Songyue Yang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Songyue Yang has a total of 34 co-authors. Here are the top co-authors by number of shared publications:
- Zhangyu Wang
Beihang University
15 shared publications - Guizhen Yu
Beihang University
14 shared publications - Peng Chen
Beihang University
8 shared publications - Bin Zhou
School of Transportation Science and Engineering, Beihang University, Beijing, China
7 shared publications - Wentao Liu
School of Transportation Science and Engineering, Beihang University, Beijing, China
5 shared publications - Junjie Zhang
Beihang University
2 shared publications
Latest Publications
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 27 (10) , 12911-12924
Rail3DNet: A Monocular Vision-Based 3-D Rail-Track Detection Network Using Rail-IoU Loss
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 27 (10) , 12911-12924
Rail3DNet: A Monocular Vision-Based 3-D Rail-Track Detection Network Using Rail-IoU Loss
- 11 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. PP (15513203) , 1-12
3DRailNet: A Multifocal Cameras Fusion Network for Long-Range 3-D Rail-Track Detection
Accurate 3-D rail-track detection is vital to the perception of the railway environments for autonomous trains. However, existing methods based on monocular image cannot capture 3-D spatial features, facing challenges in detecting 3-D rail-track in turnouts and distant scenarios. This study introduces 3DRailNet, a long-range 3-D rail-track detection network using multifocal cameras. 3DRailNet consists of two modules: disparity-based feature extraction (DFE) module and long-short rail-track detection (LSRD) module. Specifically, the DFE module utilizes multifocal images to generate a disparity image and depth image, acquiring 3-D depth features to enhance the spatial information of rail-track. Based on the 3-D depth features, the LSRD module designs a detection head for long and short focal cameras to predict the 3-D position of rail-track. Experimental results demonstrate that the mean F1 score (mF1) of our proposed 3DRailNet is 83.2%, establishing it as the state-of-the-art method in this field. All these results indicate that 3DRailNet has the potential to be readily applicable in 3-D rail-track detection in railway environments.
- 11 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. PP (15513203) , 1-12
3DRailNet: A Multifocal Cameras Fusion Network for Long-Range 3-D Rail-Track Detection
Accurate 3-D rail-track detection is vital to the perception of the railway environments for autonomous trains. However, existing methods based on monocular image cannot capture 3-D spatial features, facing challenges in detecting 3-D rail-track in turnouts and distant scenarios. This study introduces 3DRailNet, a long-range 3-D rail-track detection network using multifocal cameras. 3DRailNet consists of two modules: disparity-based feature extraction (DFE) module and long-short rail-track detection (LSRD) module. Specifically, the DFE module utilizes multifocal images to generate a disparity image and depth image, acquiring 3-D depth features to enhance the spatial information of rail-track. Based on the 3-D depth features, the LSRD module designs a detection head for long and short focal cameras to predict the 3-D position of rail-track. Experimental results demonstrate that the mean F1 score (mF1) of our proposed 3DRailNet is 83.2%, establishing it as the state-of-the-art method in this field. All these results indicate that 3DRailNet has the potential to be readily applicable in 3-D rail-track detection in railway environments.
- 30 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 1-15
UnsOcc: An Occupancy Grid-Based General Detection Network for Unstructured Scenes
Perception and understanding of dynamic situations is a critical task for autonomous driving, particularly in complex, unstructured environments such as open-pit mining sites, where targets exhibit a vast diversity of shapes and sizes. This article introduces UnsOcc, a novel occupancy grid network specifically designed to handle such challenging scenes. The method utilizes camera images and LiDAR point clouds as inputs. A Dimension Lifting Module (DLM) first unifies these multi-modal data into a consistent voxel space through depth uncertainty estimation and cross-attention mechanisms. Subsequently, a Semantic Assignment Module (SAM) integrates contextual and semantic features from the images with the voxel queries, effectively aligning 3D occupancy states with their semantic information. Furthermore, a Local Super-Resolution (LSR) module is designed to enhance the resolution of highly-attended voxels, while processing less critical areas more efficiently, thus striking an optimal balance between computational cost and detection accuracy. Validated on an unstructured site dataset, UnsOcc achieves a state-of-the-art performance of 61.64% mIoU, representing a significant 6.34% improvement over baseline methods, thereby providing an effective solution for detecting and categorizing diverse objects in unstructured sites. Additionally, the network has been engineered for deployment, effectively enhancing the general obstacle detection capability of mining trucks. The data used in this article are available at http://www.sav-lab.com/.
