Video Anomaly Detection In Crowded Scenes Based On Genetic Programming

Video anomaly detection is an extremely significant and challenging work. Not only has it a great value in the daily monitoring system, but also it is the research focus in the field of computer vision. The real-life surveillance scenes are mostly crowded scenes, such as railway stations, shopping malls, supermarkets, and these scenes have many moving objects, which have caused great difficulties for anomaly detection. How to accurately detect anomalies in crowded scenes is significant. In recent years, genetic programming is becoming a hot research direction in the field of evolutionary computing, has been applied to various fields, among which, computer vision is the most important one. However, there are few researchers about applying genetic programming to video anomaly detection. The work of this thesis is applying genetic programming to video anomaly detection. The main work is as follows:● Propose a classification method of genetic programming based on dynamical selection. In this paper, we analyze the problem of applying genetic program-ming to classification and propose a classification method of genetic program-ming based on dynamical selection. To consider the operational efficiency of the algorithm, we propose the dynamic subset selection algorithm, so as to reduce run time of the algorithm and to improve operational efficiency in the case of similar classification performance. To consider the classification performance, we propose the dynamic range selection algorithm based on static range selec-tion algorithm, which can improve the classification precision.● Propose a method for global anomaly detection based on genetic program-ming. Considering the global abnormal behaviors in crowded scenes, we pro-pose a method for global anomaly detection based on genetic programming. Firstly, we propose a new frame-level feature extraction method called Multi-Dimensional Histogram of Optical Flow, which can extract features including motion information in detail. Secondly, regarding a frame as a sample, if the frame contains abnormal behaviors, the corresponding sample is abnormal, oth-erwise is normal. After being trained, genetic programming can generate a clas-sifier for anomaly prediction. For most of the global abnormal behaviors are the moving objects in a scene, we only predict frames which contain anomalies.● Propose a method for local anomaly detection based on genetic program-ming. For local abnormal behaviors in crowded scenes, we propose a method for local anomaly detection based on genetic programming. Firstly, we propose a new frame-level feature extraction method called Multi-Frame LBP Difference and Weighted Multi-Frame LBP Difference. Compared with the traditional fea-ture extraction methods such as optical flow, the proposed methods do not re-quire complicated pre-processing steps so that it can extract features including temporal-spatial information quickly. Secondly, regarding image block of frame as a sample, if the image block contains abnormal behaviors, the corresponding sample is abnormal, otherwise is normal. After being trained, genetic program-ming can generate a classifier. Finally, with the classifier and voting mechanism, we can detect local abnormal behaviors and localize anomalies in real-time, so that we can find the abnormal regions intuitively.