Research On Road Vehicle Target Detection Method Based On Deep Learning

In recent years,artificial intelligence technology is developing at a high speed,penetrating all aspects of our lives,from driverless to online payment,and from video processing to voice processing.Especially now with the improvement of people’s living standards,people’s quality requirements,convenience requirements,comfort requirements and safety requirements are becoming higher and higher,which is particularly evident in driverless smart cars.At present,the biggest practical demand for driverless cars is to enable them to quickly,effectively and safely sense the vehicles on the road.Therefore,the research on vision-based road vehicle perception methods is of practical significance and role.In the research of road perception methods,the target detection algorithm occupies an extremely important position.However,in the traditional road vehicle detection methods,they are all 2D target detection.However,in actual environments,vehicles on the road are presented in 3D stereo Morphologically,using only 2D target detection cannot meet the rapid development of unmanned driving technology,and 2D target detection does not perform as well as 3D target detection.In the detection of 3D target vehicles,many leading researchers have combined deep learning networks and Lidar point cloud data,because deep learning networks have strong self-learning and excellent robustness,while Lidar point cloud data can Provides in-depth information about the target,which helps the regression of the target.However,due to the shortcomings of sparseness and irregularity of Lidar point cloud data,only using Lidar point cloud data in 3D target detection presents huge challenges in the positioning effect and robust performance of the target vehicle.Because of the above problems,based on fully studying the existing 3D object detection methods,this paper proposes a series of improved methods for their performance defects and verifies the feasibility of the improved methods through experiments.The main research contents and contributions of this article include:(1)First of all,this paper briefly introduces the target detection algorithm based on traditional manual feature extraction,focuses on the target detection algorithm based on deep learning,and introduces two representative 2D target detection algorithms,namely the one-stage YOLO algorithm and the two-stage Faster R-CNN algorithm.Finally,we focus on a 3D target detection method based on pure Lidar point cloud data.(2)Analyze the possibility of adding convolutional block attention model(referred to as CBAM)to the sparse embedded convolution detection algorithm(referred to as SECOND),and design an improved 3D target detection method based on CBAM model,CBAM attention The model is divided into Channel Attention(CA)and Spatial Attention(SA).The CBAM attention model connects CA and SA for processing.Comparative experiments such as single CA,single SA,parallel CA and SA are processed.(3)Finally,a new multi-scale Lidar point cloud voxel partitioning method is proposed in this paper.Lidar point cloud data is divided into voxels at multiple scales,and then voxel grids at different scales are extracted by feature extraction networks.Finally,the features at each scale are cascaded and sent to the subsequent network.This method solves the problem that target information cannot be fully utilized in a specific scene under a single-scale voxel partition.