- 30 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 1-15
UnsOcc: An Occupancy Grid-Based General Detection Network for Unstructured Scenes
Perception and understanding of dynamic situations is a critical task for autonomous driving, particularly in complex, unstructured environments such as open-pit mining sites, where targets exhibit a vast diversity of shapes and sizes. This article introduces UnsOcc, a novel occupancy grid network specifically designed to handle such challenging scenes. The method utilizes camera images and LiDAR point clouds as inputs. A Dimension Lifting Module (DLM) first unifies these multi-modal data into a consistent voxel space through depth uncertainty estimation and cross-attention mechanisms. Subsequently, a Semantic Assignment Module (SAM) integrates contextual and semantic features from the images with the voxel queries, effectively aligning 3D occupancy states with their semantic information. Furthermore, a Local Super-Resolution (LSR) module is designed to enhance the resolution of highly-attended voxels, while processing less critical areas more efficiently, thus striking an optimal balance between computational cost and detection accuracy. Validated on an unstructured site dataset, UnsOcc achieves a state-of-the-art performance of 61.64% mIoU, representing a significant 6.34% improvement over baseline methods, thereby providing an effective solution for detecting and categorizing diverse objects in unstructured sites. Additionally, the network has been engineered for deployment, effectively enhancing the general obstacle detection capability of mining trucks. The data used in this article are available at http://www.sav-lab.com/.
- 19 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-12
RailBEV: A Vision-Based Bird’s-Eye-View Representation Network for 3-D Railway Object Detection
Three-dimensional railway object detection is vital to the environmental perception of intelligent railway transportation systems. However, owing to the complexity of railway environments, critical issues such as the detection of multiscale and distant objects remain to be overcome. This study proposes a vision-based architecture called RailBEV, which introduces bird’s-eye-view (BEV) representation within 3-D railway object detection. RailBEV is mainly divided into two parallel parts: the depth supervision view transformation (DSVT) and long-range supervision (LRS) modules. Specifically, the DSVT module designed a novel depth supervision method to transform the front view (FV) images into the BEV plane. The LRS module leverages an additional 3-D detection head to guide a 2-D image backbone to learn extra 3-D object information and enhance the capability of long-range object detection. Experimental results demonstrate that the proposed RailBEV exhibits a mean average precision (mAP) of 38.6%, which is 4.4% higher than the baseline 3-D object detection network BEVDepth. Moreover, the results show that RailBEV effectively supports long-range detection and identifies multiscale and short-range objects. All these indicate that RailBEV has the potential to be readily applicable in 3-D railway object detection.
- 19 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-12
RailBEV: A Vision-Based Bird’s-Eye-View Representation Network for 3-D Railway Object Detection
Three-dimensional railway object detection is vital to the environmental perception of intelligent railway transportation systems. However, owing to the complexity of railway environments, critical issues such as the detection of multiscale and distant objects remain to be overcome. This study proposes a vision-based architecture called RailBEV, which introduces bird’s-eye-view (BEV) representation within 3-D railway object detection. RailBEV is mainly divided into two parallel parts: the depth supervision view transformation (DSVT) and long-range supervision (LRS) modules. Specifically, the DSVT module designed a novel depth supervision method to transform the front view (FV) images into the BEV plane. The LRS module leverages an additional 3-D detection head to guide a 2-D image backbone to learn extra 3-D object information and enhance the capability of long-range object detection. Experimental results demonstrate that the proposed RailBEV exhibits a mean average precision (mAP) of 38.6%, which is 4.4% higher than the baseline 3-D object detection network BEVDepth. Moreover, the results show that RailBEV effectively supports long-range detection and identifies multiscale and short-range objects. All these indicate that RailBEV has the potential to be readily applicable in 3-D railway object detection.
