Research On Pedestrian Detection Algorithm Based On Multi - Feature Fusion

Pedestrian detection is an important branch of object detection. It is widely used in videomonitoring, analysis, tracking, robotics, entertainment, supervision, intelligent transportatio n,and caring for the elderly and the disabled. Research on pedestrian detection technology hasgreat significance. The current mainstream research direction is based on machine learning,feeding features and labels extracting from training samples to classifiers. In this paper, we startfrom the multi-features combination and ensemble learning to research pedestrian detectionproblem. The main contents are as follows:(1) First, we take Dalal’s HOG+LIBSVM method as an example to introduce detectionmethod based on machine learning. The detection method based on machine learning, mainlyaffected by two factors, the feature and learning method. We compare HOG, MHOG, LBP,SIFT features and in the experiment AdaBoost is used.(2) Then we research on component-based approach. We study the effects of differentsegmentation method on detecting results. First, using one segmentation method evenly splitthe entire training windows to get small pictures, and extract features of small pictures in thesame position of each sample. Then feed them to AdaBoost classifiers to train part classifie rs.These part classifiers make up a global classifier. Using6different segmentation methods,repeat the process to get more global classifiers. For each segmentation method two globalclassifiers are built using HOG and MHOG features. Studies have shown that selecting theappropriate segmentation method to detect is better than feeding the feature extracting fromwhole window to AdaBoost classifier.(3) An ensemble approach for human detection in still images is proposed. In order toachieve better performance, global classifier built by (2) ensemble are used to detect newimages. Experimental results have shown that the proposed method can achieve betterperformance on the INRIA dataset. At the same time, we train global classifiers using6segmentation methods with HOG and MHOG feature combined and classifier ensemble is usedto detect new images. The experiment shows that the feature combination method we used isbetter.