Traffic Sign Detection Algorithm Based On Lightweight Network And System Implementation

Automatic driving is a research field with important application value,and traffic sign detection and recognition is one of the important research directions applied in the field of automatic driving.At present,with the continuous development and technical improvement of deep learning,traffic sign detection and recognition algorithm based on deep learning has become the preferred algorithm in the field of traffic sign detection and recognition because it is superior to traditional methods.Although the performance has reached a high level,due to the large amount of model parameters and high requirements for calculation and storage costs,it is difficult to deploy real-time detection at the mobile terminal.Therefore,it is necessary to reduce the amount of parameters and calculation of the model as much as possible while completing the task of traffic sign detection.In this thesis,the design of lightweight traffic sign detection algorithm is studied.The main contributions are as follows:(1)a lightweight traffic sign detection network is proposed by making four improvements to YOLOv3-tiny.Firstly,the lightweight channel attention module is introduced to strengthen the ability of network backbone to extract important information;Secondly,by fusing the deep and shallow features,a new detection scale is added to improve the detection ability of the network;Then,the network feature extraction layer is improved to deepen the network depth and improve the network feature extraction ability;Finally,the loss function is improved to optimize the calculation of positioning loss for CIo U and reduce the error.(2)In view of the problems existing in the traffic sign data set,such as too small proportion of target instances,unbalanced categories and large difference in label size,a variety of data enhancement methods are adopted,including the combination of image clipping,Mosaic and Multi-scale training,so as to make full use of and deeply mine the information of the training set and effectively improve the network performance.(3)In order to further reduce the amount of parameters and calculation of the improved model,the channel pruning algorithm based on BN layer is used to prune the improved model.Aiming at the low efficiency of manually adjusting the pruning ratio,a pruning ratio search method based on binary search strategy is proposed.(4)Knowledge distillation is used to restore the accuracy of the model with large performance loss after pruning,a knowledge distillation method based on multi branch block is proposed,which effectively improves the performance of the model with large precision loss after pruning.(5)Deploy the final model to the Jetson Xavier NX embedded development board,collect and establish the actual scene traffic sign image data set and test it.At the same time,build a traffic video real-time detection system on the target board to verify its feasibility.In this thesis,a series of comparative experiments are carried out in terms of network improvement,model compression and hardware deployment.The final algorithm model parameters are only 3.16 M and the amount of calculation is 75.89 B.The mean average p recision on Tsinghua Tencent traffic sign data set is 88.94%.The actual scene is tested,and the real-time detection is basically realized on the target board.The experimental results show that the model proposed in this paper can better complete the task of real-time traffic sign detection and recognition on the premise of small parameters and fast reasoning speed.