Anomaly Detection for Videos of Crowded Scenes based on Optical Flow Information

• Published in: 2018 3rd International Conference on Advanced Robotics and Mechatronics (ICARM) pp. 869 - 879
• Main Authors: Gao, Yuan, Cao, Hui, Xu, Shuo, Zhang, Jie, Ning, Xiaolin, Ma, Xin, Samuel Adu, Bediako
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.07.2018
• Subjects: Anomaly detection; Atom optics; Decision trees; Feature extraction; Integrated optics; optical flow information; spatial-temporal feature; Vegetation; Videos; weighted random forests
• DOI: 10.1109/ICARM.2018.8610722

Anomaly detection for videos of crowded scenes is important to social security. For crowded scenes, there are imbalanced statistical distributions for the frequency of the appearance of the anomalies and the frequency of the appearance of the normal behaviors. In this paper, an anomaly detection method for videos of crowded scenes based on optical flow information is proposed. The proposed method realizes anomaly detection in two stages. In the first stage, the videos are divided into small atoms in spatial-temporal dimensions. Here the atoms consist of the regions in the same position of the continuous frames in the videos. In each atom, spatial-temporal features extracted with the help of optical flow information are used to illustrate the spatial and temporal relationships between motion patterns. In the second stages, a weighted random forests classifier is modeled. Two weight coefficients are assigned according to the frequency of the appearance of the anomalies and the appearance of the normal behaviors in the videos. The weighted random forests classifier consists of the decision trees and the atoms are classified as anomalies or normal behaviors according to the classification result from the trees. In the process of detecting anomalies, the atoms labeled as anomalies are marked in the videos. Thus anomaly detection has been finished. The experiments are conducted on a public dataset called UCSD Ped1 and the results from experiments show that the proposed method could detect the anomalies in videos.