Hybrid Deep Learning Models for Anomaly Detection in CCTV Video Surveillance
M. Humera Khanam, Roopa Chandrika R · 2025
Due to the growing necessity of automated activity capturing in urban areas, video surveillance systems gain significance in maintaining safety of society. Traditional anomaly detection techniques, such as Support Vector Machines (SVMs), K-Nearest Neighbors (KNN), and 3D Convolutional Neural Networks (3D CNNs), have demonstrated effectiveness in detecting suspicious activities. However, these methods often struggle with high false positive rates, occlusions, varying lighting conditions, and computational efficiency. The research creates an efficient deep learning anomaly detection framework that uses MobileNetV2 for spatial feature extraction alongside Bi-LSTM for temporal sequence learning for video surveillance applications. The purpose of this research is to create an anomaly detection system that boosts performance in CCTV footage despite changing illumination levels while keeping computational speed high. The primary challenges in existing systems include the inability to capture both spatial and temporal dependencies effectively, sensitivity to illumination changes, and the computational cost of complex architectures. This study aims to overcome these limitations by leveraging lightweight yet robust deep learning models that integrate spatial and sequential feature learning. The last two architectures, namely the Custom MobileNetV2 model and the MobileNet-based Bi-LSTM model, are employed to extract spatial and both spatial and temporal features respectively for video analysis. The models examined are assessed on the Smart-City CCTV Violence Detection Dataset, revealing competently in the recognition of violence regardless of illumination. The experiment of using the proposed MobileNet-based Bi-LSTM model yields better performance than the Custom MobileNetV2 model with the testing accuracy of 94.43% as opposed to 90.17%. Thus, this work demonstrates that intermediate hybrid and optimised architectures can enhance the efficiency of video surveillance systems used in anomaly detection.