A-VAE: Attention based Variational Autoencoder for Traffic Video Anomaly Detection

Video surveillance systems are essential for an intelligent and smart traffic monitoring system. Detecting and recognizing traffic anomalies is the foremost task for the safety and security of human lives. The advancement of computer vision technology, especially in deep learning, has paved the way for accurate anomaly detection. This paper presents a novel attention-based variational autoencoder (A-VAE) architecture for detecting anomalies in traffic videos. A-VAE is designed with the help of 2D CNN and BiLSTM layers with an attention mechanism for representation learning. In addition, a shortcut connection is employed between the spatial encoder and the spatial decoder, which helps for better decoding. A-VAE is trained end-to-end on regular video sequences in an unsupervised manner. To check the effectiveness of the proposed A-VAE, two challenging real-world traffic datasets (Crossroad1 and Pedestrian) have been utilized. A-VAE delivers the AUC and EER of 87.4% and 27% in Crossroad1, and 82.2% and 31% on the Pedestrian dataset, respectively. During testing, an improved runtime of 0.0149s (∼ 67 fps) and 0.0192s (∼ 52 fps) has been achieved for the Crossroad1 and Pedestrian datasets, respectively.