Research On Technology Of Pedestrian Detection And Tracking For Vehicular Infrared Night-Vision System

With the improvement of production and living standards, the number of car ownership has increased greatly, and more and more traffic accidents have occurred.Vehicular infrared night-vision system is a part of advanced driver assistant system. It can be used both in day and night, rain and snow, fog and other inclement weather. And with the improvement of the process level, the difficulty and cost of vehicular infrared night-vision system's hardware production is decreasing. Based on the above advantages,vehicular infrared night-vision system has become the lead of the auxiliary driving system. However, most of the current vehicular infrared night-vision system just simply collect the image in front of vehicle, and then display in screen. The image has not been further processed, so the warning effect is limited. Based on this reality, this thesis studies the pedestrian detection and tracking on vehicular infrared night-vision's infrared image. The main research work is as follows:(1) Illustrated the classification and principle of infrared night-vision technology,summarized the indexes and characteristics of thermal imaging technology. In addition,the characteristics of the infrared image are analyzed from the three aspects of gray histogram, noise and resolution.(2) According to the results of infrared image characteristic analysis, the method of vehicle infrared image preprocessing is established. Firstly, the median filter is used to filter noise and histogram equalization is used to enhance the contrast of infrared image.Then, use the improved Otsu algorithm to segment the image and mathematical morphological processing to deal with the hole. Finally, mark and filter the connected domain. After the above treatment, we got the region of rough pedestrian area that is region of interests. This step can effectively improve the real-time performance of pedestrian detection algorithm.(3) Describe and demonstrate from the theoretical framework to simulation for in statistical classification algorithm of pedestrian recognition. The performance of support vector machines depends on the size of the kernel parameters. According to this feature,we combine the support vector machine and Adaboost algorithm to design the classifier.The classifier is used to judge the HOG features extracted from the region of interests after normalized to specified size to determine whether it is a pedestrian. Experiments indicate that the classifier designed in this thesis can adaptively adjust the parameters of support vector machine, which solves the problem of using one fixed kernel parameter,and also deals with the balance between precision and complexity. Thus, it has a better classification effect than the single support vector machine.(4) In the stage of pedestrian tracking, this thesis analyzed the principle and method of Mean Shift tracking algorithm and Kalman filter tracking algorithm. After analyzed the advantages and disadvantages of these two tracking methods, a tracking method combining Kalman filter with adaptive window of Mean Shift is proposed. The tracking algorithm designed in this thesis has good tracking effect in single pedestrian,pedestrian deformation and simple occlusion scene.