Hengwang Zhao

Hengwang Zhao

Auto-generated profile, based on publicly available information

Hengwang Zhao

Department of Automation, Shanghai Jiao Tong University, Shanghai, China

Research Focus & Metrics

Academic Career Length

6Years (2021-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
23 February 2021 in IEEE Robotics and Automation Letters
Verified
Journal Article
Closed access
HZHengwang Zhao
Hengwang Zhao
ZLZhidong Liang
Zhidong Liang
CWChunxiang Wang
Chunxiang Wang
MYMing Yang
Ming Yang
Point cloud registration is a key problem for robotics, computer vision, and other applications. Previous global registration algorithms are sensitive to noises or partial occlusion, while local registration algorithms are highly dependent on initial angles. To solve these problems, we propose CentroidReg, a global-to-local framework for partial point cloud registration, desensitizing noises and initial angles. The partial point clouds have consistent global information (the centroid of the complete 3D shape) and local surface structures, which motivates us to combine rotation-invariant global and local information for correspondence generating. To estimate the centroid of the complete 3D shape for partial point clouds, we propose the CentroidNet with centroid voting and maximum likelihood estimation (MLE) aggregation. Then, we use the proposed RegistrationNet to extract hybrid rotation-invariant features (HRIF) combining global information of the entire 3D shape and local structures in the local area to describe each point. Furthermore, we propose the Reliable Correspondence Generating (RCG) module to generate correspondences in the overlapping area. Finally, a weighted differentiable singular value decomposition (SVD) step is performed to estimate the final rigid transformation. The proposed framework presents state-of-the-art performance on partial point cloud registration. Experiments show that our method is robust to initial angles and noises.

13 January 2022 in IEEE Robotics and Automation Letters
Verified
Journal Article
Closed access
HZHengwang Zhao
Hengwang Zhao
HZHanyang Zhuang
Hanyang Zhuang
CWChunxiang Wang
Chunxiang Wang
MYMing Yang
Ming Yang
Point cloud registration (PCR) is a key problem for robotics, autonomous driving, and other applications. Constructing generalizable 3D descriptors and determining whether a 3D descriptor is in the overlapping area are challenging tasks in PCR. Despite the fast evolution of learning-based 3D descriptors, existing methods are either sensitive to rigid transformation and scenario changes, or can not find the appropriate descriptors in the overlapping area, which means their generalization and descriptive ability are not enough for practice applications. To solve these problems, we propose a novel neural network, named G3DOA, which jointly learns generalizable rotation-invariant 3D descriptors and their overlap scores (representing the probability in the overlapping area), to enhance the generalization ability of the descriptors across different data collected by various laser sensors. To ensure the rotation and scale invariance of the point cloud in the input stage, we estimate the Local Reference Frame (LRF) of local patches and normalize the coordinates. To learn generalizable and distinctive descriptors, we propose a novel Cylindrical LRF convolution module with multi-scaled cylindrical shells and neural layers, which hierarchically encodes and aggregates the geometric information in different cylindrical shells. Moreover, to estimate the probability of whether a point is in the overlapping area, we propose an overlap attention module that extracts co-contextual information between the feature encodings of the two point clouds. The experiments show that G3DOA trained only on an indoor dataset can be efficiently generalized to complex outdoor datasets, and the generalization ability of G3DOA outperforms state-of-the-art learning-based 3D descriptors.

31 July 2025 in IEEE Transactions on Intelligent Transportation Systems
Verified
Journal Article
Closed access
HZHengwang Zhao
Hengwang Zhao
QSQiyuan Shen
Qiyuan Shen
HZHanyang Zhuang
Hanyang Zhuang
TQTong Qin
Tong Qin
CWChunxiang Wang
Chunxiang Wang
MYMing Yang
Ming Yang
Cross-modal global localization matches visual information with pre-built LiDAR maps, which has attracted more and more attention for its low cost and potential robustness. However, the inherent modality difference between images and point clouds makes it challenging. This paper proposes a novel cross-modal global localization system, named CrossGLoc, which leverages pre-trained diffusion models and semantic cues to address this challenge. The main idea is leveraging the semantic cues shared between different modalities to bridge the modality gap, and utilizing pre-trained diffusion models to extract modality-consistent high-dimensional features guided by these semantic cues. To achieve this, ControlNet is used to generate intermediate feature maps from semantic images and semantic map projections, and a semantic categories-based feature aggregation algorithm is proposed to aggregate these feature maps into global descriptors. Furthermore, a semantic edge key points-based pose estimation algorithm is proposed to estimate the pose of retrieved image and point cloud pairs. Extensive experiments on the KITTI dataset, the KITTI360 dataset and the self-collected dataset demonstrate that the proposed method achieves state-of-the-art performance in cross-modal global localization.