Pedestrian Detection Method Based On Sparse Coding

Pedestrian detection is extensive applied to intelligent transportation,smart car driver assistance systems,intelligent monitoring,unmanned aerial vehicles detection and so on.Detecting pedestrians is still a huge challenge in road environments because of the shelters,illumination changes,gestures varieties and complex backgrounds.Currently,the popular method of pedestrian detection is the combination of feature extraction and classifier.Firstly,feature descriptor are extracted from the training samples,then a classifier model is trained by using these features,and utilized to detect the pedestrian in an image.Feature extraction is greatly influenced by training samples,so this thesis presents a method of cropping the training samples.The best window size are set 128× 52 by analyzing existing pedestrian detection window aspect ratio,this method is to retain only pedestrian information as far as possible and reduce background information to express pedestrian features.Detecting pedestrians in the road environments require a high accuracy rate and real-time.In order to improve the accuracy rate of pedestrian detection,the feature descriptor of the histogram of sparse code(HSC)is chosen,a fact that the size of the descriptor block,dictionary atoms,dictionary size and other parameters have a great effects on the expression of sample information,and ultimately affect classification results of the support vector machine(SVM)is found by analyzing the HSC,then a better performance of SVM classifier based on HSC is designed.At the same time,the Real Adaboost classifier is used to design pedestrian detection system based on HSC because the HSC + SVM method in real-time performance is not high.Two kinds of pedestrian detection methods are simulated based on MATLAB platform,the method based on SVM has 13% reduction in false positives rate and 4% improvement in detection accuracy than the gradient direction histogram method.And based on Real Adaboost method than SVM method have reduced more than 70% in time-consuming of detection.So the method based on Real Adaboost and HSC has a very good practical application prospect.