintains competitive performance (Acc = 86.14%, AUC = 94.64%), demonstrating good resilience to label noise and temporal clutter. Training and validation accuracy trends over 50 epochs Fig. 10) show rapid convergence and minimal overfitting on Virat1-RC, Virat2-RC, and IHD, with validation accuracy stabilizing in the high-80s to low-90s. For UCF-Crime Clips, validation accuracy exhibits higher variance due to its complexity, but the gap between training and validation remains controlled, indicating robust generalization. Class-specific performance was assessed using One-vs-Rest ROC curves (Fig. 12), with AUC scores ranging from 0.83 to 1.00 across all datasets and classes. Confusion matrices (Fig. 11) reveal strong separability, even for rare events. On highly imbalanced test sets like UCF-Crime, TriPath3DNet maintains high discriminative power for critical classes such as arson, burglary, and road accidents, underscoring its practical utility in real-world anomaly detection. Table 4Testing performance of TriPath3DNet. AUC is one-vs-rest (macro). All metrics in %.DatasetAccPrecRecF1AUCIHD89.4686.9789.0587.6998.07Virat1-RC93.7785.6393.6788.4897.06Virat2-RC95.3794.5496.2695.0999.42UCF-Crime86.1480.6679.2479.5794.64 Testing performance of TriPath3DNet. AUC is one-vs-rest (macro). All metrics in %. Training and validation accuracy curves of the TriPath3DNet on four surveillance datasets. Multi-class ROC-AUC curves for the proposed TriPath3DNet on four datasets. Confusion matrices of TriPath3DNet model. Quantitative comparison with contemporary approaches Table 10 presents the classification accuracy (%) of several baseline 3D CNN architectures–R(2+1)D, R3D, I3D, and X3D–alongside their temporal variants augmented with recurrent units (SRU, GRU, or LSTM), evaluated across four benchmark datasets. In most cases, naively appending recurrent modules degrades performance relative to the original backbones, suggesting that simple temporal aggregation may disrupt the spatiotemporal feature representations learned by strong 3D CNNs. Notably, I3D benefits slightly from SRU integration on UCF-Crime, but this trend is inconsistent across datasets. In stark contrast, our proposed TriPath3DNet achieves state-of-the-art results on all four benchmarks, substantially outperforming both base models and their recurrent-augmented counterparts–e.g., by +5.37% over X3D on Virat1-RC and +8.07% over R3D on UCF-Crime. This demonstrates the effectiveness of our architecture in jointly modeling spatial, temporal, and contextual cues without relying on post-hoc sequential modeling. Table 10Classification accuracy (%) of baseline models and their temporal variants across four benchmark datasets.MethodVirat1-RCVirat2-RCIHDUCF-CrimeR(2+1)D6991.1087.5588.5573.86+ SRU-51283.3580.0285.2459.72+ GRU-51285.6480.6188.8558.41+ LSTM-51284.6277.5681.9356.82R3D7088.0685.5285.5478.40+ SRU-51282.2170.4869.8866.82+ GRU-51267.3479.8877.1168.86+ LSTM-51265.1877.1372.8970.45I3D7179.4176.9876.5174.77+ SRU-51278.0274.1077.1176.36+ GRU-51276.4971.0871.0873.18+ LSTM-51277.6476.8571.9972.73X3D3081.7087.4082.8371.14+ SRU-51288.0592.0483.7365.68+ GRU-51282.9786.5483.1364.55+ LSTM-51284.4989.5886.7466.13TriPath3DNet93.7795.3789.4686.14Note: Backbones include I3D-ResNet50, R3D-18, and X3D-M. All recurrent modules use hidden size = 512. Classification accuracy (%) of baseline models and their temporal variants across four benchmark datasets. Note: Backbones include I3D-ResNet50, R3D-18, and X3D-M. All recurrent modules use hidden size = 512. We conduct a systematic evaluation of TriPath3DNet against a wide range of recent video architectures–including MViTv149, MViTv250, Video Swin Transformer51, VideoMAE52, and recurrent-augmented CNNs–all trained from scratch using an identical training protocol on the IHD, UCF-Crime, and Virat1-RC