Study On Road Object Detection Based On Convolutional Neural Network

Nowadays,with the increase of car ownership,traffic safety issues are getting more and more attention.The rapid development of computer technology provides a solution to the driving safety problem of self-driving cars,among which the research of using computer vision technology to accurately perceive the road environment is of great significance and value.With the development of convolutional neural networks,the detection accuracy is improved while the network size and computation volume are increasing,which puts the hardware performance of the vehicle embedded devices to the test.In this topic,we use convolutional neural network to study the detection of objects such as pedestrians and cars on the road in front of vehicles,and improve the network to achieve light weight while ensuring the detection accuracy.The main work of the paper is as follows.（1）Aiming at lightweight networks needed for road detection,the YOLOv4 network with better detection performance is used as the basis,and the feature extraction network CSPDarknet53 of YOLOv4 is lightweighted using the Mobile Net and Efficient Net series of networks.Comparing the five lightweight networks,the improved Mobile Net V2-YOLOv4（MV2-Y）network improves the detection speed by 63.0 %,which improves the detection realtime performance,considering the evaluation indexes such as the number of parameters,computation volume,model size and detection speed of the network.（2）In order to improve the detection accuracy,the feature fusion network is improved by incorporating an attention mechanism,augmenting the network width,and K-Means++algorithm a priori frame clustering analysis on its feature pyramid network based on MV2-Y.Specifically,the channel attention mechanism is introduced into the feature fusion network and its position and number are experimentally explored to analyze and obtain the Mobile Net V2-SE-YOLOv4（MV2-S-Y）network.The MV2-S-Y network,YOLOv4-Tiny network,YOLOv4 network and MV2-Y network are validated on PASCAL VOC,Udacity and KAIST datasets,respectively.The results show that the model size,number of parameters,and computational effort of MV2-S-Y network are significantly reduced and the detection speed is improved by1.5 times compared with YOLOv4 network.The MV2-S-Y network improves the detection accuracy by 1.3 %,2.2 % and 0.5 % on the PASCAL VOC,Udacity and KAIST datasets,respectively,compared with the MV2-Y network with less than 1.0 % increase in model size;compared with the YOLOv4-Tiny network,the detection accuracy on the PASCAL VOC,Udacity and KAIST datasets by 4.0 %,13.5 %,and 3.5 %,respectively.（3）The loss function and activation function of the MV2-S-Y network are investigated,and the EIOU loss function and Mish activation function are used to optimize the network to obtain the MV2-S-YE network.And the low-speed multi-target scene detection is performed by self-built campus road object dataset.The experimental results show that the detection accuracy of the MV2-S-YE network is 1.6 % and 2.4 % better than that of the MV2-S-YE network on the PASCAL VOC and Udacity datasets,respectively;the detection accuracy of the MV2-S-YE network on the campus road target dataset reaches 88.1 % and the detection speed reaches 44 FPS.In this study,the network structure is based on YOLOv4 network according to the demand of network lightweight and real-time required for low-speed road object detection of selfdriving cars,and the improved network has better performance on public dataset and self-built campus road object dataset.