Multi-Feature Aggregation for Semantic Segmentation of an Urban Scene Point Cloud

With the rapid development of cities, semantic segmentation urban scenes, as an important and effective imaging method, can accurately obtain distribution information typical ground features, reflecting scale level greenery in cities. There are some challenging problems point clouds including different scales, imbalanced class distribution, missing data caused by occlusion. Based on cloud network RandLA-Net, we propose networks RandLA-Net++ RandLA-Net3+. The is a deep fusion shallow features clouds, series nested dense skip connections used between encoder decoder. RandLA-Net3+ based multi-scale connection decoder; it also connects internally within decoder to capture fine-grained details coarse-grained at full scale. We incorporating dilated convolution increase receptive field compare improvement effect loss functions sample imbalance. After verification analysis our labeled scene LiDAR dataset—called NJSeg-3D—the mIoU 3.4% 3.2% higher, respectively, than benchmark RandLA-Net.