Research On Target Detection And Recognition Based On 3D Point Cloud

Compared with two-dimensional space,three-dimensional spatial information contains almost all spatial information,which provides a variety of options for better insight and understanding of the real world.Therefore,more and more researches are extended to three-dimensional space.In computer vision,object detection and recognition is the current research upsurge,although two-dimensional object detection and recognition technology has made some achievements,but because of the lack of three-dimensional information,it is difficult to expand to three-dimensional space.Point cloud data is sparse and disordered,and is not robust to complex scenes and small targets.Therefore,it is of great significance to obtain effective 3D data quickly and efficiently,and establish a set of methods that can directly use 3D point cloud data for target detection.This paper attempts to use traditional clustering and deep learningbased methods to detect three-dimensional objects,and proposes improvements to address the shortcomings of existing algorithms,optimizes the algorithm model,and conducts simulation experiments in indoor and outdoor scenarios to verify the data availability and the feasibility of the algorithm.The main research work and innovation of this paper are as follows:1.Summarize and compare the traditional target detection algorithm,and use the traditional European clustering segmentation algorithm to quickly process the point cloud to get the three-dimensional detection box of the object.As the point cloud is close and far away,the traditional clustering is prone to over-segmentation and undersegmentation.In order to solve this problem,two kinds of conditions for judging clustering segmentation and segmental clustering methods are proposed to improve the traditional algorithm.2.In order to test the effectiveness of the improved European-style clustering algorithm,this paper uses the radar point cloud data of the public data set KITTI data set for experimental verification;The binocular camera was used to obtain the point cloud data in indoor environment,and the 16-line lidar was used to collect the point cloud data in outdoor environment.The algorithm before and after the improvement was verified respectively.The experimental results show that the detection accuracy of the improved European clustering algorithm has been greatly improved on the selfcollected data.3.In order to realize 3D point cloud target identification task,deep learning network Point Net ++ is used in this paper for experimental verification.In the indoor environment,an experimental platform was set up to collect point cloud data,and aiming at the problems of poor robustness and difficult fast matching of binocular images in the process of stereo matching,a method of integrating active vision for environment perception was proposed to obtain three-dimensional point cloud data and enhance the data.The experimental results show that the self-collected data set can complete the point cloud classification and recognition task in the deep learning network.4.The existing point cloud target detection network model Frustum-pointnets is improved.The definition threshold of the mask was widened,the attention mechanism was introduced into the three-dimensional space,and the global features were added.Finally,the mask calculation Loss function was changed into Focal Loss,and the ablation test was carried out in the large 3D scene data set KITTI data set.The detection accuracy of the improved vehicle was compared with other networks,and the detection results were finally visualized.