datasets. Specifically, we employ the same input resolution (224× 224). As shown in Table 5, TriPath3DNet consistently outperforms all baselines across datasets and metrics. On Virat1-RC, it achieves 95.37% accuracy and 95.09% F1-score, surpassing even strong hierarchical transformers like MViTv2 (95.10% Acc) while yielding the highest recall (96.26%) and AUC (99.42%). On the challenging anomaly detection benchmark UCF-Crime, where most vision transformers struggle (e.g., ViViT: 4.73% F1; VideoMAE: 13.88% F1), TriPath3DNet attains 86.14% accuracy and 79.57% F1–substantially ahead of MViT V1 (67.32% F1) and MViT V2 (71.50% F1). This highlights its superior capability in modeling long-range, sparse, and semantically ambiguous events. Table 5Comparison across IHD, UCF-Crime, and Virat2-RC. Metrics are in %. Inference time is per clip.DatasetModelAccPreRecF1AUCP(M) (s)IHDMViTv14991.5691.0990.8790.7098.2936.310.0235MViTv25091.8790.4991.6490.4990.9651.230.1660Video Swin Transformer5172.5968.3371.4068.5993.3927.860.0216VideoMAE5262.3562.2463.3761.8784.9887.450.7130UniFormer-Lite5353.6152.0555.9749.4486.4639.810.0110ViViT5470.1865.7867.9966.4192.6687.450.0041SHViT5563.2562.3765.4661.6982.8938.620.0030UniFormerV25679.5282.2979.4480.0591.0722.010.0060ResNet50 + LSTM4575.0074.2574.3473.6392.2730.850.0023ResNet50 + SRU5775.9076.6175.9275.7893.7528.490.0084EfficientNetB7 + BiLSTM4461.7561.6262.9460.7688.6382.680.1228IncepResNetV3 + BiLSTM + Att42.73.4975.4367.3069.0091.4032.280.4054Video-Graph method3178.9280.6878.0178.9994.2424.560.0049TriPath3DNet89.4686.9789.0587.6998.0733.460.1290UCF-CrimeMViTv14976.1472.4965.4067.3293.7036.310.0136MViTv25078.8674.7070.7471.5095.8351.240.1300Video Swin Transformer5127.5030.809.857.8579.9027.860.0098VideoMAE5223.4119.5814.9113.8863.4387.450.2750UniFormer-Lite5336.1437.8819.7919.1671.4439.810.0074ViViT5426.3614.997.964.7362.8287.450.0039SHViT5531.3616.9611.479.7264.5338.620.0030UniFormerV25664.7763.4054.0756.3984.1522.010.0050ResNet50 + LSTM4558.6456.3749.5248.9587.8430.860.0030ResNet50 + SRU5767.9566.3055.8356.5188.9728.500.0053EfficientNetB7 + BiLSTM4441.3634.1731.0428.6480.7782.690.1303IncepResNetV3 + BiLSTM + Att42.59.5557.1151.0250.3089.3032.290.2090Video-Graph method3159.3255.7147.0748.2287.6924.560.0049TriPath3DNet86.1480.6679.2479.5794.6433.460.137Virat2- RCMViTv14994.7994.3694.6094.3799.3136.300.0237MViTv25095.1095.7595.1095.0098.6951.230.1620Video Swin Transformer5163.1061.8972.9861.4996.0027.860.0209VideoMAE5260.9353.6965.6955.6990.6887.450.9720UniFormer-Lite5368.6070.3780.1670.2394.7939.810.0121ViViT5467.5867.0677.4968.6994.8187.450.0039SHViT5573.8067.0580.8670.7794.3638.610.0030UniFormerV25680.7576.8386.9978.7996.5322.010.0056ResNet50 + LSTM4578.1574.4184.8677.7696.9830.850.0023ResNet50 + SRU5781.3377.8886.4580.1696.2328.490.0084EfficientNetB7 + BiLSTM4469.0368.3776.9870.2993.5782.680.1337IncepResNetV3 + BiLSTM + Att42.83.7981.9186.9082.8497.4032.280.4028Video-Graph method3179.8876.9187.3880.1597.1824.560.0049TriPath3DNet95.3794.5496.2695.0999.4233.460.1370MViTv1: Base 16× 4; MViTv2: Base 32× 3; Video Swin: Swin3D-T/16; VideoMAE: ViT-Base/16; UniFormer-Lite: UniFormer-ViT-Patch4×16×16; ViViT: vivit-b-16×2; SHViT: SHViT-B/512-12L (pt=2, 16×16); UniFormerV2: S-variant. Comparison across IHD, UCF-Crime, and Virat2-RC. Metrics are in %. Inference time is per clip. MViTv1: Base 16× 4; MViTv2: Base 32× 3; Video Swin: Swin3D-T/16; VideoMAE: ViT-Base/16; UniFormer-Lite: UniFormer-ViT-Patch4×16×16; ViViT: vivit-b-16×2; SHViT: SHViT-B/512-12L (pt=2, 16×16); UniFormerV2: S-variant. On IHD, TriPath3DNet achieves 89.46% accuracy with an AUC of 98.07%, demonstrating strong generalization even when competing against MViTv2 (91.87% Acc) but with significantly better calibration (AUC gap: +7.1 points). Notably, while lightweight models (e.g., SHViT, UniFormer-Lite) offer faster inference, they suffer severe performance drops–particularly on UCF-Crime–reinforcing that architectural design matters more than parameter count alone in complex video understanding tasks. These results, derived under a uniform training and evaluation setup, confirm that TriPath3DNet’s multi-path spatiotemporal modeling provides a robust and generalizable inductive bias for both action recognition and anomaly detection. To further analyze the trade-off between performance and model complexity, we conduct a focused comparison between TriPath3DNet, MViTv149, and MViTv250 on the Virat1-RC dataset under identical training and evaluation settings. As shown in Table 6, TriPath3DNet achieves the highest accuracy (93.77%) and significantly outperforms both baselines in recall (93.67%) and F1-score (88.48%), indicating stronger class-balanced discrimination–particularly important in multi-class video recognition. While MViTv2 attains a marginally higher AUC (98.94% vs. 97.06%), TriPath3DNet does so with 17.8M fewer parameters than MViTv2 (33.46M vs. 51.23M) and comparable inference latency (0.136 s vs. 0.166 s). Moreover, it improves upon MViTv1 in accuracy (+1.14%), F1 (+2.42%), and recall (+4.29%) while using slightly fewer parameters. These results demonstrate that TriPath3DNet offers a more efficient and effective architecture for clip-level action recognition on Virat1-RC, surpassing even advanced hierarchical vision transformers in key discriminative metrics. Table 6Virat1-RC clip-level comparison across accuracy, efficiency, and latency.ModelAccPrecRecF1AUCPar (M) (s)MViTv192.6384.6989.3886.0698.4436.300.0233MViTv293.5285.7988.4186.7298.9451.230.1660TriPath3DNet93.7785.6393.6788.4897.0633.460.1360Notes. Inference time reported per clip. Frame counts differ: MViTv1 (16), MViTv2 (32), TriPath3DNet (50). Abbrev.: Acc = Accuracy, Prec = Precision, Rec = Recall, AUC = Area Under ROC, Par (M) = Parameters (millions), = inference time. Virat1-RC clip-level comparison across accuracy, efficiency, and latency. Notes. Inference time reported per clip. Frame counts differ: MViTv1 (16), MViTv2 (32), TriPath3DNet (50). Abbrev.: Acc = Accuracy, Prec = Precision, Rec = Recall, AUC = Area Under ROC, Par (M) = Parameters (millions), = inference time. On Virat1-RC, TriPath3DNet delivers the best end-to-end balance: highest Accuracy (93.77%) and F1 (88.48%) driven by notably higher Recall (93.67%) than MViT-V1 (89.38%) and MViT-V2 (88.41%), indicating fewer missed alarms–critical for surveillance–while accepting a tiny Precision trade-off (85.63% vs. 85.79% for MViT-V2); although its AUC (97.06%) is slightly below V1/V2 (98.44/98.94), deployment uses fixed thresholds, so F1/Recall are more decision-aligned. It is also the most parameter-efficient (33.46M vs. 36.30M/51.23M) with latency between the baselines (0.1360 s vs. 0.0233 s for V1 and 0.1660 s for V2); normalized per frame to account for different clip lengths (16/32/50), TriPath is 1.9× slower than V1 but 1.9× faster than V2, making it the preferable choice when minimizing misses and maximizing overall decision quality under a compact model size, while V1 suits hard real-time and V2 favors high AUC at higher compute. Comparison with state-of-the-art methods Tables 7 and 8 summarize the performance of TriPath3DNet against representative state-of-the-art (SOTA) methods across the Virat1, Virat2, and UCF-Crime benchmarks. Virat1 Dataset.On Virat1, classical activity-recognition pipelines such as those by Ryoo et al65. and Zhu et al66. achieved moderate performance (61–68%). More recent deep architectures (e.g39.,) improved accuracy to 88.00%. TriPath3DNet further improves the benchmark to 93.77%, indicating enhanced robustness to viewpoint variations and inter-class similarity in real-world surveillance scenes. Table 7Comparative performance of previous methods on Virat2 and UCF-Crime datasets.