- 19 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 25 (7) , 12361-12371
CsoOcc: An Occupancy Network for Train Forward Perception Based on Cascade Sparsity Octree
The ability to perceive and reconstruct the surrounding environment in three dimensions is crucial for autonomous driving trains. With the advent of occupancy networks, it has become essential for autonomous train driving to use these networks for three-dimensional (3D) reconstruction in railway transportation. This study proposes the first railway occupancy network (CsoOcc) that considers the characteristics of railway transportation. Specifically, a cascaded optimization framework was devised to cascade the occupancy results of adjacent layers, thereby facilitating the incremental refinement of the occupancy features. A binocular attention mechanism was designed to establish correlations between binocular cameras, thereby enabling 3D reconstruction based on binocular correlations. A sparse octree optimization method was proposed to reduce the computational load on the network. Experiments based on the TWL-Occ dataset verified the CsoOcc performance, with a mean Intersection over Union of 0.632 and an Intersection over Union of 0.765, outperforming the benchmarks. The field experiments were conducted along the Hong Kong Metro Tsuen Wan Line. All the results demonstrate that the proposed occupancy network provides accurate forward 3D environmental information, thereby enhancing the safety of autonomous train navigation.
- 19 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 25 (7) , 12361-12371
CsoOcc: An Occupancy Network for Train Forward Perception Based on Cascade Sparsity Octree
The ability to perceive and reconstruct the surrounding environment in three dimensions is crucial for autonomous driving trains. With the advent of occupancy networks, it has become essential for autonomous train driving to use these networks for three-dimensional (3D) reconstruction in railway transportation. This study proposes the first railway occupancy network (CsoOcc) that considers the characteristics of railway transportation. Specifically, a cascaded optimization framework was devised to cascade the occupancy results of adjacent layers, thereby facilitating the incremental refinement of the occupancy features. A binocular attention mechanism was designed to establish correlations between binocular cameras, thereby enabling 3D reconstruction based on binocular correlations. A sparse octree optimization method was proposed to reduce the computational load on the network. Experiments based on the TWL-Occ dataset verified the CsoOcc performance, with a mean Intersection over Union of 0.632 and an Intersection over Union of 0.765, outperforming the benchmarks. The field experiments were conducted along the Hong Kong Metro Tsuen Wan Line. All the results demonstrate that the proposed occupancy network provides accurate forward 3D environmental information, thereby enhancing the safety of autonomous train navigation.
- 7 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 25 (5) , 4077-4088
Key Point Estimate Network for Rail-Track Detection
Rail-track detection is a crucial function for an active obstacle avoidance system in trains. However, existing methods face challenges in effectively detecting rail-tracks, particularly in turnout scenarios. This study introduces a novel rail-track detection approach using a key-point estimate network. The network treats the rail-track as a pair and constructs a dedicated model for detection. Additionally, a pseudo-attention mechanism leverages the detection output from previous stages, enabling the network to focus on the rail-track region. Also, a dislocation assignment mechanism is proposed to address label assignment confusion at turnouts. Moreover, a rail-track generalized IoU is also introduced, treating the rail-track as a pair and adds a correction term to enhance detection performance. Experimental results demonstrate that the proposed method achieves a remarkable mF1 score of 69.42%, establishing it as the state-of-the-art (SOTA) in this field. Furthermore, the effectiveness of the proposed method has been validated and applied in real-world testing on the Hong Kong Metro Tsuen Wan Line.
