Reliable 3D scene understanding is a fundamental requirement for intelligent machines in autonomous transportation systems, as on-board perception must remain accurate and stable across diverse environments and sensing conditions. However, LiDAR point clouds acquired in real traffic scenes are often sparse and irregular, and they exhibit heterogeneous sampling patterns that hinder consistent and fine-grained semantic interpretation. To address these challenges, this paper proposes LVCA-Net, a lightweight voxel–coordinate attention framework designed for efficient LiDAR-based 3D semantic segmentation in autonomous driving scenarios. The architecture integrates (i) an anisotropic depthwise residual module for direction-aware geometric feature extraction, (ii) a hierarchical LiteDown–LiteUp pathway for multi-scale feature fusion, and (iii) a Coordinate-Guided Sparse Semantic Module that enhances spatial consistency in a cylindrical voxel space while maintaining computational sparsity. Experiments on the SemanticKITTI and nuScenes benchmarks demonstrate that LVCA-Net achieves 67.17% mean Intersection over Union (mIoU) and 91.79% overall accuracy on SemanticKITTI, as well as 77.1% mIoU on nuScenes, while maintaining real-time inference efficiency. These results indicate that LVCA-Net delivers scalable and robust 3D scene understanding with high semantic precision for LiDAR-only perception, making it well suited for deployment in autonomous vehicles and other safety-critical intelligent systems.
RGB–thermal detection benefits from both cross-modal feature calibration and the recovery of fine-scale structural evidence. However, spatial responses produced during these operations are often used only to regulate intermediate features. We propose a stage-specific semantic–structural descriptor-conditioning framework that retains calibration and discrepancy maps from and structural-weighting and propagation maps from . At –, the four maps are jointly encoded with pyramid features to generate separate, bounded adjustments for classification and regression. On FLIR Aligned, the framework achieves mAP50:95 across six seeds, compared with for the matched baseline (paired gain points), using 7.51M parameters and 25.96 GFLOPs. It also outperforms a compute-matched ordinary control by points. Structural recovery provides the largest standalone gain. Under matched feature-path and prediction-head architectures, a six-seed four-way descriptor-group control shows a modest positive semantic–structural interaction ( points). Live descriptors exceed dual-neutral and cross-sample-permuted controls by and points. Independent retraining on M3FD yields a -point mAP50:95 gain. Stress tests further characterize sensitivity to spatially varying thermal misalignment.