Jie Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Jie Zhang has a total of 37 co-authors. Here are the top co-authors by number of shared publications:
- Junhua Sun
School of Instrumentation and Optoelectronic Engineering, Beihang University, Beijing, China
10 shared publications - Xianzhong Chen
Key Laboratory of Knowledge Automation for Industrial Processes of Ministry of Education, School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing, China
6 shared publications - Qingwen Hou
Key Laboratory of Knowledge Automation for Industrial Processes of Ministry of Education, School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing, China
5 shared publications - Lifu Deng
Key Laboratory of Knowledge Automation for Industrial Processes of Ministry of Education, School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing, China
2 shared publications - Qing Yu
School of Automation and Electrical Engineering, University of Science and Technology Beijing
2 shared publications - Ziming Ni
University of Science and Technology Beijing
2 shared publications
Latest Publications
- 1 September 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (35) , 356001
Local-to-global viewpoint selection based on geometric models for robotic 3D object scanning
Three-dimensional surface measurement of complex objects is critical for industrial component manufacturing, inspection, and sorting. Robotic arm-mounted 3D scanning systems exhibit remarkable advantages in efficiency and repeatability, while requiring effective global viewpoint selection before path planning. In this paper, we propose a primitive-based local-to-global viewpoint selection method for offline viewpoint planning of complex objects with known models. Our two-stage framework consists of bottom-up local viewpoint generation and top-down global optimization. Firstly, an initial viewpoint set is generated for surface facets. We propose a local viewpoint refinement method that groups neighboring facets into a local primitive, thereby optimizing the initial viewpoint through geometry-aware weighted adjustment of its position and orientation. Then, we establish a bidirectional viewpoint-primitive coverage relationship for global optimization, yielding a compact viewpoint set while preserving scanning coverage. By leveraging neighboring geometry and mutual coverage, our method improves viewpoint compactness and surface coverage. Unlike next-best-view methods, we focus on offline global viewpoint selection rather than online dynamic planning. Experiments on real 3D objects show our method outperforms selected representative baselines, achieving 97.81% coverage with fewer number of viewpoints, and maintains higher computational efficiency of 38.98 seconds.
- 1 September 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (35) , 356001
Local-to-global viewpoint selection based on geometric models for robotic 3D object scanning
Three-dimensional surface measurement of complex objects is critical for industrial component manufacturing, inspection, and sorting. Robotic arm-mounted 3D scanning systems exhibit remarkable advantages in efficiency and repeatability, while requiring effective global viewpoint selection before path planning. In this paper, we propose a primitive-based local-to-global viewpoint selection method for offline viewpoint planning of complex objects with known models. Our two-stage framework consists of bottom-up local viewpoint generation and top-down global optimization. Firstly, an initial viewpoint set is generated for surface facets. We propose a local viewpoint refinement method that groups neighboring facets into a local primitive, thereby optimizing the initial viewpoint through geometry-aware weighted adjustment of its position and orientation. Then, we establish a bidirectional viewpoint-primitive coverage relationship for global optimization, yielding a compact viewpoint set while preserving scanning coverage. By leveraging neighboring geometry and mutual coverage, our method improves viewpoint compactness and surface coverage. Unlike next-best-view methods, we focus on offline global viewpoint selection rather than online dynamic planning. Experiments on real 3D objects show our method outperforms selected representative baselines, achieving 97.81% coverage with fewer number of viewpoints, and maintains higher computational efficiency of 38.98 seconds.
