Wenbin Zhu
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
Wenbin Zhu has a total of 33 co-authors. Here are the top co-authors by number of shared publications:
- Jing Yuan
College of Artificial Intelligence, Nankai University, Tianjin, China
4 shared publications - Boran Wang
Nankai University, Tianjin, China and Zhongguancun Academy, Beijing, China
2 shared publications - Fei Chen
College of Artificial Intelligence, Nankai University, Tianjin, China
2 shared publications - Jiasheng He
College of Artificial Intelligence Nankai University 300350 Tianjin China
2 shared publications - Xuebo Zhang
Institute of Robotics and Automatic Information Systems and Tianjin Key Laboratory of Intelligent Robotics, Nankai University, Tianjin, China
2 shared publications - Yishuo Chen
College of Artificial Intelligence Nankai University 300350 Tianjin China
2 shared publications
Latest Publications
- 18 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Robotics
- Vol. 41 (15523098) , 3078-3098
Bridging the Gap Between Semantics and Geometry in SLAM: A Semantic-Geometric Tight-Coupling Monocular Visual Object SLAM System
Existing object-level SLAM methods often overlook the correspondence between semantic information and geometric features, resulting in a significant gap between them within SLAM frameworks. To tackle this issue, this paper proposes TiMoSLAM, a semantic-geometric tight-coupling monocular visual object SLAM system, which considers a rigorous correspondence between semantics and geometry across all steps of SLAM. Initially, a general Semantic Relation Graph (SRG) is developed to consistently represent semantic information alongside geometric features. Detailed analyses on complete constraints of the geometric feature combinations on estimation of 3D cuboid model are performed. Subsequently, a Compound Hypothesis Tree (CHT) is proposed to incrementally construct the object-specific SRG and concurrently estimate the 3D cuboid model of an object, ensuing semantic-geometric consistency in object representation and estimation. Special attention is given to the matching errors between geometric features and objects during the optimization of camera poses and object parameters. The effectiveness of this method is validated on various datasets, as well as in real-world environments.
- 18 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Robotics
- Vol. 41 (15523098) , 3078-3098
Bridging the Gap Between Semantics and Geometry in SLAM: A Semantic-Geometric Tight-Coupling Monocular Visual Object SLAM System
Existing object-level SLAM methods often overlook the correspondence between semantic information and geometric features, resulting in a significant gap between them within SLAM frameworks. To tackle this issue, this paper proposes TiMoSLAM, a semantic-geometric tight-coupling monocular visual object SLAM system, which considers a rigorous correspondence between semantics and geometry across all steps of SLAM. Initially, a general Semantic Relation Graph (SRG) is developed to consistently represent semantic information alongside geometric features. Detailed analyses on complete constraints of the geometric feature combinations on estimation of 3D cuboid model are performed. Subsequently, a Compound Hypothesis Tree (CHT) is proposed to incrementally construct the object-specific SRG and concurrently estimate the 3D cuboid model of an object, ensuing semantic-geometric consistency in object representation and estimation. Special attention is given to the matching errors between geometric features and objects during the optimization of camera poses and object parameters. The effectiveness of this method is validated on various datasets, as well as in real-world environments.
- Open Access
DEYOLO: Dual-Feature-Enhancement YOLO for Cross-Modality Object Detection
Object detection in poor-illumination environments is a challenging task as objects are usually not clearly visible in RGB images. As infrared images provide additional clear edge information that complements RGB images, fusing RGB and infrared images has potential to enhance the detection ability in poor-illumination environments. However, existing works involving both visible and infrared images only focus on image fusion, instead of object detection. Moreover, they directly fuse the two kinds of image modalities, which ignores the mutual interference between them. To fuse the two modalities to maximize the advantages of cross-modality, we design a dual-enhancement-based cross-modality object detection network DEYOLO, in which semantic-spatial cross modality and novel bi-directional decoupled focus modules are designed to achieve the detection-centered mutual enhancement of RGB-infrared (RGB-IR). Specifically, a dual semantic enhancing channel weight assignment module (DECA) and a dual spatial enhancing pixel weight assignment module (DEPA) are firstly proposed to aggregate cross-modality information in the feature space to improve the feature representation ability, such that feature fusion can aim at the object detection task. Meanwhile, a dual-enhancement mechanism, including enhancements for two-modality fusion and single modality, is designed in both DECAand DEPAto reduce interference between the two kinds of image modalities. Then, a novel bi-directional decoupled focus is developed to enlarge the receptive field of the backbone network in different directions, which improves the representation quality of DEYOLO. Extensive experiments on M3FD and LLVIP show that our approach outperforms SOTA object detection algorithms by a clear margin. Our code is available at https://github.com/chips96/DEYOLO.
