Improving long-range 3D object detection with data augmentation and attention mechanism

Abstract

Autonomous vehicles rely on LiDAR-based 3D object detection, but long-range detection remains challenging due to the sparsity and degradation of LiDAR point clouds at extended distances. This project enhances detection through data augmentation and network improvements with attention mechanisms in the PointPillar framework. Firstly, we augment the training set with simulated long-range LiDAR data to increase sample diversity. Secondly, we incorporate channel, spatial, and cross-scale fusion to strengthen representation for distant and small objects. Finally, we remove heavily occluded and low-density samples for better data quality.

Beyond performance improvement, we conduct a comprehensive error analysis to understand the effectiveness of each enhancement. Specifically, we analyze the relationships between point cloud density, distance, IoU, and confidence scores. Additionally, we quantify the number of false positive and false negative samples across 10-meter distance intervals. Experimental results on the KITTI dataset demonstrate consistent gains in average precision, especially for hard samples, along with notable reductions in both false positives and false negatives.