With the development of autonomous driving technology, autonomous vehicles have gradually moved from ideal highways to more realistic urban environments. Compared to highways, urban roads have various traffic participants. 3D object detection is of great significance to the safe driving and path planning of autonomous vehicles while LiDAR becoming an important primary 3D object detection sensor. But long-distance and small objects pose challenges for LiDAR’s data processing. Our proposed solution, Voxel R-CNN-MCA, is trained on One millioN sCenEs (ONCE) dataset in order to be closer to the urban traffic situation in China. Its baseline is Voxel R-CNN, which is an efficient voxel-based 3D detector. We add Multi-scale Coordinate Attention (MCA) to the baseline, which embeds multi-scale positional information into channel attention. Besides, residual structure is added to 3D sparse convolution backbone. We use the ONCE dataset to evaluate the proposed model. The results show that the overall mAP value of this model is 59.7%, 3.36% higher than the baseline. Our innovation considerably enhances the detection performance for long-range and small objects as compared to the baseline. Experimental results indicate that the model achieves state-of-art on the ONCE dataset.
Voxel R-CNN-MCA: Coordinate Attention for 3D Object Detection
Lect. Notes Electrical Eng.
International Conference on Autonomous Unmanned Systems ; 2022 ; Xi'an, China September 23, 2022 - September 25, 2022
Proceedings of 2022 International Conference on Autonomous Unmanned Systems (ICAUS 2022) ; Kapitel : 164 ; 1766-1776
2023-03-10
11 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
Parking method and device based on voxel coordinate system
Europäisches Patentamt | 2023
|Point-Voxel Fusion for Multimodal 3D Detection
IEEE | 2022
|