Multi Label Anomaly Detection for Crowd Scenes using Deep Learning

Abstract

quotAutomated video surveillance systems must be able to detect anomalies in congested environments, especially in conditions with a lot of pedestrian activity. This study presents the critical role of anomaly detection in crowded scenes for automatic video surveillance systems, particularly in alerting to potential issues arising from high foot traffic. The presented approach leverages advanced deep learning techniques, with a focus on two key components. newlineFirstly, introduce a novel anomaly detection and multi-label classification supervised learning method tailored for crowded scenes. For this, employ an enhanced deep learning model for multi-label anomaly classification, the Enhanced Recurrent Neural Network (E-RNN). Also, the significant contribution lies in optimizing the movement score threshold, appearance score threshold, and the number of hidden neurons in the RNN using a hybrid Elephant Herding - Grey Wolf Optimization (EH-GWO) technique, resulting in superior detection and classification accuracy. This method also achieves high accuracy compared to established approaches on benchmark datasets. However, to further enhance anomaly detection, transitioned to an unsupervised approach. newlineSecondly, in this emphasizes the importance of feature extraction by utilizing an Inception Capsule Auto-encoder (Inception-CAE) model to extract spatio-temporal features. These features are enriched with essential data to identify abnormal events. Also, calculate the reconstruction error between initial and reconstructed video frames, followed by determining the normality score relative to a threshold calculated via the Coyote Threshold Optimization algorithm (CTOA) method. The comprehensive evaluation, conducted on benchmark datasets including CUHK Avenue, UCSD Ped2, Live Video (LV), and ShanghaiTech covers various performance metrics such as ROC, AUC, accuracy, loss, model size, and time complexity. The presented approach consistently outperforms existing techniques, achieving an impressive maximum AUC of 99.1% for CUHK Avenue