PES-PointPillars: LiDAR-Based 3D Object Detection for Autonomous Driving with Directional Convolution, Adaptive Feature Fusion, and Decoupled Regression

Abstract

LiDAR-based 3D object detection for autonomous driving must balance localization accuracy with real-time inference, while sparse point measurements make small-scale objects such as pedestrians and cyclists particularly challenging to represent at long range. This paper presents PES-PointPillars, an enhanced PointPillars detector with three coordinated design changes. First, pinwheel-shaped convolution (PConv) replaces selected backbone convolutions to expand horizontal and vertical receptive fields for sparse structural patterns. Second, an Improved Inter-Layer Feature Correlation (I-EFC) module uses soft gating and adaptive thresholding to fuse multi-level features through continuous, input-dependent weights. Third, a Smooth L1-NWD (SNWD) loss applies normalized Wasserstein distance to planar position and scale while retaining Smooth L1 regression for vertical position, height, and orientation. Using the parameter settings and configuration of the original PointPillars implementation, the locally executed PES-PointPillars experiment achieves Moderate 3D average precision values of 77.1% for cars, 46.7% for pedestrians, and 62.9% for cyclists at 68.3 FPS on the KITTI validation split. Relative to the source-reported PointPillars reference, the corresponding numerical differences are 2.1, 3.2, and 3.8 percentage points. The reported component-wise and staged ablations show category-dependent gains, with the complete model providing the strongest aggregate performance among the evaluated configurations.