Research On 3D Lidar Target Detection Technology Based On Deep Learning

As an important technology in the development of intelligent technology,unmanned control platforms have been widely used in the fields of unmanned driving and mobile robots.The perception of the external environment when performing work will be a key factor in the stability of their work.Most of the existing environment sensing technologies rely on camera images to perceive the environment,which can easily lead to perceptual bias in situations such as strong illumination and large area occlusion,and thus cannot accurately identify obstacle targets.In response to the above issues,this paper uses 3D laser radar as an environmental sensing device to sense the surrounding environment of unmanned platforms,conducts in-depth research on 3D point cloud target detection algorithms,and achieves breakthrough results in algorithm optimization and target detection system design.The main research work is as follows:1.A point cloud filtering algorithm based on density clustering is proposed.Inspired by DBSCAN clustering algorithm,with target clustering as the core,a method for selecting initial clustering points has been added,and the selection method for density clustering radius has been optimized.Experiments have shown that this filtering algorithm can effectively filter out outliers and invalid points in complex point clouds in large scenes.2.An adaptive slope point cloud ground segmentation algorithm is proposed.Using the slope threshold algorithm as the main idea,the distance between the unmanned platform and the obstacle in the current area is calculated using the angle of each laser beam of the threedimensional laser radar and the setting of the drivable slope to determine whether the area is the ground,thereby dividing the ground point cloud from the non ground point cloud.Experiments show that the ground segmentation algorithm can effectively segment ground point clouds.3.Optimized the Point Pillar target detection algorithm.Aiming at the low accuracy of the Point Pillar target detection algorithm,this paper introduces an attention mechanism based on it to enhance the networks feature learning for specific regions;At the same time,the residual network mechanism is applied to the feature extraction process,and a feature pyramid structure is constructed to enhance the feature extraction ability of the network model;Finally,in order to improve the efficiency of detecting targets,the Anchor Base type detection module is replaced with an Anchor Free type detection module to increase the detection speed.Experiments have shown that the optimized target detection algorithm significantly improves detection accuracy while ensuring real-time detection.This paper mainly studies the target detection technology of an unmanned control platform using three-dimensional laser radar as an environmental sensing device in outdoor environments.It detects different road conditions and obstacles.After three steps,namely,point cloud filtering,ground segmentation,and target detection,it achieves accurate recognition of obstacles.A set of target detection system is designed and deployed to unmanned tracked vehicles,Finally,the practicality of the system was verified through physical testing.The results of this paper will lay a technical foundation for the application of three-dimensional laser radar to obstacle detection on unmanned control platforms.