| With the explosive growth of traffic information,it is difficult to manage it only by manual processing.The intelligent transportation system,which operates traffic data in a systematic and intelligent way,gradually occupies an significant position in modern traffic management.As an important component of the traffic,the use of various modern technologies to process vehicle data is a considerable part of the intelligent transportation to complete the intelligent management of the whole system.At this stage,the technology used for vehicle detection and recognition has problems such as large processing calculation amount and low efficiency.In view of the large amount of calculation in the vehicle detection and recognition technology mentioned above,this paper process an algorithm to detect vehicle targets in traffic scenes and recognize the color information and license plate authenticity using deep learning methods.The detection efficiency of vehicles has been greatly improved,and more abundant vehicle information is provided for subsequent vehicle management.The main work of this paper includes:1.Based on the YOLOv3 neural network model,realize the detection of vehicle targets in traffic scenes.The target detection model YOLOv3,which uses the deeper backbone network Darknet-53 and multi-scale features to complete vehicle target detection,by adopting appropriate pruning methods,can adapt to the needs of actual traffic scenes.While meeting the vehicle detection speed requirements,it can detect the vehicle’s front face image data containing enough vehicle information which can be directly used for vehicle information recognition without data waste.Through experimental comparison with Faster R-CNN target detection model,it is proved that the method used in the article can be superior in detection accuracy and real-time performance;2.Design an algorithm that uses the deep learning model Multi Color-Net to realize the color recognition of the car body.According to the vehicle front face detection results of the YOLOv3 network,on the basis of completing the pre-training of the two feature extraction sub-networks RGB Net and HSV Net,by extracting and splicing the RGB color features and HSV color features of the input image,the classification sub-network is used to realize the car body color recognition.Through the display of the test results on the test set,and the comparison experiments with classic network models such as Alex Net/Google Net,it can be proved that the color recognition model Multi Color-Net proposed in the article can maintain a lower computational load and achieve higher performance.The recognition accuracy rate is complete to complete the body color recognition of the vehicle in the natural scene;3.Realize the recognition of vehicle fake license plates through two neural network models.Use the YOLOv3 model to complete the vehicle front face area detection and vehicle brand recognition;use the deep learning model Multi Color-Net to complete the body color recognition;use the existing vehicle license plate recognition module to complete the license plate character recognition.And finally combining the existing vehicle license plate,body color and vehicle brand information in the vehicle attribute database,the authenticity of the license plate is determined by means of mutual verification of multiple attribute information.It has been verified that in a complex actual traffic situation,the authenticity of the license plate can be distinguished in real time while accurately distinguishing the authenticity of the license plate.This solution can learn features directly from the input image in an end-to-end manner to realize vehicle detection,body color recognition and fake license plate recognition in natural traffic scenes.Among them,vehicle front face detection is a preliminary auxiliary work of vehicle body color recognition,and data for vehicle body color recognition is obtained through vehicle front face detection;fake license plate recognition is a further extension of color recognition work.The experimental results calculated on the real data set in the article show that the proposed scheme can achieve higher accuracy and robustness compared with the existing detection and recognition methods,while maintaining a low computational complexity to meet the actual requirements. |
