Research On Road Scene Vehicle Object Detection Method Based On Deep Learning

In recent years,autonomous driving has received widespread attention due to it’s great application prospects,and environmental perception,as the biggest obstacle of the practical application of autonomous driving,it has been widely studied to obtain environmental information during vehicle driving in an efficient and low-cost manner.Although the existing 2D-based detection technology can achieve better environment perception,the detection technology has large limitations and limited information acquisition,which cannot meet the amount of information required for actual vehicle driving.With the rise of deep learning-based artificial intelligence technology,many researchers and scholars have proposed 3D target detection technology based on deep learning networks,combined with radar point cloud data and using the network’s autonomous learning characteristics.However,due to the sparse and irregular characteristics of radar point cloud data,the detection effect of 3D target detection technology which use radar point cloud data only still has many problems.In order to improve the vehicle target detection effect,this paper improves and optimizes the target detection algorithm based on the existing 3D target detection methods,and proposes a new3 D target detection method.The feasibility and effectiveness of the improved and optimized algorithm are verified by building a Li DAR platform.The main work and contributions of this paper are summarized as follows.(1)To sort out and analyze the current domestic and foreign research status of environment-aware vehicles based on deep learning target detection algorithms in the field of autonomous driving.The 2D target detection algorithms in the first stage(SSD algorithm)and second stage(Faster R-CNN)and the 3D target detection algorithm based on structured processing of radar point cloud data are summarized.(2)In response to the problems caused by the sparse radar point cloud data,long distance vehicle targets and obscured vehicle targets are mis-detected and missed,this paper introduces an attention mechanism based on the point cloud columnar fast coding target detection algorithm,and proposes a 3D target detection method that incorporates the improved attention mechanism to extract deeper features and obtain more comprehensive environmental information by combining Channel Attention(CA)and Spatial Attention(SA).In order to achieve the best detection effect,this paper also conducts comparison experiments on different order combinations of two different sub-modules,such as single CA,single SA,parallel CA and SA and series different order CA and SA.for three recognition levels of easy,medium and difficult,the research results show that the proposed optimization algorithm is more accurate than the existing basic algorithm.;and the visualization results show that the optimized algorithm has stronger robustness in detecting distant targets and obscured scenes;the accuracy of the optimized sub-module combination for CA and SA sequentially.(3)In response to the problems of high cost of existing LIDAR and to reduce the detection cost,this paper builds a low-cost LIDAR platform based on the FM continuous wave LIDAR ranging principle.Collect laser point cloud data based on the built lidar platform to verify the feasibility and effectiveness of the improved algorithm model.It is found that the developed detection platform can realize 3D target detection,the platform that based on the improved detection algorithm can effectively detect the vehicle targets on the road.