Research Of Pedestrian Detection Algorithms Based On Joint Features Of HOG And LBP

Among various pedestrian detection algorithms,due to its excellent performance in detection accuracy and processing occlusion problems,HOG-LBP pedestrian detection algorithm has attracted great attention of many scholars.However,HOG feature dimension is very high and it contains too much redundant information,because of its gradient characteristics.At the same time,HOG feature cannot describe texture features of images and cannot describe gradient spatial characteristics well.Due to its binary coding strategy,LBP feature's robustness to rotation,illumination and noise need to be improved.And the calculation process of nonlinear kernel SVM classifier is complex.It is necessary to improve the real-time detection.Therefore,based on the framework of HOG-LBP algorithm,this thesis aims at the existing problems of HOG and LBP features respectively,and combines the new improved features.At the same time,based on the classifier with lower complexity and better real-time performance,this thesis studies the pedestrian detection method.The following are the main contents of this thesis:Firstly,this thesis gives a detailed introduction of pedestrian detection technology from three aspects: research background,research significance and research status.Then,this thesis introduces and summarizes some related technologies under the framework of pedestrian detection algorithm: feature extraction algorithm,classifier pedestrian database and algorithmic evaluation method.Then,aiming at the rotation invariance and scale invariance of LBP feature need to be improved,and the high dimension of HOG feature contains too much redundant information and does not describe the texture features of images,this thesis studies the CLBC feature and the original HOG feature,proposes to extract the HOG features based on texture feature map.And through using feature serial fusion and PCA dimensionality reduction,this thesis constructs the final eigenvectorof image which is more robust to rotation and scale and has higher texture information utilization and low dimension.Combined with HIKSVM classifier,a multi-feature fusion pedestrian detection algorithm based on reduced-dimension HOG is designed.To verify the effectiveness of feature extraction and feature dimension reduction,three different feature fusion experiments are designed in this thesis.The final experimental results show that the multi-feature fusion pedestrian detection algorithm based on reduced-dimension HOG has overall superiority.Next,aiming at the shortcomings of poor description of gradient spatial properties by HOG feature and poor robustness of LBP feature to illumination and noise,this thesis studies CoHOG feature and RLBP feature on the basis of feature fusion.In this way,a classifier with high real-time cascading multi-level feature weak classifiers is constructed,and a pedestrian detection algorithm based on cascaded feature classifier is designed.To verify the validity of the extracted features and classifiers,three different feature fusion experiments are designed in this thesis.The final experimental results show that the pedestrian detection algorithm based on cascade feature classifier has overall superiority.Finally,this thesissummaries the main contents,points out the shortcomings of algorithms,and prospects the future research directions.