Method Of The Pedestrian Detection Research In Mixed Traffic

The pedestrian detection of mixed traffic is a technique of extracting the region of pedestrians in front of vehicle from images or videos with a computer in the way as human. This technique is another research hot spot after face detection in the field of computer vision. It is the key technique of intelligent driving and intelligent traffic monitoring in the complicated environment under the urban mixed traffic to detect and track the moving pedestrians. And it also plays an important role in promoting the development of intelligent vehicle and ensuring the security of the urban traffic.In this paper, the methods of pedestrian detection from video sequences based on computer vision are researched in order to ensure the safety of the pedestrian in the urban mixed traffic. It mainly considers pedestrian detection methods under two cases:the stationary camera and the moving camera.In the case of stationary camera, make a research of several algorithm of background modeling, including average background model, the background modeling algorithm based on gaussian mixture model, the background modeling algorithm based on codebook. Contrasted the performance of the algorithms, the background modeling algorithm based on codebook is researched. First, extract the foreground region of movement in front of the camera, and then determine whether there is a pedestrian in the foreground region by using some attributes of standing pedestrians. The method is of fast speed, but low accuracy. And then, the pedestrian detection based on statistical learning classification method in the case of moving camera is researched in the paper.In the case of moving camera, the Haar-like feature in statistical learning classification is researched deeply first. Besides the existing rectangular features, several rectangular features which can represent body parts are added, and the integral image method is adopted to calculate the rectangular features, which can accelerate the training and detection speed. The cascade classifier is trained based on AdaBoost algorithm. Although the training time of this method is long, the detection time turns much short. The result shows that this method can achieve good detection effect, and basically satisfy the real-time requirements. The method of histograms of oriented gradients (HOG) is analyzed, which is one of the best ways of pedestrian detection. HOG features are obtained through extracting the direction and the gradient in the edge of local area. Then the HOG vector is introduced to a linear SVM to train for an optimal classifying hyperplane to distinguish the pedestrian from other objects. The block size of traditional HOG feature is fixed, which can obtain limited features. So variable block size is adopted, and an integral vector diagram called IHOG which is similar to rectangular features is used to calculate the HOG features. This method can accelerate the training and the detection speed. It also can enhance the detection precision.Finally, a method combining the rectangular features with the HOG features is proposed, which integrates the fast speed of the rectangular features and the high accuracy of the HOG features. First, detect roughly with a classifier based the rectangular features with few layer cascade, at the time, pedestrians will be detected with plenty of false positive. And then, detect accurately with HOG features, retaining the pedestrians, rejecting the false positive from the above classifier. Each frame needn't to be searched globle by the HOG classifier, instead, the ranges to be searched are reduced to suspicious pedestrian area from the above classifier. HOG accuracy can be obtained by the combined method, and it is more than two times the speed of HOG features. Simulation results show that the combined method can speed up the detection with high accuracy.