Pedestrian Detection Based On Head-shoulder In Intelligent Video Surveillance

The intelligent video surveillance technology is a research focus in the field of computer vision. Among it, the object detection is the most basic processing procedure, and whose main purpose is to detect the object and determine its position quickly and accurately. In most instances, the pedestrian gains the most attention in video. Therefore, the main content in this paper is the research of pedestrian detection in intelligent video surveillance system.In view of the shortcomings in traditional pedestrian detection methods, a new method was put forward in this paper, which can not only ensure the high accuracy, but also meet the real-time performance in complex scenario. The main content and innovation points of the algorithm are as follows:(1) Detecting human head-shoulder instead of the whole body. The human head-shoulder was chosen as the object to be detected in this paper, rather than the whole body. There are two benefits by doing so. On one hand, it can reduce the miss rate of the system. Compared with the limbs and torso of the body, the shape of the head-shoulder has better rigidity, and it is more representative for the pedestrian. Even with the situation that the pedestrian is partly blocked in the video image frames, this method still has the potential to reduce the rate of miss detection appropriately. On the other hand, it can improve the real-time performance of the detection system. The information contained in head-shoulder is much less than that in the whole body, so the process of feature extraction is relatively simple. (2) Proposing the LW-PGD method to determine the windows to be tested quickly. Different from the traditional method of sliding window, a new algorithm of positioning head-shoulder windows to be tested was proposed in this paper, which is defined as the LW-PGD method by us. This method takes advantage of the property that the pixel gradient direction of the top of the head has a fixed scope. After finding the eligible pixels in the foreground, the location and size of the human head-shoulders regions are calculated in the original image according to the coordinates of the eligible pixels. Then, the image regions are intercepted and defined as the windows to be tested. This method can locate the windows to be tested of head-shoulder more quickly than the traditional method of sliding window, and reduces the number of windows to be tested to a great extent, with the result of improving the real-time performance of pedestrian detection.(3) Extracting the fusion feature of head-shoulder for detecting. In the pedestrian detection system, the selected feature plays an important role. On one hand, it affects the accuracy of target classification and recognition. On the other hand, its extraction process determines the efficiency of the whole system. Due to that the Histogram of Oriented Gradient feature can effectively extract the edge and contour information of the human head-shoulder and the HSV feature can quickly distinguish the color of face and hair. What’s more, proper fusion feature has better detection effect than single feature. So the combined feature of HOG and HSV was used to detect human head-shoulders in this paper.After elaborating the above innovative method, this paper got the classification decision unit by training the MIT and INRIA database with support vector machine classifier, and did experiments with the test data, which is the actual monitoring video from Britain’s Gatwick airport. The experiment results show that the innovative algorithm and method in this paper have good feasibility and effectiveness.