Research On Semantic Segmentation Of 3d Point Cloud For Large-scale Urban Scene

In recent years,with the development of intelligent robots,high-precision maps,smart cities,and other fields,the demand for large-scale urban scene understanding and intelligent perception of the environment has become increasingly high,and 3D point cloud semantic segmentation technology is the focus of research.However,there are still several problems to be solved in semantic segmentation of 3D point cloud for large-scale urban scenes.Firstly,the large scale and wide coverage of urban point cloud data pose a significant challenge to the processing speed and memory overhead of network training.Secondly,the amount of local point cloud in urban scene point cloud data is sparse,and architectural styles vary between different cities.It is impossible to accurately express urban objects only by relying on the geometric and color characteristics of point cloud.Finally,in urban point cloud data,the distribution of point cloud at the intersection of different objects is more dense than in other regions,making it easy to confuse each other during segmentation.At the same time,continuous downsampling operations during the training process will lose some important threedimensional points,making the network unable to accurately extract neighborhood features,further exacerbating the problem of blurred segmentation at the boundaries of adjacent objects.In response to the above issues,this article has mainly completed the following work:1.Aiming at the problem of large scale of urban point cloud data,which leads to difficulties in network training,a hybrid sampling method is proposed in the data preprocessing stage to perform downsampling processing on point cloud data.This method not only has the advantage of fast grid sampling speed,but also utilizes the characteristics of the farthest point sampling method that can maximize the retention of point cloud geometric characteristics,and the uniformity and stability of sampling points,ensuring the speed of subsequent network training,At the same time,it alleviates the problem of inaccurate shape feature extraction after sampling due to sparse local point cloud.2.Aiming at the problem that geometric shape features and color features cannot accurately express urban objects due to the sparse amount of local point cloud and different urban architectural styles,the normal vector features of each sampling point are introduced into the network,utilizing the feature that the normal vector features of point cloud have significant differences on urban objects with large differences in surface roughness and smoothness,effectively compensating for the shortcomings of geometric shape features and color features;A multi feature local coding module is proposed,which encodes and concatenates normal vector features,geometric features,and color features separately to avoid the problem of mutual interference between features during mixed coding.3.Aiming at the problem of blurred segmentation at the boundaries of adjacent objects,a multi feature bidirectional enhancement module is proposed,and the idea of neighborhood expansion sensing is introduced.The multi feature bidirectional enhancement module effectively solves the problem of blurred segmentation at the object boundary by learning the position offset to offset each neighborhood point according to its own feature relationship with the center point,and further improves the segmentation accuracy of the network by learning the color feature and normal vector feature offset in reverse direction;By introducing the idea of neighborhood expansion perception,a large range of 3D point features are aggregated onto a central point,effectively mitigating the impact of important 3D point loss on the network,and further solving the problem of blurred segmentation at the object boundary.Through corresponding experiments designed on the urban dataset Sensat Urban,it is verified that the proposed method for semantic segmentation of 3D point cloud in urban scenes is superior to other networks in all categories of evaluation indicators,and it is proved that the method in this paper can effectively improve the accuracy of semantic segmentation of 3D point cloud in urban scenes,with certain application value.