Design And Implementation Of Road Target Detection System Based On Deep Learning

Road target detection model is an important core component of automatic driving technology.This paper mainly studies four kinds of targets:pedestrian,small vehicle,large vehicle and traffic signal lamp.Because of its strong feature extraction and generalization ability,deep learning can effectively solve the problem of low average precision of road target detection model.However,due to the large capacity of the model,it is generally real-time performance when deployed to the equipment with limited hardware performance.Therefore,it is necessary to use pruning optimization technology to lightweight the model,Only then can it be applied to the automatic driving technology.Based on the road target detection based on deep learning,the model is lightweight in the range of acceptable range to ensure the average accuracy drop.The research work of this paper is mainly divided into the following aspects.In this paper,we use pruning optimization technology to lightweight the road target detection model based on YOLOv4,specifically,L1 regularization is used to deal with BN layer y Sparse parameters,sparse y The channel selection ability of parameters prunes the unimportant channels in the model,so as to achieve the purpose of lightweight model.Because of the instability of the pre training,sparse training and fine adjustment process,this paper adjusts the original constant learning rate training mode to the training mode that reduces the learning rate by 10 times every 5 epochs.In this way,the pre training,sparse training and fine adjustment process can operate normally,and the calculation and parameter quantity of the model are reduced after pruning,The model can reduce the consumption of hardware resources,and ensure real-time performance,and ensure that the average accuracy does not decrease or the range of decline is within acceptable range.In this paper,CSPDarknet53,the backbone network in YOLOv4 road target detection model,is modified by using the deep separable convolution and inverse residual network in mobilenetv2,and the conventional convolution operation is transformed into deep separable convolution depthwise resolution and pointwise convolution,At the same time,the conventional residual structure in CSPDarknet53 is transformed into the reverse residual structure.In order to make the model have better performance,this paper also adds passthrough layer in YOLOv2 algorithm to improve the detection ability of the algorithm for small targets,At the same time,some conventional convolutions from the three branches of the backbone network CSPDarknet53 are transformed into deep separable convolutions depthwise resolution and pointwise convolution.Finally,three convolution layers are added to the yolo head,which is used to detect large targets.These three convolution layers are not modified by depth separable convolution.After adding three convolution layers,it is found that the added convolution layer has a certain contribution to the detection ability of large targets.The experiment shows that the model can improve the reasoning speed of the model while ensuring the average accuracy.This paper presents the web deployment of road target detection model and makes an experimental analysis.The analysis results show that the proposed optimization method can improve the detection ability of road targets.