Research On Vehicle Detection And Vehicle Type Recognition In The Scene Of Traffic Checkpoint

Traffic checkpoints are specific monitoring locations set by traffic management departments at urban roads or highway entrances and exits.They can shoot and record motor vehicles passing through the checkpoints to achieve monitoring and management of complex road traffic conditions.In recent years,the state has attached great importance to the construction of intelligent transportation systems and vigorously promoted the informationization of transportation.As an important part of intelligent transportation systems,the traffic checkpoint system has important applications in vehicle information monitoring,suspected vehicle tracking,and false license plate identification.This thesis focuses on the vehicle face image obtained in the scene of traffic checkpoint,and conducts research work on vehicle target detection and fine-grained vehicle type recognition.The main research work and innovations are as follows:(1)The vehicle detection problem in the scene of traffic checkpoint is studied.This thesis designs a vehicle detection algorithm based on the YOLO v3 network in the target detection algorithm:1)In order to solve the problem of high missed detection rate and false detection rate of images taken at night,a dark light image processing algorithm is designed to improve the vehicle detection effect at night;2)According to the requirements of vehicle detection in the scene of traffic checkpoint,the anchor boxes are determined by cluster analysis for the similar size of vehicles;3)The method of random input of multi-resolution images improves the detection accuracy of multi-scale targets and enhances the robustness of the algorithm.The average detection accuracy of the constructed vehicle detection algorithm for vehicles is 91.69%.(2)The key point detection model of car face based on ensemble of regression trees is established.The model is based on a gradient boosting tree and a cascade structure,which can locate the key point coordinates of the outer contour of the car face.Through experimental verification,the average pixel offset between the key points detected by the model and the actual key points is about 9 pixels,which can accurately locate the key point positions of the car face image;On this basis,a car face alignment algorithm is designed.This algorithm can remove the unnecessary background information of the car face image.Subsequent experiments have verified the effect of car face alignment on improving the accuracy of vehicle type recognition.(3)A multi-branch convolutional neural network model is designed to achieve fine-grained vehicle type recognition.The car face image is divided into multiple regions according to components,and the corresponding branch convolutional neural network is used for feature extraction,and the high-level semantic features extracted by each branch convolutional neural network are used for feature fusion for vehicle type recognition.The model achieved a recognition accuracy of 91.81%on the self-built car face dataset(a total of 24155 images,including 146 different models).