ABNORMAL CROWD BEHAVIOUR DETECTION BASED ON DEEP LEARNING AND SPARSE REPRESENTATION

Zhendi Gai, Dongmei Liu, Faliang Chang, and Nanjun Li

Keywords

Abnormal behaviour detection, histogram of optical flow, saliency information, deep learning, sparse representation

Abstract

This article proposes a new method combining improved principal component analysis network (PCANet) and sparse representation for abnormal crowd behaviour detection on the consideration that “abnormal events are not “normal events. First, the histogram of optical flow (OF) is extracted as the temporal features according to the Horn–Schunck OF algorithm, and the saliency information of the video frame is calculated as spatial features. Then, the improved PCANet network is used to extract the high-level features from these low-level features, and a sparse dictionary of normal crowd behaviour is built based on the high-level features. Finally, the high-level features of the video frame to be tested are sparsely reconstructed to get the reconstruction cost, and frames are classified as normal and abnormal by comparing to the reconstruction cost. Experiments on the UMN dataset show that the proposed method captures the high-level features of the crowd behaviour effectively and outperforms other state-of-the-art methods.

Important Links:



Go Back