- 9 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Automation Science and Engineering
- Vol. 23 (15455955) , 8072-8085
LGF-Net: Local Geometry-Aware Fusion Network for Cross-Modal 3D Semantic Segmentation
3D LiDAR semantic segmentation is crucial for autonomous driving, yet its performance is constrained by point cloud sparsity. While cross-modal Knowledge Distillation (KD) from camera images can improve accuracy, its potential is limited by shallow fusion mechanisms in teacher networks. To address this, we propose LGF-Net, a framework that enhances KD efficiency through superior 3D geometric representation and effective cross-modal fusion. The framework introduces a Local Structure-aware Encoder (LSE) that extracts explicit geometric descriptors via eigenvalue decomposition of local neighborhood covariance matrices, providing discriminative surface priors for the 3D backbone. Building on these priors, the Grouped Bottleneck Attention Fusion Module (GBA) employs a deformable cross-attention mechanism where geometric features serve as positional encodings to adaptively aggregate semantically relevant 2D image features for each 3D point, moving beyond rigid point-to-pixel projection and yielding a high-performance teacher network. Multi-scale KL divergence-based distillation subsequently transfers the teacher’s cross-modal knowledge to the LSE-enhanced student network, which relies solely on point clouds during inference. Experiments on SemanticKITTI demonstrate significant accuracy improvements, and its generalization is further validated on our newly constructed industrial dataset, IIR-MMSD, confirming the effectiveness of the proposed method.
- 9 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Automation Science and Engineering
- Vol. 23 (15455955) , 8072-8085
LGF-Net: Local Geometry-Aware Fusion Network for Cross-Modal 3D Semantic Segmentation
3D LiDAR semantic segmentation is crucial for autonomous driving, yet its performance is constrained by point cloud sparsity. While cross-modal Knowledge Distillation (KD) from camera images can improve accuracy, its potential is limited by shallow fusion mechanisms in teacher networks. To address this, we propose LGF-Net, a framework that enhances KD efficiency through superior 3D geometric representation and effective cross-modal fusion. The framework introduces a Local Structure-aware Encoder (LSE) that extracts explicit geometric descriptors via eigenvalue decomposition of local neighborhood covariance matrices, providing discriminative surface priors for the 3D backbone. Building on these priors, the Grouped Bottleneck Attention Fusion Module (GBA) employs a deformable cross-attention mechanism where geometric features serve as positional encodings to adaptively aggregate semantically relevant 2D image features for each 3D point, moving beyond rigid point-to-pixel projection and yielding a high-performance teacher network. Multi-scale KL divergence-based distillation subsequently transfers the teacher’s cross-modal knowledge to the LSE-enhanced student network, which relies solely on point clouds during inference. Experiments on SemanticKITTI demonstrate significant accuracy improvements, and its generalization is further validated on our newly constructed industrial dataset, IIR-MMSD, confirming the effectiveness of the proposed method.
- 6 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 1-10
Beam-Steering Radar Imaging for Blast Furnace Burden Surface Based on Sparse Bayesian Learning in Time-Constrained Scenario
Accurate measurement of the burden surface plays a crucial role in optimizing blast furnace charging, reducing carbon emissions, and decreasing energy consumption. During the production process, the burden distribution cycle dynamically changes. When abnormal conditions occur in the blast furnace, the burden charge cycle is shortened, creating a time-constrained scenario for measurement of the burden surface. Increasing the speed of beam steering leads to the problem of sparse aperture in radar echoes. Sparse aperture results in a lack of radar echo consistency, leads to the “scallop” effect and affecting the accuracy of reconstruction. To address the problem, we propose a novel Bayesian imaging method for burden surface based on the beam steering radar mechanism. Firstly, the mechanism model of beam steering radar and Gaussian-Laplace mixed guidance prior are introduced, and the measurement matrix of beam steering radar is proposed. Secondly, an alternative function is introduced to approximate the Gaussian likelihood, which leads to a relaxation of the posterior density of the model parameters to improve the speed of reconstruction. Finally, the accuracy and efficiency of the proposed fast algorithm are verified by experimental results obtained from measured data. The experimental results on the W-shaped burden surface show that our method achieves 49%, 74%, and 56.1% improvements in RMSE, MAE, and Entropy, respectively, compared to the traditional sparse Bayesian learning method.
