HD Video Pedestrian Detection Research

Pedestrian detection, as the key point of research in the field of security and protection monitoring, also presented one of the most complicated processes. Pedestrians generally being the main targets in the surveillance cameras, pedestrians detection was a pre-processing of other computer vision processing such as abnormal behavior analysis, action recognition processing, and had a crucial impact on the accuracy of the whole system. Since the accurate identification of pedestrian target was difficult, if higher accuracy to be acquired, the speed of the system would be sacrificed. But the speed of the pedestrian detection was particularly important for the processing of HD videos. Now the common problem of pedestrian detection lay in the accuracy and performance that could’t have it both ways.To the questions of the existing pedestrian detection, this paper researched a new speed-up method of the pedestrian detection, and implemented a pedestrian detection system applied in HD video, which was based on the HOG features combined with pattern classification algorithms, and had the overall architecture and several key algorithm opitimized. What’s more, we transplanted the most time-consuming part of system to the GPU, using the powerful parallel computing ability of the GPU to parallel the system to accelerate. The main work of this paper was as follows:1. This paper made use of the moving target detection as the pre-process of pedestrian detection. We narrowed the detection range of the image to be detected and got acceleration through detecting the moving target to obtain the object region which was intersted in. This paper introduced some classical moving target detection algorithm and proposed hybrid biaxial mapping method which was based on background subtraction method, using color information entropy and HSV color dimension to detect and filtered foreground pixels, and then made use of the biaxial mapping method to get moving target areas quickly.2. Based on pattern recognition, pedestrians detection algorithm was generally composed of the image feature and the pattern classifier, so this paper introduced several popular image features and pattern classifiers, and theoretically analyzed a hybrid classifier which utilized SVM classifier as the weak classifier and cascaded by Adaboost to improve the accuracy.3. The algorithm of the HOG features combined with the pattern classifier was the basis of pedestrian detection in this paper, based on which this paper used the moving target detection as pre-process, four class gaussian template to speed up the calculation of HOG feature, adjusted the structure of the SVM classifier to simplify support vector, and finally transplanted the most time-consuming part of system to the GPU, using the powerful parallel computing ability of the GPU to parallel the system to accelerate.The experimental results showed that the method proposed by this paper largely improved the accuracy and efficiency simultaneously.