| Pedestrian detection is a fundamental problem in the field of intelligent video analys is, and it is not only significant in research but also useful in practical applications. Pedestrians, compared with the vehic les, are not rigid, and have complicate gestures. So the detection process is susceptible to be degraded by different positions or actions. On the other hand, the complexity of various pedestrian detection algorithms has made the limitations of the hardware platform to do pedestrian detection, so transplanting pedestrian detection algorithm on the portable embedded platforms is also a challenge.This paper describes the basis of the principles and processes in relevant pedestrian detection algorithms. After that, proposed two different kinds of pedestrian detection strategy which are based on the foreground segmentation and feature classifier for two different occasions: indoor and outdoor scenes.About the foreground segmentation, in order to meet the needs of integrity and robustness in pedestrian detection, proposed an enhancement algorithms of foreground segmentation by motion-based contour. Combined with superpixels and Otsu algorithm, improving the result of the foreground segmentation in low contrast areas. The point is to get the key region where pedestrian passed by, the key region consists of superpixels, then calculating the adaptive threshold for foreground segmentation by Otsu algorithm.In terms of feature classification, briefly introduced the HAAR, HOG, LBP. Due to the common pedestrian detection operator HOG and LBP has some shortcomings, further, using and learning the Centrist operator(based on the gradient and texture) to improve such problems, and giving the performance of the cascade testsThen, according to the characteristics of DM8168 embedded platform, transplanting the pedestrian detection algorithm which used the Centrist feature, and adding main color extraction module of pedestrian, finally, completing the optimization works. |
PDF Full Text Request