Guided Model-based LiDAR Super-Resolution for Resource-Efficient Automotive scene Segmentation

High-resolution LiDAR data is essential for 3D semantic segmentation in autonomous driving. However, the high cost of advanced LiDAR sensors limits their widespread adoption. Affordable alternatives like 16-channel LiDAR generate sparse point clouds, resulting in reduced segmentation performance. To address this, we present the first end-to-end framework that integrates LiDAR super-resolution (SR) and segmentation tasks. This framework achieves context awareness through a novel joint optimization during training, ensuring that the SR process integrates semantic guidance from the segmentation task and preserves the details of smaller classes. Furthermore, the proposed framework incorporates a novel SR loss function to enhance the network’s focus on regions of interest. The lightweight model-based SR network achieves a significant reduction in number of parameters compared to state-of-the-art LiDAR SR models, enabling efficient integration with segmentation networks. Experimental results demonstrate that our method delivers segmentation performance comparable to networks using high-resolution and high-cost 64-channel LiDAR data.