Anomalous Event Detection and Localization Using Stacked Autoencoder

Anomalous event detection and localization from the crowd is a challenging problem to the computer vision community. It is an important aspect of intelligent video surveillance. Surveillance cameras are set up to monitor anomalous or unusual events. But, the majority of video data, related to normal or usual events, is accessible. Thus, analysis and recognition of anomalous events from huge data are very difficult. In this work, an automated system is proposed to identify and localize anomalies at local level. The proposed work is divided into four steps, namely preprocessing, feature extraction, training of stacked autoencoder and anomaly detection and localization. Preprocessing step removes background from video frames. To capture the dynamic nature of foreground objects, magnitude of optical flow is computed. Deep feature representation is obtained over the raw magnitude of optical flow using stacked autoencoder. Autoencoder extracts high-level structural information from motion magnitudes to distinguish between normal and anomalous behaviors. The performance of proposed approach is experimentally evaluated on standard UCSD and UMN dataset developed for anomaly detection. Result of the proposed system demonstrate its usefulness in anomaly detection and localization compared to existing methods.