The Research On Pedestrian Detection Based On Gradient And Texture Feature Selection

Pedestrian detection is a key application fields of computer vision, which has received extensive attention of people. Quickly and accurately pedestrian detecting from the static image currently has good application prospect.In this thesis, combining several primary feature extraction method, using SVM (Support vector machine, lib-SVM) as classifier, we studied how to classify static samples of pedestrians. In order to improve the detection rate and reduce the waste of time, we further extract the secondary characteristics (using such as singular value decomposition, PCA) from the primary feature of sample, and using the INRIA and Daimler databases to verify the effect of the method. Through comparing processing time and error detection rate, we also analyzed the performance of each features. This article take the gradient histogram feature and texture feature as primary characteristics, the process is listed as follows:Firstly, extracting the gradient histogram features (Histograms of Oriented Gradients, HOG) and texture characteristics of the Local Binary Pattern (LBP) as primary characteristics of the samples. Combining the two primary features as a new composite characteristics (HOG-LBP).The experimental results show that the combination of characteristics can effectively improve detection rate.Secondly, using PCA on the two primary characteristics to get the new low dimensions features. Then using K-SVD to learning the over complete dictionary, and get two primary features of sparse expression. Experimental results show that after the dimension reduction of secondary characteristics, consumption of time is greatly reduced. At the same time, detection rate is not much loss.Thirdly, in order to further explore the combination effect of gradient features and texture features, we combine the two secondary characteristics extracted using PCA as new features. Will likewise, K-SVD is also used to extract two kinds of new features. Experimental results show that the combination of secondary characteristics, compared with the combination of the primary characteristics, can greatly reduce the testing timed, At the same time, the error inspection rate have a certain degree of reduction.Finally, based on the gradient histogram feature, the two features in the low dimensional space are combined. Do the same experiment using texture features, experimental results show that the composite characteristics can effectively improve the detection rate.