LSR-RIBNet: A Novel LiDAR Super-Resolution Model for Scene Semantic Segmentation in Outdoor Environments

LiDAR-based scene semantic segmentation is crucial for unmanned ground vehicles working in outdoor environments. However, the performance of scene semantic segmentation is often poor due to the low vertical resolution of commonly used LiDAR sensors. To achieve high-resolution LiDAR data for semantic segmentation tasks, this paper proposes a novel LiDAR super-resolution network named LSR-RIBNet. To make full use of the range, intensity, and azimuth information of raw LiDAR measurements, LSR-RIBNet takes the 2D projection images of LiDAR points as inputs and predicts a high-resolution 3D LiDAR point cloud. Comparative experiments on public datasets not only show the effectiveness of the proposed LSR-RIBNet model, but also verify the performance improvement of scene semantic segmentation tasks with the LiDAR super-resolution results.