TITLE:
A Novel Pillar Feature Encoder for Pillar-Based 3D Object Detection in Point Clouds
AUTHORS:
Huijun Yu, Shuai Kang, Zhihao Zou
KEYWORDS:
Point Cloud, 3D Object Detection, Pillar Feature Encoder, Multi-Attention
JOURNAL NAME:
Open Journal of Applied Sciences,
Vol.15 No.3,
March
18,
2025
ABSTRACT: Point cloud-based 3D object detection is a key technology in autonomous driving and mobile robot perception systems. However, the sparsity and irregularity of point cloud data result in poor performance of existing methods in detecting small objects at long distances and occluded objects. This paper proposes a Novel Pillar Feature Encoder to address feature encoding challenges in Pillar-based 3D point cloud object detection, improving the detection of occluded and small objects, especially at long distances. This method converts point cloud data into pillar features through voxelization and designs two convolutional neural network branches: Point Feature Encoding and Pillar Feature Encoding. The former extracts point features within local pillars, while the latter extracts global pillar features, which are then fused to resolve the problem of occlusion-related information loss, thus enhancing the detection accuracy of occluded objects. A Multi-attention mechanism is introduced to enhance the focus on key point features and learn optimal channel weights, thus improving the detection of small objects at long distances. We conducted experiments on the PointPillars network framework using the KITTI dataset for training and testing. The results show that the improved algorithm significantly enhances the average precision (AP) for 3D detection of Cars, Pedestrians, and Cyclists on the KITTI dataset, demonstrating exceptional performance in detecting occluded objects and small targets at long distances, thus validating the effectiveness of the proposed method.