Open AccessDEYOLO: Dual-Feature-Enhancement YOLO for Cross-Modality Object Detection
Object detection in poor-illumination environments is a challenging task as objects are usually not clearly visible in RGB images. As infrared images provide additional clear edge information that complements RGB images, fusing RGB and infrared images has potential to enhance the detection ability in poor-illumination environments. However, existing works involving both visible and infrared images only focus on image fusion, instead of object detection. Moreover, they directly fuse the two kinds of image modalities, which ignores the mutual interference between them. To fuse the two modalities to maximize the advantages of cross-modality, we design a dual-enhancement-based cross-modality object detection network DEYOLO, in which semantic-spatial cross modality and novel bi-directional decoupled focus modules are designed to achieve the detection-centered mutual enhancement of RGB-infrared (RGB-IR). Specifically, a dual semantic enhancing channel weight assignment module (DECA) and a dual spatial enhancing pixel weight assignment module (DEPA) are firstly proposed to aggregate cross-modality information in the feature space to improve the feature representation ability, such that feature fusion can aim at the object detection task. Meanwhile, a dual-enhancement mechanism, including enhancements for two-modality fusion and single modality, is designed in both DECAand DEPAto reduce interference between the two kinds of image modalities. Then, a novel bi-directional decoupled focus is developed to enlarge the receptive field of the backbone network in different directions, which improves the representation quality of DEYOLO. Extensive experiments on M3FD and LLVIP show that our approach outperforms SOTA object detection algorithms by a clear margin. Our code is available at https://github.com/chips96/DEYOLO.
- 3 December 2024
- Published by Springer Nature
- p. 236-252
DEYOLO: Dual-Feature-Enhancement YOLO for Cross-Modality Object Detection
- 3 December 2024
- Published by Springer Nature
- p. 236-252
DEYOLO: Dual-Feature-Enhancement YOLO for Cross-Modality Object Detection
- 23 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 72 (00189456) , 1-15
Adaptive Soft Encoding: A General Unsupervised Feature Aggregation Method for Place Recognition
Place recognition is an important problem in environment perception, which can help simultaneous localization and mapping (SLAM), as well as scene understanding of mobile robots. For place recognition, feature aggregation plays an important role in the representation of environments. In this paper, an adaptive soft-encoding (ASE) method is proposed for aggregating numerous features into low-dimensional feature vectors, which includes a training phase and an encoding phase. For the training phase, the information along each dimension of the features is evaluated which is further employed to assign all dimensions into different subdimensional intervals. After that, subdimensional Gaussian mixture models (SD-GMMs) are fitted by features in the subdimensional intervals and organized into an ASE tree, of which the tree nodes hierarchically store the parameters of the SD-GMMs to reflect the distribution of the features. For the encoding phase, features are fed into the root node of the ASE tree. The weight of each node is calculated from top to bottom to generate multiple aggregated feature vectors with different description capabilities. In addition, an ASE-based framework for place recognition is proposed in which local descriptors are extracted from sensor data (e.g., images or LiDAR data), and then aggregated into global descriptors by ASE. Finally, owing to the hierarchical structure of the ASE tree, a hierarchical matching strategy of the global descriptors is designed to recognize places efficiently. Experimental results demonstrate that feature vectors aggregated by ASE have strong distinguishability, and the ASE-based place recognition method achieves good accuracy and efficiency. The code of ASE can be accessed at https://github.com/wdyiwdwd/ASE-Encoder.
