Real-time highway traffic analysis and generative adversarial network based anomaly detection

Abstract

This thesis presents the design and development of an advanced video surveillance system aimed at enhancing highway traffic monitoring and anomaly detection. The proposed system integrates multiple traffic monitoring capabilities including multi-axle vehicle classification, speed estimation, and vehicle counting, which are essential for understanding and managing vehicular activity on highways. These features are individually developed using deep learning and computer vision methodologies to ensure accurate and real-time performance even in challenging environments. The system is further extended by incorporating an anomaly detection module based on Constrained video anomaly detection GAN (CVAD-GAN), which leverages latent space modeling to learn the normal patterns in traffic video frames. The CVAD-GAN framework reconstructs video frames and uses pixel-wise reconstruction error in conjunction with discriminator feedback to detect deviations from the learned normal behavior, thereby identifying anomalies in real-time video streams. During testing, reconstruction is used to localize abnormal regions in video frames, making the system robust and practical for deployment.