DS-RangeNet: Lightweight Dual-Stream LiDAR Semantic Segmentation for Industrial Indoor Environments
TL;DR
DS-RangeNet is introduced, a lightweight range image network that processes geometry and material-sensitive intensity in separate streams that handles shallow fusion and weak similarity after separate encoding and progressively stronger alignment during fusion.
Abstract
Real-time LiDAR semantic segmentation for industrial AGVs must distinguish repeated structures and weak glass returns while remaining robust to sensor-dependent intensity and tight edge computing budgets. We introduce DS-RangeNet, a lightweight range image network that processes geometry and material-sensitive intensity in separate streams. The geometry stream uses voxel-PCA descriptors, while the intensity stream uses normalized range, local intensity statistics, boundary strength, and intensity curvature. A lightweight convolutional attention block handles shallow fusion, whereas intensity–geometry cross-attention (IGCA) links deep features by estimating affinity within the guiding stream and routing values from the other stream. Centered kernel alignment (CKA) and normalized cross-covariance reveal weak similarity after separate encoding and progressively stronger alignment during fusion. On the site disjoint UBPC-9 test split, DS-RangeNet reaches 73.2% mIoU with 5.69 M parameters and 37 ms end-to-end latency on Jetson AGX Orin. A nine-fold leave-one-environment-out evaluation obtains 71.0% mIoU. The evaluation further spans SemanticPOSS and SemanticKITTI, five random seeds, 21 corruption conditions across seven families, standard cross-attention and convolution controls, a 60 min Jetson run, and cross-sensor transfer.
