LiDAR Pointcloud Semantic Segmentation Research

In recent years,developments in the field of autonomous driving technology and artificial intelligence have led to an increased interest in autonomous driving systems to sense object information in real time.3D LIDAR has become one of the most widely used sensors in autonomous driving due to its high accuracy,high frequency,and integrity of view.After preliminary research,this paper found that the current graphics processing units(GPUs)in embedded devices are lacking in terms of computing power,and data communication technology cannot support cloud server to remotely process a large amounts of LIDAR point cloud data.However,the spherical projection algorithm,can effectively increase the processing speed of the neural network on the SemanticKITTI point cloud data to more than 10 scans/s.Furthermore,point cloud semantic segmentation systems that have adopted the spherical projection preprocessing scheme,such as SqueezeSeg and RangeNet++,present the following three weaknesses:(a)the system directly transplants the 2D convolutional neural network for real images for feature extraction of point cloud projection images,thus it lacks a targeted design based on the characteristics of point cloud projection images,and there is still space to improve the performance of the semantic segmentation neural network;(b)the system lacks an improved design for processing sparse mechanical LIDAR point clouds and solid-state LIDAR point clouds,resulting in insufficient generalization capability of the network for processing various types of LIDAR point clouds;(c)the problem of insufficient ability of the system to extract small-scale target features is not addressed.To address the above problems,this paper in-depth investigates key technologies such as image channel information processing,small-scale target recognition,and sparse feature extraction.Subsequently,this paper constructs three frameworks of point cloud semantic segmentation for three types of LiDAR,aiming to improve the processing accuracy and precision of point cloud semantic segmentation systems for different types of data.The main work accomplished and the innovations achieved in this paper are listed as follows.1.To address the problem of different types of channel tensor information between cloud projection images and real images,this paper designed a Pointwise Convolutional Expansion Module(PCEM)for feature transformation.The application of this sub-module to the base modules of the semantic segmentation network reduces the number of network parameters and makes the system more suitable for processing projected image data.In this paper,experiments were conducted using projection images generated by conversion of the SemanticKITTI opensource point cloud dataset.The results demonstrated that the system achieves a processing speed of 19.4 frames/second on the RTX3090 platform,with a system accuracy,or Mean Intersection over Union(mIoU),of 49.3%.Compared to RangeNet++,the processing speed increased by 58.3%while the accuracy only decreased by 2.9%.2.For the problem of low accuracy of semantic segmentation of small-scale objects,this paper designed a Multi-scale Spatial Attention Mechanism Module(MSAM).It enables to adjust the local feature weight region of the multi-scale target after projection of the fitted original data.The results of comparison experiments with other algorithms on SemanticKITTI indicated that the system achieves the processing speed of 17.3 frames/second,with an mIoU of 51.8%.Compared to SqueezeSegV3,it represents a 183.3%increase in speed,and only the 4.1%decrease in accuracy.In addition,MASM has improved the segmentation accuracy of small-scale targets by 16.9%compared to the original system.3.To address the lack of feature extraction capability for sparse point clouds and small-scale targets,this paper proposed a Coordinate Fusion Module(CFM).This module fuses intermediate features with the original coordinate information to enhance local features,removing the feature transformation convolution layer from the base module to reduce the number of coordinate information fusion.In addition,to verify the performance of the algorithm,a sparse open-source dataset and a selfpicked sparse point cloud dataset were constructed.The results revealed that the mloU of the system reaches 44.1%,which is 11.5%and 2.9%better than RangeNet++ and SqueezeSegV3 respectively.4.According to the characteristic that solid-state lidar generates the same number of point cloud data per frame in a scene,an adaptive resolution spherical projection scheme is proposed in this paper.It enables the extension and compression of projected images to preserve the complete information of the point cloud.Comparative experiments on the PandaSet solid-state LIDAR point cloud open-source dataset confirmed that the projection improvement scheme improves the performance of the semantic segmentation system for solid-state LIDAR point clouds with an mIoU of 16.9%.There is an 11.8%and 5.6%improvement compared to RangeNet++and SqueezeSegV3,and a 12%improvement compared to the fixed resolution projection processing scheme.