- 6 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 1-10
Beam-Steering Radar Imaging for Blast Furnace Burden Surface Based on Sparse Bayesian Learning in Time-Constrained Scenario
Accurate measurement of the burden surface plays a crucial role in optimizing blast furnace charging, reducing carbon emissions, and decreasing energy consumption. During the production process, the burden distribution cycle dynamically changes. When abnormal conditions occur in the blast furnace, the burden charge cycle is shortened, creating a time-constrained scenario for measurement of the burden surface. Increasing the speed of beam steering leads to the problem of sparse aperture in radar echoes. Sparse aperture results in a lack of radar echo consistency, leads to the “scallop” effect and affecting the accuracy of reconstruction. To address the problem, we propose a novel Bayesian imaging method for burden surface based on the beam steering radar mechanism. Firstly, the mechanism model of beam steering radar and Gaussian-Laplace mixed guidance prior are introduced, and the measurement matrix of beam steering radar is proposed. Secondly, an alternative function is introduced to approximate the Gaussian likelihood, which leads to a relaxation of the posterior density of the model parameters to improve the speed of reconstruction. Finally, the accuracy and efficiency of the proposed fast algorithm are verified by experimental results obtained from measured data. The experimental results on the W-shaped burden surface show that our method achieves 49%, 74%, and 56.1% improvements in RMSE, MAE, and Entropy, respectively, compared to the traditional sparse Bayesian learning method.
- 4 March 2026
- Published by Elsevier in Neurocomputing
- Vol. 679
MFI-Net: Multi-stage feature interaction network for fine-grained object recognition in point clouds
- 4 March 2026
- Published by Elsevier in Neurocomputing
- Vol. 679
MFI-Net: Multi-stage feature interaction network for fine-grained object recognition in point clouds
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (4) , 6128-6136
A Pose Measurement Method for Vision Sensor in Slender Structures
The 3-D multiview internal reconstruction of the slender pipeline is crucial for the internal surface inspection in industrial fields, where the pose of the reconstruction sensor directly affects the global stitching accuracy of the multiple views. However, conventional methods suffer from cumulative errors over a long distance. This article proposes a pose measurement method via multisensor fusion for the vision reconstruction system in a slender structure. The process integrates a laser rangefinder for displacement measurement and an inclinometer for attitude determination, with spatial correlation achieved through a monocular camera tracking the laser spot projected onto a flat plate. Accordingly, we propose a combined calibration algorithm for the pose measurement system, using an auxiliary camera and a double-sided target. The experimental validation on a 200-mm-wide and 6-m-long platform shows that the proposed multisensor fusion method achieves an root mean square error (RMSE)of 1.34 mm for 3-D marker point reconstruction at a moving distance of 5 m.
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (4) , 6128-6136
A Pose Measurement Method for Vision Sensor in Slender Structures
The 3-D multiview internal reconstruction of the slender pipeline is crucial for the internal surface inspection in industrial fields, where the pose of the reconstruction sensor directly affects the global stitching accuracy of the multiple views. However, conventional methods suffer from cumulative errors over a long distance. This article proposes a pose measurement method via multisensor fusion for the vision reconstruction system in a slender structure. The process integrates a laser rangefinder for displacement measurement and an inclinometer for attitude determination, with spatial correlation achieved through a monocular camera tracking the laser spot projected onto a flat plate. Accordingly, we propose a combined calibration algorithm for the pose measurement system, using an auxiliary camera and a double-sided target. The experimental validation on a 200-mm-wide and 6-m-long platform shows that the proposed multisensor fusion method achieves an root mean square error (RMSE)of 1.34 mm for 3-D marker point reconstruction at a moving distance of 5 m.
- Open Access
- 15 December 2025
- Published by Iron and Steel Institute of Japan in ISIJ International
- Vol. 65 (13) , 2204-2212
Electromagnetic Scattering Characteristics of Random Roughness Burden Surfaces in Blast Furnace Aerosol Dusts
Addressing the propagation characteristics of radar waves in the dusty environment of blast furnaces, this paper develops an electromagnetic scattering model for the burden surface that incorporates random roughness and stockline shapes. By obtaining roughness parameters from actual samples, we establish a geometric model of the burden surface that integrates both form and texture. Utilizing the radiative transfer equation and integral equation methods, we construct a composite scattering model for the dust-burden surface interaction. The study examines the effects of surface roughness and slope on scattering characteristics and, by comparing with measured data, verifies the model's effectiveness, providing theoretical support for the inversion of radar data.
