Research On Vehicle Attribute Detection And Recognition Based On Deep Learning

With the rapid development of the city,vehicle identification is a very key module in the intelligent traffic management system.Combined with the actual traffic management scene,this paper proposes to divide the vehicle attributes into appearance attributes and safety attributes.Safety attributes mainly refer to whether the driver abides by the traffic rules,fasten the seat belt and answer the phone;Vehicle appearance attribute recognition this paper mainly studies vehicle type and vehicle color recognition.From the perspective of application,this paper studies the vehicle appearance attribute recognition and safety attribute recognition for the actual traffic scene.After improvement based on yolov3,it mainly solves the detection and recognition problems of high similarity of some vehicle models,false detection and large size difference of target objects(1)In order to solve the problem that the fog and rain weather will block the target and blur the image in bad weather,a generative countermeasure network based on attention mechanism is used to enhance the image quality,so that the feature information of the target area extracted by attention network is shared to generative network and discriminant network during model training,Experiments are designed to prove that the algorithm can effectively alleviate the problem of low accuracy of target detection and recognition due to the loss of image information.(2)Aiming at the problems of high similarity of vehicle type and insufficient pixels of small target image,a cascade morphological multi feature extraction network based on residual network is proposed.By widening the network module,the target features are extracted with a wider feature plane during feature extraction operation,and the extraction ability of a single residual module on the amount of feature information is improved.The feature aggregation module of each dimension and the feature transfer module based on attention mechanism are cascaded to effectively divide the features of different network level,and achieve the effect of multi feature sharing through multi-level connection.The target area is significantly enhanced and transferred to the deep layer of the network for application.Finally,a experiment is designed to verify the improved algorithm have a effective improvement on the detection accuracy of similar targets and small targets.(3)Aiming at the problem of inaccurate location of prediction bounding box caused by polarization of large and small targets,the detection layer of yolov3 is optimized,and eiou is proposed to replace the original loss function to evaluate the position relationship between prediction bounding box and target.That is to say,by increasing the loss calculation of side length difference,the optimization of prediction bounding box width and height in model training is towards reducing and comparing with the real bounding box width and height High value of the direction.By setting up a control experiment to verify,after using focal eiou loss to correct the error of location information,the experimental results show that the optimization can effectively improve the accuracy of target positioning.