Research On Road Scene People Detection Based On HOG And LBP Feature

In recent years,with the number of vehicle increasing in China,the traffic has become more and more crowded,and the number of traffic accident has risen rapidly.Researches show that the traffic accident victims mostly are pedestrians.So the need of developing an assisted-driving system that can avoid pedestrians is urgent.As one of the core technologies,people detection has attracted more and more attention.Aiming at the deficiency of single feature in people detection,a method based on feature fusion and data dimension reduction is designed to improve the accuracy and the speed of detection.The image has preprocessed to prepare for the following work.The sample size is normalized.The denoising results of several filtering methods are compared.The road image is equalized by histogram.The results show that the median filter has the best effect on salt pepper noise and Gaussian noise.The contrast of the original image is improved by histogram equalization.A classification algorithm based on Histograms of Oriented Gradients(HOG)and Support Vector Machine(SVM)is designed.The HOG feature of the sample is extracted.The SVM classifier is trained based on four kernels.The experiment compares their classification accuracy and speed.The results show that SVM classifier based on linear kernel has the highest accuracy and the fastest speed.A feature fusion algorithm is designed to improve the shortage of single HOG feature.The uniform pattern Local Binary Pattern(LBP)feature which has the best performance is selected to serial fusion with HOG.In order to solve the problem of high feature dimension of the fused feature of HOG-LBP,Principal Component Analysis(PCA)was used to reduce the dimension.The experiment shows that the classification accuracy of the 1700 dimensional HOG-LBP-PCA feature is higher than the single HOG feature.The classification speed is higher than HOG-LBP which has not been dimensional reduced.Finally,the multiscale detection algorithm based on image pyramid and the window fusion algorithm based on greedy strategy has been studied.An experimental platform is built based on industrial camera and Open CV.The results of HOG-LBP-PCA and single HOG are compared.The results show that the improved feature has better performance in accuracy,recall and Receiver Operating Characteristic(ROC)curve.The above content shows that the algorithm designed in this paper can effectively improve the detection accuracy and optimize the detection speed.And it shows a certain reliability on the experimental platform.