Jiyuan Hu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Jiyuan Hu has a total of 7 co-authors. Here are the top co-authors by number of shared publications:
- Haibin Liu
Beijing University of Technology
3 shared publications - Huanjie Wang
Beijing University of Technology
3 shared publications - Wenshuo Wang
Beijing Institute of Technology
3 shared publications - Chao Wei
School of Mechanical Engineering, Beijing Institute of Technology, Beijing, China
1 shared publication - Chuan Hu
Shanghai Jiao Tong University
1 shared publication - Weida Wang
Beijing Institute of Technology
1 shared publication
Latest Publications
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 27 (11) , 14106-14117
Toward Trustworthy Autonomous Driving Through the Quantification and Propagation of Uncertainty From HD Maps to Trajectory Prediction
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 27 (11) , 14106-14117
Toward Trustworthy Autonomous Driving Through the Quantification and Propagation of Uncertainty From HD Maps to Trajectory Prediction
- 8 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. PP (00189545) , 1-15
Integration of Uncertainty-Aware Map Estimation and BEV Features for Robust Trajectory Prediction in Autonomous Driving
While the substantial cost of constructing and maintaining high-definition (HD) maps motivates the pursuit of online map generation for improved efficiency and cross-domain adaptability, current online maps inevitably suffer from geometric noise, occlusions, and other imperfections. These impacts can drive trajectory predictions beyond feasible boundaries and accumulate over extended horizons, thereby compromising the trustworthiness of downstream modules. The explicit modeling and robust handling of map uncertainty, however, remain largely unexplored. To address this gap, an uncertainty-aware trajectory prediction framework is presented for autonomous driving that integrates bird's-eye-view (BEV) features and quantifies map uncertainty to adaptively balance HD map geometric constraints against BEV evidence. This dual constraint paradigm, marrying uncertainty quantification with geometric priors of map structure, yields robust trajectory prediction in complex, real-world environments. Experimental results demonstrate that our method effectively contends with uncertainty across diverse, challenging settings and markedly improves both the accuracy and reliability of downstream trajectory prediction, with particularly notable improvements in highly interactive scenes and regions characterized by poor map quality.
- 8 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. PP (00189545) , 1-15
Integration of Uncertainty-Aware Map Estimation and BEV Features for Robust Trajectory Prediction in Autonomous Driving
While the substantial cost of constructing and maintaining high-definition (HD) maps motivates the pursuit of online map generation for improved efficiency and cross-domain adaptability, current online maps inevitably suffer from geometric noise, occlusions, and other imperfections. These impacts can drive trajectory predictions beyond feasible boundaries and accumulate over extended horizons, thereby compromising the trustworthiness of downstream modules. The explicit modeling and robust handling of map uncertainty, however, remain largely unexplored. To address this gap, an uncertainty-aware trajectory prediction framework is presented for autonomous driving that integrates bird's-eye-view (BEV) features and quantifies map uncertainty to adaptively balance HD map geometric constraints against BEV evidence. This dual constraint paradigm, marrying uncertainty quantification with geometric priors of map structure, yields robust trajectory prediction in complex, real-world environments. Experimental results demonstrate that our method effectively contends with uncertainty across diverse, challenging settings and markedly improves both the accuracy and reliability of downstream trajectory prediction, with particularly notable improvements in highly interactive scenes and regions characterized by poor map quality.
- 30 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 74 (9) , 13865-13877
Robust and High-Precision Point Cloud Registration Method Based on 3D-NDT Algorithm for Vehicle Localization
Point cloud registration plays a pivotal role in LiDAR-based vehicle localization, as its robustness and accuracy directly influence the map generation quality and localization precision. The 3D Normal Distributions Transform (3D-NDT) is a competitive registration algorithm that excels in handling noise and dynamic changes in complex environments. However, its performance is hindered by the blurring of local point cloud features caused by voxelization. To address this limitation, this paper proposes an improved 3D-NDT registration method that leverages point cloud normal vector segmentation and voxelization techniques to optimize the registration process, improving both accuracy and matching scope. Initially, a plane point cloud clustering technique is employed for preliminary voxel partitioning. This is followed by plane point cloud subdivision, which further refines the voxels to ensure a uniform voxel cell size, thereby appropriately weighting each voxel cell's contribution to the objective function. The proposed method employs a dual-stage voxel division strategy, utilizing two distinct voxel sizes tailored to different optimization phases. In the first stage, a larger grid size broadens the matching scope and enhances the efficiency of preliminary registration. The resulting output serves as the starting point for the refined registration phase, where a smaller grid size is used to improve matching precision. Experimental results demonstrate that the proposed method achieves a 50% reduction in median translation error (from $1.0 \times 10^{-3}$ m to $0.5 \times 10^{-3}$ m) and a 50% reduction in median rotation error (from 0.01° to 0.005°) compared to the original 3D-NDT algorithm. Furthermore, it significantly improves matching robustness and accuracy by extending the effective matching range by 100% for translation (from 1.2 m to 2.4 m) and 25.6% for rotation (from 39° to 49°).
- 30 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 74 (9) , 13865-13877
Robust and High-Precision Point Cloud Registration Method Based on 3D-NDT Algorithm for Vehicle Localization
Point cloud registration plays a pivotal role in LiDAR-based vehicle localization, as its robustness and accuracy directly influence the map generation quality and localization precision. The 3D Normal Distributions Transform (3D-NDT) is a competitive registration algorithm that excels in handling noise and dynamic changes in complex environments. However, its performance is hindered by the blurring of local point cloud features caused by voxelization. To address this limitation, this paper proposes an improved 3D-NDT registration method that leverages point cloud normal vector segmentation and voxelization techniques to optimize the registration process, improving both accuracy and matching scope. Initially, a plane point cloud clustering technique is employed for preliminary voxel partitioning. This is followed by plane point cloud subdivision, which further refines the voxels to ensure a uniform voxel cell size, thereby appropriately weighting each voxel cell's contribution to the objective function. The proposed method employs a dual-stage voxel division strategy, utilizing two distinct voxel sizes tailored to different optimization phases. In the first stage, a larger grid size broadens the matching scope and enhances the efficiency of preliminary registration. The resulting output serves as the starting point for the refined registration phase, where a smaller grid size is used to improve matching precision. Experimental results demonstrate that the proposed method achieves a 50% reduction in median translation error (from $1.0 \times 10^{-3}$ m to $0.5 \times 10^{-3}$ m) and a 50% reduction in median rotation error (from 0.01° to 0.005°) compared to the original 3D-NDT algorithm. Furthermore, it significantly improves matching robustness and accuracy by extending the effective matching range by 100% for translation (from 1.2 m to 2.4 m) and 25.6% for rotation (from 39° to 49°).