Research On Objection Detection Of Road Scene Based On Deep Learning

Accompanied by the breakthrough of artificial intelligence technology with deep learning as the core,it is gradually used in the field of computer target detection.Aiming at the road scene target detection task,this paper focuses on directions of 2D target detection and 3D target detection,taking the YOLOV3 algorithm and Point RCNN algorithm as the research object,and puts forward the following improvements for its shortcomings.(1)Firstly,the application scenarios of computer vision,especially target detection technology,are introduced,and the development status of target detection technology at home and abroad is described briefly;Then it introduces the deep learning theory,including convolution layer,pooling layer,full connection layer,activation function,etc.;Then it discusses the principle of mainstream feature extraction network model and several important target detection network models,such as Faster RCNN、SSD、YOLOV2;Finally,it compares several common deep learning frameworks.(2)Aiming at the problems of missed detection,false detection and low confidence of the original YOLOV3 algorithm in road scene target detection,his paper proposes an improved YOLOV3 algorithm,which combines theoretical algorithms such as grouped convolution and Kmeans clustering to improve the feature extraction module and anchor frame parameters.First,an Im Res Next module is proposed to replace the Res Net module in the original YOLOV3 network to improve the feature extraction ability of the feature extraction network;then it is aimed at the YOLOV3 loss function;then the bounding box coordinate loss module is improved;and then a DIo U is proposed for the NMS module-Softer-NMS module;finally use the Kmeans algorithm to cluster the sample set to generate anchor frame parameters.We use the KITTI data set for training and verification experiments to show that the improved algorithm has higher detection efficiency and better accuracy.At the same time,it also achieves good detection results in real scenes,thus proving that the improved algorithm IYOLOV3 proposed in this chapter is more effective than the original algorithm.The accuracy and generalization performance have been improved,and the AP values of "Car","Cyclist",and "Pedstrain" are increased by 1.12%,1.23%,and 0.65% respectively.(3)The Point RCNN algorithm for 3D target detection based on lidar data is easily affected by the sparsity of point cloud data.Therefore,this paper proposes the following improvements to the algorithm: Firstly,Sk Net structure based on attention mechanism is integrated into the feature extraction module,then improve the loss function of angle,and finally KITTI data set is used for training and verification.The experimental results prove that the improved algorithm IPoint RCNN proposed in this chapter has improved accuracy and generalization performance compared with the original algorithm.The AP values of "Car","Cyclist",and "Pedstrain" are increased by 1.29%,0.94%,and 0.56%respectively.