Research On Algorithmof Target Detection And Tracking In Traffic Surveillance Video

Intelligent video analysis is a kind of comprehensive technology that gains mutidisplinary research achievement which include fields like Image Processing, Pattern Recognition, Computer Vision and so on. Recently, intelligent monitoring technology has a wide range of applications in transportation, military guidance, medical diagnose and also in public places, residential area, schools. The thesis mainly studies the moving target detection and tracking algorithms in traffic monitor video which has static background and obtained by monocular camera. At last, it gives a accurate result of traffic flow and the specific works are as follows:.For target detection,. merit or demerit of several background modeling methods are analysis and several background updates algorithm are thoroughly discussed firstly, at the basis of which, background subtraction algorithm is selected to detect moving vehicles. Secondly, grayscale foreground image is self-adaption binaryzation by modified Otsu algorithm , which gains a good effect in different illumination condition. Then some necessary processing were done on the binary image of foreground like empty filling, connected domain identification, noise removing and shadow removing.For the description of the target feature, firstly, several feature like interested points, texture, color, contour of targets are elaborated. Secondly, SIFT was selected to describe interested points and then cars with different color or size are tested, which not only shows the extraction process but also verify the reliability especially on cars with more texture. Then, the paper conducts in-depth research on HOG and color histogram and give a new quantitative measures for gradient direction which was proved effective. The paper makes a combination of two kinds of features through one-dimensional vector. At last, the new quantitative measure was proved effective after several experimentations.For targets tracking, Kalman filter is proposed to accomplish effective tracking and then the features are adjusted in the real time on the basis of reliability of different matching features. Thus, SIFT feature is selected on the condition of bad illumination or unreliable color. On the contrary , the integrated color histogram is selected on the condition of better illumination or less texture information. In addition, a counting algorithm is designed for multiple objects, which have a higher accuracy after proved.