Research On Lidar Moving Target Detection Technology

With the increasing demand for environmental perception and target recognition in the security field,the use of three-dimensional point cloud data for target detection and tracking is gradually gaining its advantages.Compared with traditional cameras,three-dimensional point clouds provide accurate information about the three-dimensional coordinates of each point,which can accurately describe the shape and position of objects,and help to estimate parameters accurately in environmental perception and target recognition tasks.Therefore,the technology of dynamic target detection and tracking based on three-dimensional point clouds has significant importance in the security field.However,three-dimensional point cloud data contains a large amount of background and noise interference,and the characteristics of targets in three-dimensional point clouds are vastly different from those of traditional visual features,which pose significant challenges to three-dimensional point cloud dynamic target detection and tracking.This paper proposes a series of processing methods for the key issues in three-dimensional point cloud processing,including dynamic point cloud registration and denoising,3D object detection and tracking,etc.,and the main research contents are as follows:Firstly,the basic theory and technology of 3D point cloud processing are introduced.Two representative deep learning networks,Point Net and Point Net++,are first introduced,which are basic network models for tasks such as classification and segmentation of point cloud data.Then,the basic methods of 3D point cloud detection and tracking are analyzed,including the Point RCNN-based 3D object detection algorithm and the Center Point-based3 D object detection and tracking algorithm.These algorithms have high accuracy and practicality in the field of 3D point cloud object detection and tracking.By analyzing and sorting out the above content,this article provides a theoretical basis and technical background for subsequent research.Second,in view of the problem of a large number of backgrounds in 3D point cloud data,the research on the static background elimination method based on the improved NDT algorithm based on density weighting and the fusion of multi-frame background point cloud data was carried out.Under the premise of retaining the target features,the amount of point cloud data is greatly reduced,and the registration and background removal of dynamic point clouds are effectively realized.Aiming at the problem of point cloud noise in dynamic environment,a noise model is established,and a dynamic point cloud denoising algorithm based on gradient field is proposed.Quantitative experiments,qualitative experiments and ablation experiments show that the point cloud noise removal effect is 0.66% higher than that of the advanced Socre Net method,which verifies the effectiveness of the proposed algorithm.Third,aiming at the problem of low object detection accuracy in 3D dynamic point cloud,a 3D point cloud moving target detection and tracking method based on multi-scale Center Proposal Network(CPN)is proposed.Robust object detection is achieved through multi-scale feature extraction,center proposal generation,and multi-frame feature fusion.In addition,an improved Transformer method based on the CPN network is introduced.The improved Transformer decoder includes a multi-head self-attention mechanism and a multihead cross-attention mechanism.A dynamic attention weight mechanism is also introduced to make full use of the original point cloud features and context information.,so as to improve the accuracy of target detection.Quantitative experiments and qualitative results analysis show that,compared with the existing methods,the method proposed in this paper improves the detection accuracy of the advanced Center Point method in object detection and tracking tasks,and provides an effective solution for the security field.