(a) Virat2 Accuracy DatasetAuthor [Ref]# ClsAcc (%)M. Alijanpour et al58.685.60Y. Ghadi et al59.982.55B. Ayhan et al31.492.50B. Ayhan et al31.662.86T. Mahmud et al60.1271.80S. Hendi et al38.1275.22X. Wang et al61.1251.91Y. Li et al62.674.59Y. Li et al39.1294.20TriPath3DNet695.37(b) UCF-Crime Accuracy DatasetAuthor [Ref]# ClsAcc (%)M. Kumar et al42.562.04M. Qasim et al57.1491.24L. Vuyyuru et al63.1478.30N. Mumtaz et al43.648.85W. Ahmed et al40.1447.34Y. Hao et al35.1437.90A. Singh et al64.1482.91X. Zhou et al33.1481.90TriPath3DNet1486.14 Comparative performance of previous methods on Virat2 and UCF-Crime datasets. Table 8Comparative performance of the proposed TriPath3DNet model with previous works on the Virat1 and UCF-Crime datasets.(a) Virat1 Accuracy ComparisonAuthor [Reference]Accuracy (%)Y. Li et al39.88.00M. S. Ryoo et al65.61.70Y. Zhu et al66.68.00N. M. Nayak et al67.61.50TriPath3DNet93.77(b) UCF-Crime AUC Comparison (Frame-level WS-VAD)Author (et al.)AUC (%)Qasim et al57.91.64Ahmed et al40.80.60Tan et al68. (RGB)82.69Tan et al68. (Flow)85.56Tan et al68. (RGB+Flow)86.71Karim et al16.86.94Maqsood et al32.82.00Zhou et al17.87.24Wu et al19.88.08Majhi et al46.91.58Fan et al49.86.19Lv et al14.86.06Zhang et al18.86.22Ye et al21.86.55Wang et al22.85.07TriPath3DNet94.64 Comparative performance of the proposed TriPath3DNet model with previous works on the Virat1 and UCF-Crime datasets. Virat2 Dataset. On the Virat2 dataset, prior works such as STDRN39 reported strong performance (94.20%) under a 12-class setting, while earlier CNN-based and handcrafted-feature approaches achieved accuracies ranging from 51.91% to 92.50%. Under a 6-class configuration, TriPath3DNet achieves 95.37% accuracy, surpassing previously reported results in comparable class settings. The improvement can be attributed to the complementary short-term, long-term, and temporal-difference pathways that jointly model both transient motion cues and sustained contextual patterns. UCF-Crime Dataset. For UCF-Crime, recent weakly supervised video anomaly detection (WS-VAD) approaches report frame-level AUC values typically in the 86–92% range19,46,57. Transformer-based and hybrid MIL frameworks achieve competitive performance but often rely on long untrimmed sequences and heavy temporal modeling. Under our fixed-length clip protocol, TriPath3DNet attains a macro AUC of 94.64%, exceeding the strongest reported baselines in Table 8. Importantly, TriPath3DNet maintains this performance while using a lightweight R3D-18 backbone (33.46 M parameters) and requiring only 129 ms per 50-frame clip on an 11 GB GPU. Ablation analysis To evaluate the contribution of each component in TriPath3DNet, we conduct an ablation study on the IHD and UCF-Crime datasets (Tables 9, Fig. 13 and Fig. 14). The full model (A1: short + long + TDE) achieves the highest accuracy–89.46% on IHD and 86.14% on UCF-Crime–with strong AUC (>94%) and consistent inference time (129–137 ms per clip). Removing TDE (A2) causes the largest performance drop on UCF-Crime (-7.50% accuracy), confirming its critical role in capturing subtle or sparse motion cues. Ablating long-term (A3) or short-term (A4) branches also degrades performance: short-term cues are especially vital for IHD (-5.12%), while long-term context benefits UCF-Crime (-7.05%), underscoring the necessity of multi-scale temporal modeling. Alternative fusion strategies (B1, C1), max pooling instead of average pooling (D1), and omission of Kinetics pretraining (E1) all yield inferior results. Notably, removing pretraining (E1) reduces UCF-Crime accuracy by 25.46%, highlighting the importance of large-scale initialization. Table 9Unified ablation on IHD and UCF-Crime: Test Accuracy (%), Precision (%), Recall (%), F1 (%), and AUC (%) per variant. Inference time columns are removed.Var.Model Config.IHDUCF-CrimeAccPrecRecF1AUCAccPrecRecF1AUCA1Full (short + long + TDE)89.4686.9789.0587.6998.0786.1480.6679.2479.5794.64A2w/o TDE87.9586.4786.9486.4597.5978.6471.7169.5369.6693.57A3w/o Long-Term87.6586.1287.4186.3597.1279.0977.8969.7371.3093.25A4w/o Short-Term84.3483.9984.2283.8696.2579.7775.7669.6971.0894.18B1Concat + Conv3D82.8380.7782.0180.4896.2377.7376.7064.8370.2893.20C1Fusion after layer186.1486.1885.2585.2297.0976.3671.6568.7768.2892.06D1Max pooling (avg replaced)85.2482.9685.4183.8497.2970.9168.0961.5962.7989.46E1No Kinetics pretrain76.8180.3475.0476.4593.3360.6861.0946.7748.3086.35 Unified ablation on IHD and UCF-Crime: Test Accuracy (%), Precision (%), Recall (%), F1 (%), and AUC (%) per variant. Inference time columns are removed. Test accuracy of TriPath3DNet ablation variants (A1–E1) on IHD and UCF-Crime datasets. Each variant represents a specific architectural modification for evaluating the contribution of temporal streams. Test metrics of TriPath3DNet ablation variants (A1–E1) on IHD and UCF-Crime datasets. Conclusion This work presents TriPath3DNet, a novel, efficient, and interpretable 3D CNN architecture designed specifically for real-time, multi-class classification of fixed-duration surveillance video clips. By integrating three complementary temporal pathways–short-term motion, long-term context, and Temporal Difference Encoding (TDE)–TriPath3DNet effectively captures both transient events and prolonged anomalous behaviors, enabling robust detection under challenging real-world conditions such as occlusion, poor lighting, and adverse weather. Extensive evaluation across four datasets–including the newly curated Virat1-RC, Virat2-RC, and UCF-Crime, as well as our proprietary (IHD)–demonstrates that TriPath3DNet achieves state-of-the-art performance, with up to 95.37% accuracy, 99.42% AUC, and 137 ms per 50-frame clip. Notably, it outperforms both lightweight CNNs and recent vision transformers (e.g., MViT, VideoSwin) by significant margins, despite having only 33.46 M parameters. This highlights a key insight: while transformers excel at modeling long, data-rich sequences, they are often overkill for short, sparsely sampled surveillance clips (e.g., at 5 fps) where local spatio-temporal dynamics dominate–and where carefully designed CNNs can offer superior efficiency and accuracy. Ablation studies confirm that each temporal branch contributes meaningfully to performance, with gains attributable to architectural design rather than model scale. Grad-CAM visualizations further reveal that TriPath3DNet produces spatially precise and temporally coherent attention maps, enhancing interpretability–a critical requirement for security-critical deployments. To our knowledge, TriPath3DNet is the first framework to jointly address the core operational constraints of industrial asset-protection systems: (i) fixed-length, motion-triggered video inputs, (ii) the detection of subtle, low-motion anomalies–such as loitering or zone intrusion–under real-world conditions including occlusion, variable lighting, and environmental clutter. By aligning architectural design with edge–cloud deployment requirements and the 2-minute alert SLA, our work bridges the persistent gap between academic video analytics and practical surveillance needs, establishing a new efficiency–accuracy benchmark for real-time, deployable video classification. Limitations. The current design of TriPath3DNet is optimized for single-activity, fixed-duration surveillance scenarios typical of asset-protection systems, where each motion-triggered clip contains one primary event of interest. Consequently, its performance may degrade on clips containing multiple concurrent or sequential activities, as the model assumes a