Global Abnormal Event Detection Based on Compact Coefficient Low-Rank Dictionary Learning

Abstract

In this paper, an approach to detect global abnormal events is presented, which is based on a compact coefficient low-rank dictionary learning (CCLRDL) algorithm. Similar with sparse representation, the aim of the approach is to achieve the reconstruction coefficients over the normal bases. First of all, the histogram of maximal optical flow projection (HMOFP) is extracted from a set of normal training frames to describe the movements of the crowd. Secondly, after a process of selecting the training samples, the inexact augmented Lagrange multiplier (ALM) algorithm is utilized to obtain a low-rank dictionary. And then, by using the ALM algorithm the reconstruction coefficients of testing samples are acquired. Finally, reconstruction cost (RC) is introduced to detect whether a frame is normal or not. The experiment results on the well-known UMN dataset and the comparisons to the most popular methods show our algorithm is promising.