Research On Dynamic Target Detection Algorithms In Intelligent Visual Surveillance

As the most basic and critical step in intelligent video surveillance, dynamic target detection has a direct impact on the applicability of an intelligent video surveillance system. It is a closely-watched front subject in the field of computer vision, and has achieved fruitful achievements. However, due to the complexity of monitoring scene, to extract accurate foreground information from the video is still facing with many challenges. Research on how to improve the effect of dynamic target detection has important theoretical significance and extensive application value.In this thesis, the problem of poor performance of dynamic target detection in challenging situations with shadow, ghost, local illumination change, camera jitter and so on is studied deeply. Embarking from different research methods, two improved dynamic target detection algorithms are proposed. Through experiments on indoor and outdoor test video, and compared with other algorithms, the effectiveness of the proposed algorithm is verified. The main work of the thesis is as follows:1. Aiming at the problem that ViBe algorithm has a poor performance on restraining ghost and removing shadow, an improved ViBe algorithm based on double background model is proposed. Caching pre-K-frame image of video, the method of random sampling and improved average are used to build two background models:foreground detection model and shadow detection model which are used for foreground detection and shadow removal respectively. In foreground detection phase, the range of sample extraction is expanded to improve foreground detection model’s reliability, then the ghost area can be restrained; Compare with the minimum matching number can reduce the amount of calculation; Foreground detection model is updated by replacing the false sample. In shadow removal phase, we compare interest foreground area obtained from foreground detection with shadow detection model, combine the fast normalized cross correlation function with LBP texture feature information to detect and eliminate the shadow, then the detection effect can be enhanced. Experimental results demonstrate that the improved algorithm can effectively restrain ghost, remove shadow interference, and make the test results more accurately.2. The traditional Gauss Mixed Model (GMM) only uses color feature as the basis for dynamic target detection, which often causes problems like incomplete detection result, poor detection effect on local illumination change situations, be sensitive to camera jitter and so on. Therefore, a method of dynamic target detection based on region combination feature is proposed in this thesis. Firstly, texture, color and position characteristics of the region are combined into a unified feature vector. Secondly, the combination feature of each pixel is modeled by GMM. Finally, the moving objects are detected by background subtraction method. Experimental results demonstrate that the improved algorithm can improve the detection effect and meet the real-time requirements.