The Research On Vehicle Type Classification And Posture Recognition Of Vehicles Marked As Targets

With the increase in the number of cars and the frequent occurrence of traffic accidents in China,the increasing demand for traffic safety has made researchers attach great importance to the field of automobile safety.In recent years,the rapid development of vehicle intelligent driving and active safety technology has played an important role in improving the safety of vehicles.Through active sensor technology,it can recognize the pose and type of the vehicle.Vehicle type classification is the key task of intelligent transportation systems,and vehicle posture detection is an important technology for assisted driving systems in the field of intelligent drive.Statistical research on traffic accidents shows that in the case of a collision between two vehicles,the vehicle type and the relative motion posture of the two vehicles have an important impact on the collision damage of the vehicle,which in turn has a great impact on the damage of the occupants in the vehicle.By identifying the type and posture of the vehicle,better protection of the occupants in the vehicle can be achieved,which has important significance.Based on this,we use the machine vision technology which commonly used in intelligent driving systems,combining the YOLO algorithm,B-CNN algorithm,improved HOG + LBP feature extraction and support vector machine SVM and other methods.First select the target vehicle frame,and then classify the vehicle type and posture of the vehicle in the target frame.The specific focuses and innovations of the research are as follows:1.Vehicle detection.We studied the YOLOv3 algorithm,re-clustered the target candidate boxes using the K-means clustering method to obtain the optimal number of candidate boxes,and used the vehicle picture to train the feature extraction network again to obtain the vehicle recognition model of YOLOv3.2.Classification of vehicle type.We selected the BCNN network which has high classification accuracy and used it to classify vehicles of different sizes and shapes.Aiming at the confusing situation of BCNN models at different angles,we propose a multi-angle feature fusion classification method.It uses BCNN’s convolutional network to extract vehicle features at different angles,and then linearly fuse the extracted multi-angle features.The fused features are used to train the vehicle classifier,which improves the accuracy of vehicle classification at different angles.3.Classification of vehicle posture.In view of the current lack of vehicle posture annotation information and uneven distribution of vehicle posture in the current vehicle data set,we adopted the method of obtaining vehicle posture data from the perspective of the 3D software camera.And for the small sample pose training set currently collected,we propose an adaptive weighted HOG-PCA algorithm,which is combined with LBP features to jointly extract vehicle pose features,and finally divides the vehicle pose into eight different by SVM classification category.In order to meet the occupant’s adaptive protection under active and passive safety conditions,and provide data for the prediction of vehicle collision conditions.This article mainly proposes a method that can classify the vehicle type and posture of the vehicle by combining the three algorithms,and through the road image data set experimental test,this method can accurately identify and frame the target vehicle,and classify the vehicle type and vehicle posture,so as to provide useful information outside the vehicle for the safety detection part of the occupant in the vehicle,which has good innovation and application value.