Anomalous event detection and localization in dense crowd scenes

Recognizing and localizing anomalous events in crowd scenes is a challenging problem that has attracted the attention of researchers in computer vision. Surveillance cameras record scenes that require an automated examination to identify anomalous events. Existing approaches in the field have utilized different feature descriptors, modeling methods, and recognition strategies to accurately and efficiently detect anomalies in the scene. Existing techniques in the field have focused mainly on performing global frame-level identification of abnormal events. Only a small number of studies have considered locating abnormal action in the frame. Proposed methods are also often evaluated on scenes that contain a sparse number of individuals performing abnormal and normal staged acts. This research aims to detect and locate anomalies in a structured and unstructured dense crowd scene. The proposed model first detects moving objects and individuals in the scene using a deep convolutional neural network and tracks objects and individuals using spatial and temporal features. Then, spatial-temporal features are extracted from consecutive frames of interest points. The extracted features include the histogram of optical flow, velocity and direction of moving objects, and other features that can indicate sudden motion change. A support vector machine model is then used to classify abnormal events into one of seven classes. The proposed methodology is evaluated on Hajj2 dataset that has 18 videos and 7 different types of abnormal events.