Skip to content
Open access

DLP-Pillars: A LiDAR 3D Object Detection Method With Adaptive Re-Encoding and Multidimensional Feature Enhancement

2026 · IEEE Access · Vol 14, pp. 171856-171874 · 0 citations · 36 references

Abstract

3D object detection is a core task in intelligent perception, and LiDAR point clouds provide precise spatial geometric and scale information. To address issues such as high sparsity, non-uniform distribution, and insufficient feature representation of distant objects in LiDAR point clouds, this paper proposes DLP-Pillars, a lightweight 3D object detection method based on PointPillars. The method first introduces a Dynamic Sub-grid Re-encoding (DSRE) mechanism, which performs fine-grained spatial re-encoding of sparse regions by jointly modeling point cloud density and distance information, thereby mitigating the loss of geometric information for distant objects. Subsequently, a Large-Kernel Gated Depthwise Separable Convolution (LKGD) module is incorporated into the deeper layers of the backbone network to expand the receptive field and enhance contextual aggregation, while a learnable dynamic residual gating mechanism is employed to suppress background noise. Finally, we propose a Pillar-aware Spatial-Channel Attention (PiSCA) module to enhance sparse feature representations and further suppress background interference. Experimental results show that the proposed method outperforms the baseline PointPillars on the KITTI dataset, achieving 3D detection AP scores of 76.75%, 48.05%, and 63.99% for the Car, Pedestrian, and Cyclist classes, respectively, under Moderate difficulty, while maintaining high inference efficiency and a lightweight model design. Ablation studies further validate the effectiveness of each module and their synergistic benefits.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.