Anomaly Detection in Surveillance Videos Using Slowfast with Grad-CAM Visualizations

Automatic identification of anomalies in surveillance videos is a growing area of research due to its potential to assist in real-time crime prevention and public safety monitoring. Despite the development of several video anomaly detection algorithms, accurately identifying suspicious or illegal activities in real-world environments remains a challenging task due to variations in scene context, ambiguous behavior patterns, and the rarity of anomalous events. To address this, automatic anomaly detection using deep learning has emerged as a promising solution, with methods such as CNNs and RNNs becoming common; however, they often struggle to jointly capture complex spatial-temporal patterns in real-world surveillance footage. This study proposes a framework based on the SlowFast network-a dual-pathway 3D convolutional architecture that captures both slow-moving spatial semantics and fast temporal motion cues from video sequences. The model is trained and evaluated on the UCF-Crime dataset using a $\mathbf{7 5 \% - 1 5 \% - 1 0 \%}$ train-validation-test split. Our approach demonstrates promising performance, achieving an overall classification accuracy of 81.05 % in detecting anomalous activities in unseen test dataset. Furthermore, Grad-CAM is integrated to provide visual insights into the model's decision-making process. This enhances the transparency and interpretability of the model, allowing human operators to better understand, and validate automated surveillance systems.