Advanced Video Anomaly Detection with Enhanced Feature Learning Using Deep Convolutional Neural Nets and Context Mining Approach

The increasing deployment of CCTV surveillance systems necessitates intelligent anomaly detection for identifying unauthorized actions such as shoplifting, sudden sprints, and unusual object appearances. Traditional techniques rely on extensive labelled datasets, leading to high computational costs and impractical real-time applications. This research introduces a novel method integrating three pretrained Convolutional Neural Network (CNN) models for tracking, object detection, and anomaly detection, ensuring highly efficient feature extraction while minimizing manual labelling and computational overhead. A Denoising Autoencoder (DAE) reconstructs expected behaviour from usual activity patterns, highlighting anomalies through red bounding boxes for enhanced interpretability. The UCSD Pedestrian Dataset (Ped1 and Ped2) serves as the evaluation benchmark, where normal pedestrian movement is distinguished from anomalies such as bicycle intrusions or abrupt object appearances. Additionally, context mining techniques extract spatial and temporal behaviour features, improving discrimination between normal and anomalous events. The proposed approach significantly reduces computational complexity while enhancing real-time detection, achieving an accuracy of 96.4 %, an F1-score of 92.9 %, and an AUC-ROC of 97.1 %, with a processing rate of 32 FPS, ensuring scalability in large-scale surveillance environments. Future work will explore action recognition and key point detection to refine human behaviour characterization and extend applicability across diverse video datasets to further strengthen robustness and real-world implementation.