The Research Of Some Technologies For Human Abnormal Behavior Detection In Certain Scenarios

The techniques for human abnormal behavior detection are the important research content in the intelligent monitoring system, the primary ways to improve the intellectuality and have the widespread application prospect in the public safety and the economy. So many scholars commit themselves to the human abnormal behavior recognition research and have achieved some results. But the human abnormal behavior detection technology has yet to mature and has many challenges. This paper makes a deep analysis in research status of the human abnormal behavior detection. It mainly aims to abnormal behaviors in specified monitoring scenes, such as loitering, bending, tripping and fighting. The study is to detect abnormal behaviors by being based the technology of tracking algorithm and human behavior recognition, and it increases the efficiency of the abnormal behavior detection. The main works are given as follows.(1) The modified Differential Structural Similarity (DSSIM) is presented in this paper. The algorithm can effectively detect abnormal behavior of pedestrians. Because of the real-time and robust to varying light conditions, the tracking algorithm of DSSIM is chosen in the paper. Based the DSSIM, the luminance computational method is modified and it is combined with the GM(1,1) gray prediction model. It improves the tracking stability and robust to the shelter. So it can detects the loitering. By making affine transformation of the tracking window during the tracking process, the tracking window can be rotated to detect bending and tripping. The experiment results show the abnormal behavior of pedestrian such as loitering, bending and tripping in the simple scenario can be effectively detected.(2) This paper presents a fighting detection algorithm based on hybrid features. By incorporating histograms of quadruples optical flow gradient into local spatio-temporal features, an effective coding scheme called Locality-constrained Linear Coding (LLC) is used for the features. Our results confirm that the fighting detector only with linear support vector machine for sequence classification performs well. Compared with the fighting detector using Bag-of-Words(BOW) framework and local spatio-temporal features, the detector proposed in the paper improves the recognition accuracy.(3) By adopting the algorithm presented in the paper, we develop a system of human abnormal behavior detection with high quality performance. The system based the MFC frame and the OPENCV library is not only simple and practical but also can detect the abnormal behavior effectively.