Target Detection And Recognition Of Ground Natural Scene Based On Laser Point Cloud

With the rapid development of new technologies such as artificial intelligence,cloud computing and big data,more and more artificial intelligence and automation technologies are beginning to be applied to all aspects of production and life.In the 3D field,autonomous driving technology is in the ascendant.The key technologies are the detection and recognition of ground natural scene targets based on laser point cloud.These problems are also highly academic in research.One of the mainly difficulties lies in the complex environment during the detection process,the target detection is susceptible to occlusion,point cloud density changes,noise and other factors.After the detection and acquisition of the target point cloud,the difficulty lies in the ability of the existing descriptors is limited,so the recognition task cannot be completed well.This thesis starts from the two aspects of detection and recognition,and proposes the following work:In the detection stage,the multi-plane ground point fitting method is used to replace the original single-plane ground point fitting method.The selection of the seed points is also based on the height prior,and the speed is increased while the ground points are fitted more accurate;in the process of point cloud clustering,a point cloud clustering method based on LRF region growth is proposed.When the local reference frame(LRF)is constructed,the processing of noise and point cloud density is increased,which improves the robustness of the algorithm.In addition,the MCMD algorithm is used to improve the growth of seed points,and the higher quality domain points are obtained.The shape,distribution,reflection intensity and viewpoint features are used to construct the classifier to rough classify the cluster targets,which suppress false alarm.After completing the detection task and obtaining the target point cloud to be identified,this thesis adopts the most effective method currently,namely 3D convolutional neural network for recognition.In view of the low efficiency of the original 3D convolution,this thesis proposes a 3D convolutional neural network OctCNN which is more efficient in time and space utilization,based on the 3D convolution operation based on Octree structure.For the over-fitting problem,two improved networks Attribute-boosted OctCNN and SubVoxel OctCNN based on multi-task learning are proposed,which improve networks' generalization ability.In addition,combined with the advantages of the ESF descriptor,the traditional descriptor and neural network description are combined for the first time,and a new network architecture named as Fused OctCNN is proposed,which makes the recognition accuracy further improved.