single dominant class per input. Similarly, the architecture is not designed for long, untrimmed videos; both accuracy and computational efficiency are expected to suffer due to limited temporal receptive fields and fixed pooling structures. These limitations reflect intentional operational trade-offs rather than fundamental flaws, but they do restrict applicability to broader video understanding tasks. Future Work. We plan to pursue three directions: (i) extending the framework to localize and classify multiple sequential events in longer, untrimmed videos; (ii) exploring hybrid architectures that fuse TriPath3DNet with lightweight vision transformers (e.g., MViTv1) to leverage global attention while preserving temporal efficiency; and (iii) developing motion-aware filtering mechanisms to suppress false alarms caused by environmental nuisances (e.g., swaying trees, fluttering flags)–which account for over 90% of false positives in current cloud-based systems and severely limit scalability. Given that MViTv1 achieves competitive AUC on IHD (98.29% vs. 98.07%) at ~10× higher speed, such hybrid models could yield a Pareto-optimal solution for both short and long surveillance sequences. Supplementary Information Supplementary Information. Acknowledgements The authors gratefully acknowledge Pro-Vigil Surveillance for providing proprietary dataset used in this research, with the condition that it shall not be made publicly available. Appendix Author contributions Author Contributions Conceptualization: Mohanarangan K; Methodology: Mohanarangan K; Data curation: Mohanarangan K; Writing – original draft: Mohanarangan K; Writing – review and editing: Mohanarangan K; review and supervision: Palanisamy P. Funding This research did not receive any specific funding from any organization. Data availability The Virat1-RC, Virat2-RC, and UCF-Crime-Clip (UCF-CC) datasets used in this study are derived from publicly available sources and have been carefully curated to suit the experimental requirements of this research. These curated versions are available upon reasonable request for academic or non-commercial research purposes. Interested researchers may contact the corresponding author to obtain access. In contrast, the proprietary IHD dataset contains real-world surveillance footage involving sensitive content and privacy concerns; therefore, it is not publicly shareable due to confidentiality constraints. Code availability The source code used in this study is available and shared at https://doi.org/10.5281/zenodo.18718931. Declarations Competing interests The authors declare no competing interests. Ethical approval Not applicable. Informed consent Not applicable. Footnotes All supplementary materials are accessible via the publisher’s website or directly linked with the article. Dataset-level distribution statistics are provided in Appendix Table 12. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary Information The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-52403-z. References Associated Data Supplementary Materials Supplementary Information. Data Availability Statement The Virat1-RC, Virat2-RC, and UCF-Crime-Clip (UCF-CC) datasets used in this study are derived from publicly available sources and have been carefully curated to suit the experimental requirements of this research. These curated versions are available upon reasonable request for academic or non-commercial research purposes. Interested researchers may contact the corresponding author to obtain access. In contrast, the proprietary IHD dataset contains real-world surveillance footage involving sensitive content and privacy concerns; therefore, it is not publicly shareable due to confidentiality constraints. The source code used in this study is available and shared at https://doi.org/10.5281/zenodo.18718931.