| In recent years,driving a car to travel because of its safety and convenient gradually become the most important way to travel,and intelligent transportation system has received more and more attention in China because of the increasing number of cars.The detection,recognition and tracking of vehicles play an important role in intelligent transportation system.However,the detection,recognition and tracking of vehicles still face many difficulties due to the complex and changeable driving environment and the changing shape of the vehicle itself as it moves.Therefore,this thesis discussed vehicle detection,recognition and tracking technology from the deep learning theory which is more popular in recent years,detecting the front vehicle and its corresponding license plate by the vehicle detection algorithm in urban lane video,then combined with the vehicle license plate character recognition algorithm to identify the target,finally using the vehicle tracking algorithm to achieve multi-scale fast tracking of vehicle target.The main work of this thesis is as follows:(1)Based on the characteristics of the fast detection speed of YOLOv4-tiny detection algorithm,the vehicle and license plate data set is constructed to train it,and it is applied to the vehicle and license plate detection.In order to improve the performance of YOLOv4-tiny in vehicle and license plate detection,after understanding the structure and principle of attention mechanism net—SENet,SENet is used in YOLOv4-tiny algorithm.The experimental results show that the Recall,mean Average Precision and mean Intersection over Union of YOLOv4-tiny embedded in SENet in the vehicle and license plate detection are 94.54%,97.55% and 84.49%,respectively,which are 4.68%,2.67% and 0.75% higher than YOLOv4-tiny.(2)A deep learning framework based on CRNN model is selected for accurate unsegmented recognition of license plate characters,In order to avoid the traditional license plate recognition algorithm complicated process and more error accumulation.Aiming to solve the problem that the feature extraction capability of the feature extraction network VGG16 model in CRNN model is not enough,for more accurate identification of license plate characters,the modified DenseNet model is selected to replace VGG16 model.Experimental data indicate that the CRNN model using DenseNet as the convolution layer has 95.20% and 99.23% accuracy in the complete license plate recognition and character recognition respectively,which is 3.6% and1.07% higher than the CRNN model using VGG16 as the convolution layer.(3)The Kernel Correlation Filtering(KCF)tracking algorithm is used for real-time tracking of moving vehicles in video sequences.Aiming to solve the problem that the tracking performance of the KCF tracking algorithm is not good when the target scale and attitude change,in the case of real-time algorithm,the KCF tracking algorithm which is combined with the improved YOLOv4-tiny algorithm is used for stable tracking of moving vehicles.Experimental results show that the KCF tracking algorithm with improved YOLOv4-tiny for multi-scale tracking of target vehicles has a higher average IoU value and a smaller target center position error in the tracking process.(4)PyCharm software and Python development language were used to complete the above experiments,and the advantages of each improved algorithm were verified by multiple indicators. |