- 15 December 2025
- Published by Iron and Steel Institute of Japan in ISIJ International
- Vol. 65 (13) , 2204-2212
Open AccessElectromagnetic Scattering Characteristics of Random Roughness Burden Surfaces in Blast Furnace Aerosol Dusts
Addressing the propagation characteristics of radar waves in the dusty environment of blast furnaces, this paper develops an electromagnetic scattering model for the burden surface that incorporates random roughness and stockline shapes. By obtaining roughness parameters from actual samples, we establish a geometric model of the burden surface that integrates both form and texture. Utilizing the radiative transfer equation and integral equation methods, we construct a composite scattering model for the dust-burden surface interaction. The study examines the effects of surface roughness and slope on scattering characteristics and, by comparing with measured data, verifies the model's effectiveness, providing theoretical support for the inversion of radar data.
- 22 October 2025
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 36 (10) , 105013
Semantic-to-instance top-down analysis of complex industrial pipelines in 3D point cloud scenarios
Pipelines are common and critical components in some complex industrial equipment, characterized by intricate bending patterns and hierarchical distributions across irregular surfaces. Segmenting pipeline instances from the background for further inspection is a challenging task. To address this challenge, we propose a comprehensive top-down framework that follows a semantic-to-instance integrated task flow for pipeline analysis. We first design a deep neural network model that utilizes both positional embedding and pipeline-aware geometric features to distinguish bent and long pipeline point clouds from the background. Then, we introduce an enhanced axis-growing algorithm that incorporates robustness optimization strategies to achieve precise instance-level axis reconstruction and recover key geometric parameters. Furthermore, we develop a simple yet effective model for accurate minimum clearance measurement between adjacent pipelines. The experimental results on real-world point cloud datasets demonstrate the effectiveness of the proposed framework. The pipeline segmentation achieves a superior accuracy of 95.89%. The minimum clearance measurement reaches mean accuracies of 0.028 mm for straight pipelines and 0.08 mm for bent pipelines, respectively.
- 22 October 2025
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 36 (10) , 105013
Semantic-to-instance top-down analysis of complex industrial pipelines in 3D point cloud scenarios
Pipelines are common and critical components in some complex industrial equipment, characterized by intricate bending patterns and hierarchical distributions across irregular surfaces. Segmenting pipeline instances from the background for further inspection is a challenging task. To address this challenge, we propose a comprehensive top-down framework that follows a semantic-to-instance integrated task flow for pipeline analysis. We first design a deep neural network model that utilizes both positional embedding and pipeline-aware geometric features to distinguish bent and long pipeline point clouds from the background. Then, we introduce an enhanced axis-growing algorithm that incorporates robustness optimization strategies to achieve precise instance-level axis reconstruction and recover key geometric parameters. Furthermore, we develop a simple yet effective model for accurate minimum clearance measurement between adjacent pipelines. The experimental results on real-world point cloud datasets demonstrate the effectiveness of the proposed framework. The pipeline segmentation achieves a superior accuracy of 95.89%. The minimum clearance measurement reaches mean accuracies of 0.028 mm for straight pipelines and 0.08 mm for bent pipelines, respectively.
- 30 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-10
Multimodal Transformation for Small-Scale 3-D Object Detection in Industrial Scenarios
Detecting small and information-scarce objects within complex 3D backgrounds remains a critical yet challenging task in industrial scenarios. While existing multi-modal approaches leverage the cross-modal complementation between point clouds and intensity data to improve object representation, they still face limitations in addressing multi-modal fine alignment and achieving precise 3D positional regression. This paper proposes a new 3D object detection network based on feature fusion between point clouds and multi-view images. First, we propose a dense multi-modal feature fusion module that establishes point-to-pixel fine correspondence and effectively integrates multi-modal feature channels. Then, we design a normalized 3D positional embedding generator to enhance a transformer-based detection head, which improves localization accuracy through refined positional encoding of the fused features. Experimental results on a multi-modal industrial dataset demonstrate that the proposed method achieves state-of-the-art performance with an AP of 0.903 and a Recall of 94.10%, representing 6.35% and 3.58% improvements in comparison to the existing optimal method. Specifically, the method achieves optimal performance of 2.66 mm and 0.85 mm on the proposed metrics 3D mATE and 3D mASE, representing 19.64%, and 29.75% improvement over the existing optimal method, respectively.
