Robust Spatio-Temporal Anomaly Detection in Video Surveillance Using Deep Learning: A 3-Layered Convolutional Autoencoder with Temporal Regularity Learning

Anomaly detection in video surveillance is a critical application of deep learning in computer vision with significant implications for public safety and security. This paper presents an enhanced 3-Layered Convolutional Autoencoder (3L-CAE) combined with temporal regularity learning and ConvLSTM architecture for robust detection of unusual activities in surveillance videos. The proposed approach addresses the challenge of high-dimensional video data processing through an innovative spatio-temporal feature learning framework. Experimental validation on five benchmark datasets (Avenue, UCSD-Ped1, UCSD-Ped2, Subway Entrance, Subway Exit) demonstrates superior performance with accuracy rates of 91.67% (UCSD-Ped1), 92.57% (UCSD-Ped2), and 91.42% (Avenue), significantly outperforming existing CNN, RNN, and 3D-CNN approaches. The system achieves a computational efficiency of 3.9 seconds per 1000 frames while maintaining an AUC-ROC of 0.956 and 0.945 on benchmark datasets, making it suitable
for real-time surveillance applications.

Keywords: Anomaly detection, Convolutional autoencoder, ConvLSTM, Spatio-temporal learning, Video surveillance, Deep learning, Temporal regularity