LPSF-LiDARNet: Log-Polar Spatiotemporal Fusion-Based LiDAR Point Cloud Semantic Segmentation for Autonomous Driving

This paper tackles the challenges of sparse and unevenly distributed of ring-like mechanical LiDAR point clouds in road environment perception for autonomous driving by proposing the LPSF-LiDARNet framework. The framework enhances fine-grained 3D semantic segmentation through inter-frame spatiotemporal feature enhancement and balanced voxel sampling. Temporal window fusion is achieved via multi-frame stacking, integrating complementary temporal features to mitigate single-frame incompleteness. Spatially, log-polar coordinate voxel sampling leverages spatial distribution patterns to improve feature consistency. Additionally, adaptive heterogeneous convolution kernels with dynamic attention mechanisms are introduced, combined with semantic-level data augmentation, to optimize feature extraction for sparse points and rare samples. The framework demonstrates superior performance over existing models in experiments using SemanticKITTI and local datasets, validating the effectiveness of its spatiotemporal fusion strategy and sampling mechanism. Ultimately, the model achieves a segmentation accuracy of 96.4% and a frame rate of 49.2 FPS on local datasets, outperforming existing methods. By addressing critical issues in ring-like mechanical LiDAR systems, this work provides a robust solution for intelligent agents to achieve real-time 3D environmental understanding in complex dynamic scenarios, offering significant practical value for autonomous navigation and decision-making systems.