- 7 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 25 (5) , 4077-4088
Key Point Estimate Network for Rail-Track Detection
Rail-track detection is a crucial function for an active obstacle avoidance system in trains. However, existing methods face challenges in effectively detecting rail-tracks, particularly in turnout scenarios. This study introduces a novel rail-track detection approach using a key-point estimate network. The network treats the rail-track as a pair and constructs a dedicated model for detection. Additionally, a pseudo-attention mechanism leverages the detection output from previous stages, enabling the network to focus on the rail-track region. Also, a dislocation assignment mechanism is proposed to address label assignment confusion at turnouts. Moreover, a rail-track generalized IoU is also introduced, treating the rail-track as a pair and adds a correction term to enhance detection performance. Experimental results demonstrate that the proposed method achieves a remarkable mF1 score of 69.42%, establishing it as the state-of-the-art (SOTA) in this field. Furthermore, the effectiveness of the proposed method has been validated and applied in real-world testing on the Hong Kong Metro Tsuen Wan Line.
- 3 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 24 (12) , 19661-19672
MRCNet: Multiresolution LiDAR-Camera Calibration Using Optical Center Distance Loss Network
The integration of cameras and LiDAR sensors has emerged as a promising approach to enhance environmental perception and 3D reconstruction capabilities in autonomous vehicles and robotic systems. Precise extrinsic calibration is of paramount importance to achieve effective multi-sensor fusion applications. Traditional calibration methods often rely on manual procedures and specific calibration targets, which can be time-consuming and prone to errors. In contrast, Convolutional Neural Networks (CNNs) have shown potential in devising end-to-end calibration systems, leveraging their ability to extract robust features automatically. In this paper, MRCNet, an online end-to-end LiDAR-camera calibration network, which overcomes the limitations of traditional methods and previous CNN-based approaches, is proposed. This article introduces a multi-resolution feature extraction module, enabling the extraction of comprehensive and informative features from RGB images and depth images derived from point clouds. Additionally, the optical center distance loss, a novel concept that accounts for the camera’s optical imaging characteristics, facilitating more effective feature extraction is incorporated. MRCNet is the first online calibration network that considers the influence of camera imaging properties. This paper employs an iterative refinement process to progressively estimate the calibration error, allowing online extrinsic estimation. Evaluation tests on the KITTI Odometry dataset demonstrate the superior performance of MRCNet compared to existing learning-based methods, achieving a mean absolute calibration error of 0.350cm in translation and 0.033° in rotation. Furthermore, ablation studies validate the effectiveness of the modules of MRCNet. The code for MRCNet will be made publicly available at: https://github.com/AlexWang0214/MRCNet.
- 3 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 24 (12) , 19661-19672
MRCNet: Multiresolution LiDAR-Camera Calibration Using Optical Center Distance Loss Network
The integration of cameras and LiDAR sensors has emerged as a promising approach to enhance environmental perception and 3D reconstruction capabilities in autonomous vehicles and robotic systems. Precise extrinsic calibration is of paramount importance to achieve effective multi-sensor fusion applications. Traditional calibration methods often rely on manual procedures and specific calibration targets, which can be time-consuming and prone to errors. In contrast, Convolutional Neural Networks (CNNs) have shown potential in devising end-to-end calibration systems, leveraging their ability to extract robust features automatically. In this paper, MRCNet, an online end-to-end LiDAR-camera calibration network, which overcomes the limitations of traditional methods and previous CNN-based approaches, is proposed. This article introduces a multi-resolution feature extraction module, enabling the extraction of comprehensive and informative features from RGB images and depth images derived from point clouds. Additionally, the optical center distance loss, a novel concept that accounts for the camera’s optical imaging characteristics, facilitating more effective feature extraction is incorporated. MRCNet is the first online calibration network that considers the influence of camera imaging properties. This paper employs an iterative refinement process to progressively estimate the calibration error, allowing online extrinsic estimation. Evaluation tests on the KITTI Odometry dataset demonstrate the superior performance of MRCNet compared to existing learning-based methods, achieving a mean absolute calibration error of 0.350cm in translation and 0.033° in rotation. Furthermore, ablation studies validate the effectiveness of the modules of MRCNet. The code for MRCNet will be made publicly available at: https://github.com/AlexWang0214/MRCNet.