Research Of The Pedestrian Detection Algrithm Based On The Hybrid Fearures

With the increasingly rich and colorful of the social life, the pedestrian detection in public place is required intensively such as airport, station, market, residential community and so on. Pedestrian detection aims at detecting pedestrian using calculating power of computers. Because of the complexity of the background, light intesity, pedestrian clothing decorations and pedestrian postures, detecting pedestrian accurately and quickly is still a hot issue of the current research.In this area, one of the famous algrithms uses feature descriptor to extract the pedestrian feature and machine learning to train classifier, then identify pedestrian by classifier which is adopted to detect pedestrian in this thesis.This thesis focuses on feature descriptor and classifier. The histogram of oriented gradient proposed by Dalal and Triggs which has strong description ability is used through the comparison of all kinds of feature descriptors. The histogram of oriented gradient has some shortcomings which includes large amount of calculation, high dimension of features, no use of the image texture information and so on. In order to reduce dimension and take into account of the texture information of the image, a hybrid feature which combines the pyramid model and the entropy information. The support vector machine (SVM) is used in this thesis because of its fast classification rate and simple structure. Most of the common used SVM is linear kernel SVM which have a faster classification rate but a lower detection rate. Nonlinear kernel SVM has a better detection rate but a slower classification rate. Considering these issues, a nonlinear kernel called intersective kernel is adopted to realize SVM. Then the additive kernel SVM is realized on the basis of the intersective kernel SVM. These two SVMs can maintain the strong calssification ability of the nonlinear kernel and the classification speed of linear kernel. Finally, hybrid fearures based on combination of pyramid histogram of oriented gradient and entropy fearures combines respectively with intersection kernel SVM and addition kernel SVM to increase the performance of pedestrian detection algrithm.The classifier of the algrithm is trained in INRIA pedestrian library to test the algorithm performance through the comparison of DET and ROC curves. Experimental results show that the pedestrian detection proposed in this thesis has a higher detection rate and a faster detection speed, even under the conditions of complex background, pedestrians with more gestures and small block.