Research And Implementation On Pedestrian Detection Technology

With the development of intelligent video surveillance, driver assistant, city security, intelligent robots, pedestrian detection technology become a focus in the field of machine vision. But the pedestrians are not rigid objects, with a variety of posture. For complex background and other objects bring, the pedestrian detection has become a difficult problem. The main work of thesis are as follows:(1) This thesis studies the classical HOG feature and compared the DET curve of HOG feature in different coordinates respectively. Based on HOG characteristics of different coordinates(the dimension decline), respectively, in the linear kernel function and Gauss kernel function, to experiment. Experiments show that the fusion of HOG in the linear kernel function, regardless of whether the PCA dimension reduction, have shown a good effect, but in the Gauss kernel function, the effect is not obvious.(2) This thesis introduces the deformable component model(DPM model) which is widely concerned, and the HOG feature of DPM model have been greatly improved, which has a better detection rate.(3) Pedestrian re-identification has gradually became a hot issue. This thesis uses the spatial color histogram of different segmentation to experiment, it show that the horizontal segmentation has a better recognition rate.(4) The pedestrian detection and tracking system, based on SVM+HOG and the Gauss background modeling module, is realized by VC++.