Multi-LiDAR-Based 3D Object Detection via Data-Level Fusion Method

Abstract

With the rapid development of artificial intelligence, the application prospect of the global perception system that can cover large-scale scenarios in smart cities is becoming increasingly extensive. However, due to the sparse point cloud data at the remote end and the complex logic of result-level stitch, most of the current LiDAR-based global perception technology performs poorly. Based on the above problems, we first propose a technology for mosaic point clouds from the data level, improving remote targets’ point cloud density. Secondly, a new point cloud detection model based on the deep learning framework is proposed, which enhances the feature extraction ability of small targets with sparse point clouds at the intersection of perception stations by focusing on features by attention mechanism. Moreover, we add the direction accuracy by changing the heading angle prediction to interval prediction. Finally, to verify the effectiveness of our method, we propose the public dataset VANJEE Point Cloud, which is collected in the real world. Our algorithm has improved the global trajectory fusion rate from 91.7% to 95.6%. Experiments prove the effectiveness of the proposed method in this paper.