- 23 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 72 (00189456) , 1-15
Adaptive Soft Encoding: A General Unsupervised Feature Aggregation Method for Place Recognition
Place recognition is an important problem in environment perception, which can help simultaneous localization and mapping (SLAM), as well as scene understanding of mobile robots. For place recognition, feature aggregation plays an important role in the representation of environments. In this paper, an adaptive soft-encoding (ASE) method is proposed for aggregating numerous features into low-dimensional feature vectors, which includes a training phase and an encoding phase. For the training phase, the information along each dimension of the features is evaluated which is further employed to assign all dimensions into different subdimensional intervals. After that, subdimensional Gaussian mixture models (SD-GMMs) are fitted by features in the subdimensional intervals and organized into an ASE tree, of which the tree nodes hierarchically store the parameters of the SD-GMMs to reflect the distribution of the features. For the encoding phase, features are fed into the root node of the ASE tree. The weight of each node is calculated from top to bottom to generate multiple aggregated feature vectors with different description capabilities. In addition, an ASE-based framework for place recognition is proposed in which local descriptors are extracted from sensor data (e.g., images or LiDAR data), and then aggregated into global descriptors by ASE. Finally, owing to the hierarchical structure of the ASE tree, a hierarchical matching strategy of the global descriptors is designed to recognize places efficiently. Experimental results demonstrate that feature vectors aggregated by ASE have strong distinguishability, and the ASE-based place recognition method achieves good accuracy and efficiency. The code of ASE can be accessed at https://github.com/wdyiwdwd/ASE-Encoder.
- Open Access
- 19 July 2018
- Published by Frontiers Media SA in Frontiers in Neurology
- Vol. 9, 583
The Epidemiological Characteristics of Stroke in Hunan Province, China
Previous studies have shown that Hunan Province has a high incidence of stroke and a high proportion of intracerebral hemorrhage (ICH). Considering the changes over the past three decades, little is known about the current epidemiological characteristics of stroke in Hunan Province. In 2013, a cross-sectional study was conducted at seven national disease surveillance points (DSPs) in Hunan Province. A multistage cluster sampling method was used to select a representative sample. A total of 21,156 participants aged 20 years and older were examined. Among the 21,156 participants, the number of prevalent strokes, incident strokes and deaths was 307, 87, and 36, respectively. The 2010 China census-standardized prevalence, incidence and mortality were 1191.0 per 100,000 people [95% confidence interval (CI) 1044.8–1337.2], 333.6 per 100,000 person-years (95% CI 255.7–411.5) and 129.7 per 100,000 person-years (95% CI 81.1–178.3), respectively. Ischemic stroke (IS), ICH, subarachnoid hemorrhage (SAH), and stroke of undetermined type (UND) constituted 50.6, 41.4, 5.7, and 2.3% of all incident stroke cases, respectively. Tianxin, Liuyang, Wuling, and Hongjiang have high proportions of ICH (61.5, 58.3, 60, and 50%, respectively). Hypertension is the most common risk factor for prevalent stroke (71.34%), followed by smoking (30.62%) and alcohol use (25.73%). In conclusion, Hunan Province has an extremely heavy stroke burden. The high proportion of ICH is not limited to the Changsha community; it represents an important issue for all of Hunan Province.
- 19 July 2018
- Published by Frontiers Media SA in Frontiers in Neurology
- Vol. 9, 583
Open AccessThe Epidemiological Characteristics of Stroke in Hunan Province, China
Previous studies have shown that Hunan Province has a high incidence of stroke and a high proportion of intracerebral hemorrhage (ICH). Considering the changes over the past three decades, little is known about the current epidemiological characteristics of stroke in Hunan Province. In 2013, a cross-sectional study was conducted at seven national disease surveillance points (DSPs) in Hunan Province. A multistage cluster sampling method was used to select a representative sample. A total of 21,156 participants aged 20 years and older were examined. Among the 21,156 participants, the number of prevalent strokes, incident strokes and deaths was 307, 87, and 36, respectively. The 2010 China census-standardized prevalence, incidence and mortality were 1191.0 per 100,000 people [95% confidence interval (CI) 1044.8–1337.2], 333.6 per 100,000 person-years (95% CI 255.7–411.5) and 129.7 per 100,000 person-years (95% CI 81.1–178.3), respectively. Ischemic stroke (IS), ICH, subarachnoid hemorrhage (SAH), and stroke of undetermined type (UND) constituted 50.6, 41.4, 5.7, and 2.3% of all incident stroke cases, respectively. Tianxin, Liuyang, Wuling, and Hongjiang have high proportions of ICH (61.5, 58.3, 60, and 50%, respectively). Hypertension is the most common risk factor for prevalent stroke (71.34%), followed by smoking (30.62%) and alcohol use (25.73%). In conclusion, Hunan Province has an extremely heavy stroke burden. The high proportion of ICH is not limited to the Changsha community; it represents an important issue for all of Hunan Province.