The Research Of Pedestrian Detection System For Images

The great commercial value of pedestrian detection system in intelligent monitoring system and intelligent traffic system makes it a very hot topic of the field of target detection. The diversity and non-rigid shape of the pedestrians significantly increase the difficulty of effective pedestrian detection. Both video based pedestrian detection system and image based pedestrian detection are studied in the recent years. The paper studied the pedestrian detection system for images. And the study focused on increasing the detecting accuracy and reducing the detecting time. In this paper, the pedestrian detection system is divided into two parts:classifier training and pedestrian detecting.In the training phase of the classifier, the histogram of oriented gradients (HOG) feature is introduced in detail for no single feature has been shown to outperform HOG. But the high-dimensional feature space of the HOG leads to a high computational cost and bw detecting speed of the classifier. In this paper, we proposed a modified HOG cascaded with the principal component analysis (PCA) algorithm. Redundant dimensions are eliminated from the HOG feature vector by the introduction of PCA, leading to improved detection accuracy and speed of the classifier.In the detection phase, the sliding-window method is employed in pedestrian detection for single images. This approach scans the whole image. Each window is classified as pedestrian or non-pedestrian. The approach is simple and flexible. But it is time-consuming as a large number of non-pedestrian windows need to be classified. In order to improve the efficiency and accuracy of pedestrian detection in the detection phase, we proposed an ROIs segmentation method based sliding window detecting algorithm to remove parts of the background area, which reduced the number of non-pedestrian windows for detection, leading to increased detecting speed and accuracy in detection phase.In this paper, the simulation experiments of the pedestrian classification and detection in still images are implemented on a PC, which runs MATLAB code with LIBSVM library. The experimental results show that redundant dimensions are eliminated from the HOG feature vector by the introduction of PCA, leading to improved detection accuracy and speed. The image segmentation reduces the number of detection windows, resulting in a great reduction in detection time of a single image. In summary, our method yields good pedestrian detecting performance.