Abnormal Events and Behavior Detection in Crowd Scenes Based on Deep Learning and Neighborhood Component Analysis Feature Selection

Abstract

In the last few years, surveillance cameras have been massively distributed both indoors and outdoors in public, due to security concerns creating a need to monitor unexpected actions or activities in a scene. An intelligent automated approach is highly required to detect anomalies from the scene as well, to save the time and cost required by laborers to detect the anomalies manually from monitor screens. In this research, we propose a deep learning-based approach to detect abnormal events and behavior from surveillance videos in crowd scenes. Thus, by using the keyframe extractor to localize the keyframes that contain important information from video frames. The selected keyframes are used to compute the optical flow values of magnitudes, orientations, and velocities of each keyframe to generate several 2D templates. Then, the obtained 2D templates are supplied to a pre-trained model ‘AlexNet’ to extract high-level features. The Neighborhood Component Analysis (NCA) feature selection method is applied to select the appropriate features, then use these features to generate a classification model via Support Vector Machine (SVM) classifier. Results are evaluated on several public datasets, along with a new dataset that we built it that contains different videos covering abnormal events and behaviors. The obtained results proved that the proposed method outperforms other methods.