Deep learning-based 3D target detection algorithm

Abstract

Automated vehicle driving requires a heightened awareness of the surrounding environment, and detecting targets is a crucial element in reducing the risk of traffic accidents. Target detection is essential for autonomous driving. In this paper, we improve the CenterPoint 3D target detection algorithm by introducing a self-calibrating convolutional network into the 2D backbone network of the original algorithm. This enhancement improves network extraction speed and feature extraction capability. Additionally, we improve the two-stage refinement module of the original algorithm by extracting feature points from the multi-scale feature map rather than the single-scale feature map. This approach reduces the loss of small target feature information, and we build a data enhancement module to increase the number of training samples and improve the network model’s robustness. We validate the algorithm on the KITTI dataset and analyze domestic data visualizations. Our results show that the bird’s-eye view mAP detection accuracy of the algorithm when the target is a vehicle has improved by 1.68%, and the 3D target mAP detection accuracy has improved by 1.02% compared with the original algorithm.