Anomaly Detection in Crowded Scene

Abstract

Monitoring abnormal events in a crowded scenes is essential these days, especially with the increase in surveillance cameras in most, if not all, places. This has made the field of computer vision an active field of research in recent periods. Identifying abnormal events in a crowded scene as soon as possible is essential, especially from the security side. It is a cumbersome and challenging task for humans because there are many surveillance cameras and overcrowding these days. Computer vision can be used to solve this problem and get accurate and high results. In this thesis, a convolutional autoencoder neural network architecture for anomaly detection in videos has been proposed. The proposed convolutional neural network model has been trained to recognize anomaly frames. The performance of the proposed network has been evaluated using Avenue [1] and UCSD [2] standard datasets designed to identify anomalous events in crowded scenes. Experimental results showed that the proposed model outperforms state-of-the-art methods achieving an accuracy of 71.96% and 89.52% on Avenue [1] and UCSD Peds2 [2] datasets respectively. A detailed analysis of the experiments used to choose the parameters is presented along with a comparison with other methods that used machine learning methods with an average accuracy of 69.31% and 63.90% on Avenue [1] and UCSD Peds2 [2] datasets respectively.