Research On Object Classification And Recognition Method Based On GM-APD LiDAR

LiDAR is a sensor that actively probes using a laser beam as the carrier,which can obtain the three-dimensional point cloud image of the target and has all-weather,high-precision imaging detection capabilities.The LiDAR that uses the Geiger-Mode Avalanche Photo Diode(GM-APD)detector can perform high-sensitivity single-photon imaging on distant targets and has been widely used in the field of precision guidance.Target recognition is the core technology to improve precision guidance capabilities.However,when GM-APD LiDAR detects and images targets,it is easily affected by atmospheric interference,obstruction by obstacles,and the intermixing of multiple targets,leading to problems such as reduced signal-to-noise ratio in point cloud images and decreased accuracy in target classification recognition.This paper conducts research on point cloud denoising based on gravitational feature functions,ground point cloud segmentation based on plane fitting,object classification based on Hough transformation,object extraction based on morphological filtering,occluded object compensation based on Lucas-Kanade Gaussian pyramid,point cloud object recognition based on AGF-Transformer,and object recognition method based on intensity image-optimized dual deviation angle feature histogram.This enhances the denoising precision of noise suppression algorithms,reduces the influence of multiple mixed targets on classification algorithms,improves the resolution of point cloud images,strengthens the target recognition algorithm’s ability to represent target features,and achieves the goals of improving point cloud image quality and increasing target classification recognition accuracy.This has significant implications for comprehensively improving the precision strike capability of precision-guided weapons and the success rate of combat missions.The specific research contents are as follows:1)The GM-APD LiDAR point cloud denoising method is studied.Noise in LiDAR point clouds typically divides into two types: sparse outlier noise and dense outlier noise.Traditional denoising algorithms tend to consider dense outlier noise as real targets,leading to subpar noise removal.To address this issue,this study delves into the working mechanism of the GM-APD array LiDAR and the distribution characteristics of point clouds from multiple targets.We analyzed the causes of sparse and dense outlier noise in point clouds and established a spatial distribution model for noisy point clouds.Using this model,we investigated the gravitational feature representation method for point clouds and proposed a point cloud noise suppression method based on gravitational feature functions.By incorporating gravitational reference functions,we achieved rapid and precise removal of dense outlier noise.Our method was tested through simulations and experiments on public datasets as well as point clouds acquired outdoors using a GM-APD LiDAR.We achieved an average denoising accuracy of 0.942,an average denoising recall rate of 0.895,and an average processing time of 0.183 s.Compared to the EAR and WLOP methods,there’s at least a 12.5% improvement in average denoising accuracy,13.7% in average denoising recall rate,and 9.9% in average processing speed.The results indicate that the proposed point cloud noise suppression method based on gravitational feature functions effectively filters out dense outlier noise,providing high-quality point cloud images for subsequent target classification and identification."2)The multi-object classification extraction and obscured object compensation were studied.LiDAR-captured point clouds of target scenes contain a significant amount of ground point clouds,which greatly impede the speed of subsequent target extraction and classification.To address this,we propose a ground point cloud segmentation method based on plane fitting.By mapping point clouds to a sector grid space,we reduce the number of algorithmic iterations,achieving fast and precise ground point cloud segmentation.When multiple targets are intermixed,the boundaries between targets are indistinct,and different types of targets are easily perceived as a single entity.In response to this,we explored a target classification method based on the Hough Transform.Using morphological filtering,we extracted boundary information of adjacent targets to distinguish different types of targets,enabling accurate and rapid multi-target classification.When the extracted targets are obscured,it can lead to loss of feature information and a decline in recognition accuracy.To recover the lost feature information from obscured targets,we proposed an occlusion compensation method based on the Lucas-Kanade Gaussian Pyramid.By calculating the residual between the estimated and actual diffusion functions,we can correct the obscured parts of the target,enabling fast and accurate compensation.Our methods were tested on the KITTI dataset and on multi-target outdoor scenes with 40% occlusion,achieving a classification extraction localization Io U of 0.990,an average gradient of obscured target compensation of 6.536,and a computation time of 0.048 s.Compared to the RANSAC,GPF,and SPGraph methods,the classification extraction localization Io U,average gradient for occluded target compensation,and average computation time improved by at least 10.6%,92.3%,and 21.8% respectively.The results show that our proposed algorithm can rapidly and accurately classify and compensate for multiple obscured targets,providing data support for subsequent target identification methods."3)The point cloud object recognition method based on AGF-Transformer is studied.Traditional Graph Convolutional Neural Networks(GCNNs)utilize fixed-layer convolutional kernels to represent the local feature information of an object.When the target structure is complex,its representational dimension is limited,leading to missing feature information.To address this issue,an adaptive factor is introduced into the graph convolutional layer,allowing the nearest neighbor graph modeling dimension to be adaptively adjusted based on the complexity of the target structure.This achieves an adaptive dimensionality increase of the point cloud features,enriching the point cloud’s feature information.Local feature representation methods can easily fall into local optima due to the repeated encoding of local region features during the convolution process.By incorporating a Transformer layer,the point cloud’s local neighborhoods and global features are associated,which not only resolves the problem of local optima but also enhances the network’s overall representational capability of the target,enabling high-speed,high-precision recognition.Simulations and experimental tests on the Model Net40 dataset and outdoor point clouds captured using GM-APD LiDAR achieved class-average accuracy of 96.9% and overall average accuracy of 98.4%.Compared to methods like3 DShape Net Parts,DGCNN,and Adaptiveconv,there was at least a 9.6% and 7.8%increase in class-average and overall average accuracy,respectively.The results indicate that the proposed point cloud recognition method based on the adaptive graph convolution Transformer can achieve efficient and high-precision object recognition,meeting the needs of practical applications.4)The object recognition method based on intensity image-optimized dual deviation angle feature histogram is studied.In the process of object recognition,when prior knowledge is lacking,the accuracy of object recognition methods based on deep learning decreases.To address this issue,this study investigates a local reference coordinate system calibration method based on feature point sampling,using a dual bias angle feature histogram to represent point cloud features.To address the issue of incomplete feature representation of point cloud images,the GM-APD detector,capable of simultaneously capturing target intensity images and point clouds,is utilized.The Fourier contour information of the intensity image enriches the target feature representation capability.Subsequently,by integrating the ratio discrimination of the nearest and second-nearest neighbor distances,the target is recognized through feature matching.This study conducted simulations and experimental tests on the B3 R,Model Net40 datasets,and outdoor point clouds captured using GM-APD LiDAR,achieving an average recall rate of 0.953,average precision of 0.806,and average computation time of 0.528 s.Compared to the SHOT and SDASS methods,there was at least a 17.6%,19.2%,and 13.2% increase in average recall rate,precision,and computation time,respectively.The results demonstrate that the proposed algorithm can achieve high-speed,high-precision recognition of targets through feature matching techniques,even when prior knowledge of the target is insufficient,meeting the needs of practical applications.