Research On Target Detection Method Based On Point Cloud And Deep Learning

Target detection is an important branch of image processing and machine vision.In the field of autonomous driving,multiple cameras and lidars are usually used as the eyes of the vehicle to achieve target perception tasks.Comparing the two-dimensional image data collected by the camera with the point cloud data collected by the lidar,it is found that the latter can more accurately reflect the three-dimensional spatial information,so the point cloud data has particularly important value and significance in the three-dimensional target detection task.Due to the huge amount of point cloud data and a large amount of redundant information,existing deep learning methods usually directly use point cloud data as input,which wastes computing resources to a certain extent,affects the overall efficiency of the deep learning model,and detects small target objects Poor performance.In response to these phenomena,this paper takes the three-dimensional target detection model PointRCNN network as the main research object,and designs two solutions to solve the problems of computational efficiency and poor detection performance of small targets.The first option is to combine the traditional point cloud processing method with the deep learning model,and the second option is to design a multi-branch feature extraction structure for the deep learning model.Both schemes have achieved certain experimental results,and have certain reference significance for target detection tasks in actual autonomous driving scenarios.The specific research content of this article is as follows:(1)The traditional point cloud processing technology and the point cloud processing technology based on deep learning are introduced,and the basic ideas,advantages and disadvantages of these two methods of processing point cloud data are mainly analyzed.In view of the limitations of traditional point cloud processing methods that are fast but not accurate,and deep learning methods have high accuracy in processing point cloud data,but the large amount of calculations,a plan to combine these two methods to perform three-dimensional target detection tasks is proposed.(2)Aiming at the characteristics of three-dimensional target detection tasks in autonomous driving scenarios,a grid-based ground segmentation algorithm is designed to filter out invalid ground information.The experimental results show the effectiveness of the algorithm.At the same time,the comparison and analysis of this algorithm with other ground segmentation algorithms prove the speed advantage of the algorithm.(3)The two experimental schemes of ground segmentation+deep learning model and ground segmentation+3D spatial clustering+deep learning model fully prove the feasibility and effectiveness of combining traditional point cloud processing methods with deep learning models.Under different difficult scenarios,These two schemes have obtained different experimental effects.(4)Aiming at the single phenomenon of PointRCNN model feature extraction,using the idea of feature fusion,adding voxel-based Voxel feature branch and voxel-based Pillars feature branch to the model,and splicing different features to achieve multi-branch features Fusion.The target detection performance of the final model,especially the detection performance of small targets,has a certain improvement.The innovative points of the research content in this article can be summarized as:propose the idea of combining traditional point cloud processing methods with deep learning models,use traditional methods to simplify the point cloud scene,and make the deep learning module focus on the target to be detected;design a grid-based ground The segmentation algorithm makes up for the large calculation and slow speed of RanSc and other algorithms;a multi-branch feature extraction structure is designed for the PointRCNN model,which enhances the feature expression ability of the model.