Multi-feature Extraction And Vehicle Identification Research Of Vehicle Objects In Monocular Video

Research on multi-feature extraction and vehicle type recognition in monocular video is an important part of intelligent transportation systems research.This thesis does research on multi-feature extraction and vehicle type recognition in monocular video,the main contents and innovations are as follows:1.Establishment of standard models of feature model library.Establishing the standard models of feature model library is a basic work of this research.The library is constructed from original physical features,vehicle photos and structure features of more than 2000 vehicle types,providing a basis for vehicle type recognition.2.Monocular camera calibration.According to the national road line standard,based on the correspondence between the corner pixel coordinates of video image lane and the three-dimensional coordinates,Zhang Zhengyou camera calibration algorithm is used to realize a still camera calibration.3.The three-dimensional vehicle geometry feature extraction and vehicle type recognition.Kmeans algorithm is used to find the cluster centers of four vehicle types in the standard models of feature model library.Firstly the deformable three-dimensional vehicle model is built.Secondly the three-dimensional model is projected onto the image plane according to the monocular camera calibration results,and then the matching evaluation function between vehicle model projection and 2D vehicle object in the image plane is constructed.Thirdly,the estimation of distribution algorithm is used for optimization on the matching evaluation function to find the optimal solution corresponding to the geometric parameters of the target vehicle.Vehicle type recognition is achieved through these geometric parameters,and the average recognition rate is 89.2%.4.Vehicle face feature extraction and recognition.An improved adaboost algorithm is used to detect vehicle face.Eigenfaces,FisherFaces and Extended LBP methods from human face recognition are used for vehicle face feature extraction and recognition,vehicle type recognition is achieved by vehicle face recognition and the accuracy rate of vehicle face recognition is 70.83%.Multi-feature extraction includes three-dimensional vehicle geometry feature extraction and vehicle face feature extraction.Vehicle type recognition includes rough recognition through three-dimensional geometric feature and finer recognition through vehicle face recognition in each category.