A 3D Object Detection Framework for Autonomous Vehicles Exploiting Dual-Relation Pillar Encoding

This study addresses critical limitations in pillar-based 3D object detection systems through three key innovations. First, our Dual-Stage Pillar Feature Encoding (DSPFE) module establishes inter-pillar geometric relationships using centroid displacement vectors, resolving under-segmentation issues in occluded scenarios. Second, we implement adaptive channel recalibration through squeeze-excitation blocks, achieving 18.7^% computational complexity reduction compared to conventional attention mechanisms. Third, the proposed orientation-aware loss function incorporates sinusoidal encoding for yaw angle prediction, demonstrating 4.2% improvement in pedestrian detection accuracy. Experimental validation on KITTI dataset reveals AP gains of 1.10% (cars), 3.78% (pedestrians), and 2.23% (cyclists) over PointPillars baseline under moderate difficulty settings. The framework achieves real-time performance of 42 FPS on NVIDIA Jetson AGX Xavier, confirming practical deploy ability.