Lightweight Design And Deployment Of Object Detection Model For Traffic Scenes

In traffic scenes,the targets exhibit characteristics such as high complexity and high density,which demand object detection models used in traffic scenes to possess accurate detection capabilities and fast response abilities.Model lightweighting is a critical technology for deploying object detection models,achieved through the removal or deletion of redundant parameters in deep network models,compressing the network model to reduce its size without compromising its accuracy.With the development of object detection technology,autonomous driving becomes possible.However,how to apply theoretical object detection models to real traffic scenarios is still an urgent problem to be solved.Most deep learning-based object detection methods have high computational complexity and memory consumption,making it difficult to deploy on edge devices with limited computing resources and memory.Considering the requirements for model size and response speed in actual traffic scenarios,this study uses the one-stage target detection model Yolov5 s as the benchmark model,and performs a series of lightweight designs on it to deploy the lightweight model.A lightweight model is obtained through model compression technology.First of all,this work can significantly reduce the width and depth of the network by channel pruning the network and pruning the unimportant parts.Then use model quantization technology to further compress the volume of the model.Finally,knowledge distillation is used to retrain the pruned model to improve detection accuracy.The main research work includes:(1)For the problem that the target detection model deployed on the embedded computing platform occupies a large amount of memory and has a slow response speed,this study first selects Yolov5 s as the benchmark model,and performs channel pruning on the convolutional layer and batch normalization layer.The network weight matrix is clustered and quantified to generate a lightweight model Yolov5s-Light with fewer parameters and smaller size.(2)Aiming at the accuracy loss of the model after the lightweight model,this paper proposes a knowledge distillation framework called VT distillation,which uses the lightweight model after model pruning and quantization as the student model,and uses the teacher network Yolov5 l The knowledge learned from the self-built virtual data set Virtual Ped Cross-720 guides the training of the lightweight model in the real data set and corrects the performance of the lightweight model Yolov5s-Light.Then the final model Yolov5s-Compression is deployed to the computing platform for performance testing to prove the feasibility of the lightweight method in this paper.This paper generates a lightweight model Yolov5s-Light after performing pruning and quantization operations on the Yolov5 s model,and conducts a series of comparative experiments on model pruning and network quantization respectively to prove the performance of the pruning and quantization schemes;then use the proposed The VT framework,a knowledge distillation framework combining virtual and real,corrects the model performance of Yolov5 sLight,and builds a virtual dataset Virtual-Ped Cross-720 according to the data type and structure of the real dataset City Persons,and uses it in the VT framework Train the teacher network and the student network to get the final model Yolov5s-Compression,and conduct a series of experiments on the real dataset City Persons and the virtual dataset Virtual-Ped Cross-720 to illustrate the effect of the VT framework on improving the accuracy of small models;Finally,the corrected model is deployed to the development board for actual testing.The results show that the model lightweight solution in this paper is suitable for the target detection model and is suitable for deployment on the edge computing platform.