Anomaly detection in video using two-part sparse dictionary in 170 FPS

Abstract

Automatic supervision of crowd behavior with the aim of detecting abnormal movements has become important in the field of public places' security and protection. Crowd congestion is not fixed in public places, thus we need an algorithm that can perform powerfully in high and low crowd congestions. Generally, there are two different methods for analyzing crowd behavior: the method which is based on tracking moving objects in which every person or moving object is traced separately in the scene, and holistic method which investigates the crowd as a whole. In this paper, our goal is to identify the abnormal behaviors in public places through applying holistic methods (only pedestrians are presents in these places). Crowd behavior is modeled as a collection of basis; identifying and locating the abnormal behaviors are done by Sparse Coding. In real videos, the frames are perspective, and in this research we propose a solution to this problem based on the dataset we use; naming two-part sparse dictionary. General training features are saved in a two-part dictionary, and test movements are analyzed through rebuilding the extracted features from the test video based on the available dictionary which is formed in an unsupervised way using sparse combinations. High errors on this stage show the lack of suitable rebuilding of the test video based on available behaviors in the dictionary, so the algorithm detects and locates abnormal behaviors. Proposed algorithm is performed on UCSD datasets, ROC curve is calculated and EER values are 0.29 and 0.35 respectively. The results show the ability of the proposed algorithm for real time detection of abnormal behaviors.