Real-Time Semantic Segmentation Method Of LiDAR Point Cloud In Large-scale Outdoor Scene

As an essential branch of autonomous mobile robots’ environmental perception and understanding,3D Li DAR point cloud semantic segmentation must strike a balance between semantic segmentation accuracy and embedded platforms computing resources.Accurate segmentation results enable the robot to perform tasks like grasping,positioning,and path planning with more precision.Due to the difficulty of traditional heuristic geometric feature information extraction in meeting the increasingly complicated application environment and the diversity of categories,data-driven feature learning was developed.However,State-of-the-art methods convert Li DAR into regular 2D images or Cartesian grid for processing,which reduces the amount of computation caused by the unstructured point clouds.However,image-based methods inevitably cause inevitably alters the 3D geometric topology,while Cartesian grid-based methods ignore the density inconsistency of outdoor Li DAR,thus limiting their segmentation ability,especially for small instances such as pedestrians and bicycles.The real-time semantic segmentation of Li DAR point clouds in large-scale outdoor scenes is the emphasis of this research.The data source,representation method,and feature learning of the Li DAR point cloud all need to be taken into account when designing an effective semantic segmentation algorithm.Therefore,this thesis first designed an efficient Li DAR point cloud representation method.And further designed and implemented a real-time semantic segmentation method of Li DAR point clouds.Specifically,it includes the following aspects:(1)For the problem of representation method of Li DAR point clouds in outdoor scenes,A point cloud representation method based on 3D conical grid is proposed,which is divided into three dimensions: horizontal distance,inclination angle and azimuth relative to the Li DAR coordinate system,so as to balance the distribution of points in the unit grid.Experiments show that the proposed algorithm balances the inner points of the grid better than voxel and cylindrical grid on the outdoor Li DAR point cloud dataset,and has less quantization error.In the process of point cloud downsampling,the presented algorithm can better maintain the geometric information of the point cloud.(2)For the problem of semantic segmentation of high-resolution sparse grid,a new framework for the Li DAR segmentation base on 3D conical grid and sparse convolution network(Spconv3D)is proposed,where conical partition is used to solve the sparsity and density inconsistency of Li DAR.Reparameterization Spconv3 D to fully learn the geometric properties of grid,while improving memory usage and computational efficiency in the model inference stage.Experiments show that the proposed method outperforms the State-of-theart methods on the outdoor Li DAR point cloud dataset,especially for few categories and close-range points can be accurately identified,which is suitable for operation in complex scenes.(3)For the problem of high-resolution point cloud and efficient 3D grid fusion,a new framework for the Li DAR segmentation base on point-cone interactive fusion is proposed.To boost network inference speed,this approach uses point high-resolution information,sparse conical grids’ efficient local learning capabilities,and structural re-parameter technology.In the conical grid branch,designed an adaptive fusion network to obtain a greater feeling.An attention fusion network is designed in the feature interaction between points and grids to solve the problem that the trilinear interpolation method used in the process of data interaction is computationally expensive due to random memory access,and adjacent interpolation lacks the precise coordinate relationship between the two.In addition,an improved data enhancement method with mixed examples is also proposed to alleviate the imbalance of data samples during training and avoid the problem of overlapping augmented data.Experiments show that the proposed method has better segmentation accuracy and meets the realtime requirements of the algorithm.