- 30 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-10
Multimodal Transformation for Small-Scale 3-D Object Detection in Industrial Scenarios
Detecting small and information-scarce objects within complex 3D backgrounds remains a critical yet challenging task in industrial scenarios. While existing multi-modal approaches leverage the cross-modal complementation between point clouds and intensity data to improve object representation, they still face limitations in addressing multi-modal fine alignment and achieving precise 3D positional regression. This paper proposes a new 3D object detection network based on feature fusion between point clouds and multi-view images. First, we propose a dense multi-modal feature fusion module that establishes point-to-pixel fine correspondence and effectively integrates multi-modal feature channels. Then, we design a normalized 3D positional embedding generator to enhance a transformer-based detection head, which improves localization accuracy through refined positional encoding of the fused features. Experimental results on a multi-modal industrial dataset demonstrate that the proposed method achieves state-of-the-art performance with an AP of 0.903 and a Recall of 94.10%, representing 6.35% and 3.58% improvements in comparison to the existing optimal method. Specifically, the method achieves optimal performance of 2.66 mm and 0.85 mm on the proposed metrics 3D mATE and 3D mASE, representing 19.64%, and 29.75% improvement over the existing optimal method, respectively.
- 13 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Antennas and Wireless Propagation Letters
- Vol. 24 (2) , 369-373
Electromagnetic Scattering of Blast Furnace Burden Surface With Stochastic Properties
Microwave radars have been utilized as detectors within blast furnaces to accomplish the measurement of the surface shape of the furnace burden. However, the additional physical information contained in the radar echoes has not yet been fully explored. The electromagnetic scattering model of the burden surface can serve as the theoretical foundation for the inversion of radar data. In this letter, firstly, the statistical values of the roughness parameters of the burden surface are obtained by measuring typical samples. Secondly, an equivalent dielectric constant model is proposed for the coexistence of coke, ore, water and gas on the burden surface. Finally, the back-scattering coefficient of the burden surface was calculated by the integral equation method (IEM) based on the roughness parameters and geometric characteristics of the burden surface. The effects of charge size and volume fractions of ore, coke and water on the electromagnetic scattering characteristics were analyzed. The experimental results indicate that for the blast furnace burden surface where coke, ore, and gas coexist, when the angle of incidence is $30^{\circ }$ and the root mean square height is less than 0.03 meters, the back-scattering is enhanced as the roughness of the burden surface increases. In comparison, for the surface of coke material, the moisture content has a greater impact on electromagnetic scattering than the incorporation of ore with the same volume fraction.
- 13 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Antennas and Wireless Propagation Letters
- Vol. 24 (2) , 369-373
Electromagnetic Scattering of Blast Furnace Burden Surface With Stochastic Properties
Microwave radars have been utilized as detectors within blast furnaces to accomplish the measurement of the surface shape of the furnace burden. However, the additional physical information contained in the radar echoes has not yet been fully explored. The electromagnetic scattering model of the burden surface can serve as the theoretical foundation for the inversion of radar data. In this letter, firstly, the statistical values of the roughness parameters of the burden surface are obtained by measuring typical samples. Secondly, an equivalent dielectric constant model is proposed for the coexistence of coke, ore, water and gas on the burden surface. Finally, the back-scattering coefficient of the burden surface was calculated by the integral equation method (IEM) based on the roughness parameters and geometric characteristics of the burden surface. The effects of charge size and volume fractions of ore, coke and water on the electromagnetic scattering characteristics were analyzed. The experimental results indicate that for the blast furnace burden surface where coke, ore, and gas coexist, when the angle of incidence is $30^{\circ }$ and the root mean square height is less than 0.03 meters, the back-scattering is enhanced as the roughness of the burden surface increases. In comparison, for the surface of coke material, the moisture content has a greater impact on electromagnetic scattering than the incorporation of ore with the same volume fraction.