Enhancing Point Cloud Feature Utilization for 3D Object Detection

doi:10.1109/LSP.2025.3626268

Enhancing Point Cloud Feature Utilization for 3D Object Detection

2025-01-01

0

PRE

AI

DOI：10.1109/LSP.2025.3626268

Original

Original request for help

Save

Abstract

En 中文

Traditional voxelization methods use many artificial components to handle sparse and uneven point clouds, which may lead to spatial discretization distortion. In addition, the fixed convolution size limits the detector's ability to capture feature correlations, resulting in low feature utilization. To adderss this challenge, this paper proposes a two-stage LiDAR 3D object detector, Pillar-CT3D++. First, it introduces a self-attention encoding module that uses a Transformer to process voxelized point cloud information. By capturing the position information of point cloud objects within voxels through multi-dimensional position encoding and an encoder-decoder mechanism, the quality of feature suggestions is improved. Second, a feature weight-aware module is proposed, which designs multi-scale pooling and group attention mechanisms in the 2D backbone network to adaptively recalibrate the feature responses of channels, generating more representative feature information. The detection performance of Pillar-CT3D++ is validated through experiments on the KITTI dataset and compared with existing detectors. Specifically, on the KITTI dataset, the proposed model outperforms the baseline CT3D by 0.43%, 1.04%, and 0.87% in detecting simple, medium, and difficult-level objects, respectively.

Keywords:

Feature extraction

Point cloud compression

Encoding

Three-dimensional displays

Object detection

Convolution

Vectors

Detectors

Decoding

Attention mechanisms

3D object detection

autonomous driving

LiDAR

point cloud

Enhancing Point Cloud Feature Utilization for 3D Object Detection

Abstract

Journal

Organization