Geometry and density-aware PointPillars for lightweight 3D object detection

Abstract

To address missed detections of distant sparse objects in LiDAR 3D detection and inference latency on edge devices, we propose GD-Pillars, a lightweight algorithm. First, a Geometry and Density-Aware Pillar Feature Encoding (GD-PFE) network enhances local geometric descriptions and compensates for sparse target features via a log-scaled residual attention mechanism. A feature alignment layer is integrated to effectively mitigate covariate shifts caused by this non-linear scaling. Second, an efficient structural reparameterization backbone (Rep-Backbone) leverages multi-branch training and folds into a single-path convolution during inference, boosting spatial representation with zero extra latency. On the KITTI dataset, GD-Pillars maintains 41.9 FPS while improving moderate-difficulty pedestrian 3D AP by 2.03% and cyclist Average Orientation Similarity (AOS) by 2.52%, substantially enhancing comprehensive 3D perception without increasing edge computing burdens.