中国科学院自动化研究所机构知识库

With the development of intelligent and connected vehicles, RSU (roadside unit) sensors are playing an increasingly important role for environment perception. For vehicle detection in autonomous mine, lack of diversity data on RSU LiDAR limits the application of deep learning based methods. To solve this issue, a voxel-based background filtering module is introduced into 3D object detectors for vehicle detection with RSU LiDAR in mine environments. The proposed background filtering method models average height and the number of points for each voxel as Gaussian distribution to generate a background table. To address the impact of the false negative points of the background filtering module, we also propose a multivariate Gaussian loss to model bounding box uncertainty. The predicted covariances between variates help to learn the relationship between the missed parts and the visible ones. Besides, a background filtering based data augmentation method for vehicle detection is also proposed in this paper. Three RSU LiDAR datasets with different terrains in the BaoLi mine area are used for comprehensive experiment evaluations. Experiments show that the proposed background filtering module and multivariate Gaussian loss can significantly improve the generalization ability and performance of several state-of-the-art 3D detectors on different terrain data. Moreover, most background voxels are filtered out, the inference time of the 3D detectors is about 2x faster. Besides, the effectiveness of the proposed data augmentation method is also demonstrated.