Haar Characteristics LBP Text Feature For Pedestrian Detection

Pedestrian detection is one of the research hotspots in the field of computer vision, which is widely used in intelligent transportation, intelligent robots, and human motion analysis. With advances in technology, people begin to put it into the emerging areas in aerial image, the victim rescue, and more. Because the pedestrian is easily influenced by shot scene,pose, shelter, dimension or other factors, the pedestrian detection has become the current research difficulty. The current mainstream research method, starting from machine vision, is extracting features from a lot of samples, and then the pedestrian detection problem is converted into a pattern recognition problem by the way of machine learning. The research work and innovation in this paper mainly includes the following parts:Improved HLBP texture feature.Compared to Local Binary Pattern(LBP), the Haar-like LBP(HLBP)effectively reduce the noise by using local statistical way in the pedestrian detection.However, when calculating the eigenvalues in HLBP texture feature, the center point is not involved andmissing the information of center point. In order to solve this problem, we propose an improved HLBP(IHLBP)feature, this method includethe central point with maximum weight. before extracting the IHLBP feature,by using two-dimensional discrete Haar wavelet transform and two level decomposition, we can get three different scales image. Then extracted the three scales image of IHLBP feature and series connected,finally getting Multiple IHLBP feature(MIHLBP).In the INRIA Person data set using support vector machine(SVM)test. The experimental show that this method can effectively improve the recognition rate.MIHLBP texture feature based on HSV color space.A lot of information will be lost when the image transforms from color to grayscale, thus the MIHLBP features extracted from gray image can not deep mining the image texture. In order to solve this problem,a method extracting MIHLBP feature based on HSV color space was proposed. Firstly, images were converted from RGB space to HSV space. Secondly, MIHLBP were extracted from H, S, V channel respectively. Thirdly, the final MIHLBP-HSV feature was obtained by normalizing and seriating three MIHLBP characteristics acquired from second stage. Tests were conducted in INRIA Person dataset using SVM classifier. Experimental results show that this approach achieved higher recognition rate reaching up to 98.5% and had a better performance when compared to HOG, HPG-LBP, LGP-LBP feature.