Anomaly Detection In Surveillance Videos With Variable Temporal Observation

On developing a deep learning-based approach for recognizing anomalies in surveillance videos under variable temporal observation. The reported five-fold cross-validation results show a mean accuracy of 78% at both 50% and 100% video observation. At 50% observation, the model achieves a macro F1-score of 74.4%, indicating that it can classify the selected action categories using partial video information. The evaluation across different observation levels also shows that the highest mean accuracy, 79.6%, is obtained at 75% observation. These results demonstrate the potential of the proposed architecture for partial-video action recognition. However, the model has limitations. Its evaluation is based on a reduced dataset containing three classes: Normal, Road Accidents and Robbery. The limited diversity of the dataset may affect the model's ability to generalize to different surveillance environments. Furthermore, recognizing an action from a partial video does not necessarily establish that the model can predict an anomaly before the event begins.

[1] R. E. Sri, Y. E. Mulyanto, P. I. K. Eddy, and P. M. Hery, “Human Activity Classification Using Deep Learning Based on 3D Motion Feature,” Machine Learning with Applications, vol. 12, p. 100461, June 2023, doi: 10.1016/j.mlwa.2023.100461. [2] A. Hussain, T. Hussain, W. Ullah, and S. W. Baik, “Vision Transformer and Deep Sequence Learning for Human Activity Recognition in Surveillance Videos,” Computational Intelligence and Neuroscience, vol. 2022, pp. 1–10, Apr. 2022, doi: 10.1155/2022/3454167. [3] I. Lamaakal, C. Yahyati, Y. Maleh, K. El Makkaoui, and I. Ouahbi, “An Explainable Hybrid CNN-transformer Model for Sign Language Recognition on Edge Devices Using Adaptive Fusion and Knowledge Distillation,” Scientific reports, vol. 16, no. 1, p. 7143, Mar. 2026, doi: 10.1038/s41598-026-38478-8. [4] J. Manoharan and Yuvaraj Sivagnanam, “A Novel Human Action Recognition Model by Grad-CAM Visualization With Multi-Level Feature Extraction Using Global Average Pooling With Sequence Modeling by Bidirectional Gated Recurrent Units,” International Journal of Computational Intelligence Systems, vol. 18, no. 1, May 2025, doi: 10.1007/s44196-025-00848-x. [5] H. Wang, Y. Wang, S. Yan, X. Du, Y. Gao, and H. Liu, “Merge-and-Split Graph Convolutional Network for Skeleton-Based Interaction Recognition,” Cyborg and Bionic Systems, vol. 5, Jan. 2024, doi: 10.34133/cbsystems.0102. [6] A. Saha, S. Gupta, S. K. Ankireddy, K. Chahine, and J. Ghosh, “Exploring Explainability in Video Action Recognition,” arXiv.org, Apr. 13, 2024. https://arxiv.org/abs/2404.09067 . [7] C. Roy et al., “Explainable Activity Recognition in Videos Using Deep Learning and Tractable Probabilistic Models,” ACM Transactions on Interactive Intelligent Systems, vol. 13, no. 4, pp. 1–32, Dec. 2023, doi: 10.1145/3626961. [8] F. Tempel, D. Groos, E. A. F. Ihlen, L. Adde, and I. Strümke, “Choose Your Explanation: A Comparison of SHAP and Grad-CAM in Human Activity Recognition,” Applied Intelligence, vol. 55, no. 17, Nov. 2025, doi: 10.1007/s10489-025-06968-3. [9] J. Garcia-Torres Fernandez, “Interpretability in Video-Based Human Action Recognition: Saliency Maps and GradCAM in 3D Convolutional Neural Networks,” Nordic Machine Intelligence, vol. 4, no. 2, pp. 19–25, Aug. 2024, doi: 10.5617/nmi.10587. [10] L. Wu, C. Zhang, and Y. Zou, “SpatioTemporal Focus for Skeleton-Based Action Recognition,” Pattern Recognition, vol. 136, p. 109231, Apr. 2023, doi: 10.1016/j.patcog.2022.109231.