Yang Songyue, an Outstanding Hundred Talents Postdoctoral Fellow at Beihang University, has long been engaged in research on intelligent perception in rail transit. He has led projects such as the National Postdoctoral Researcher Support Program (Grade B), the National Natural Science Foundation Youth Fund, and the Postdoctoral Science Foundation General Program. He has published over 10 SCI/EI-indexed papers, including five as the first author in top journals such as IEEE TITS, IEEE TVT, and IEEE TIM. He has also filed or been granted over 10 invention patents. He won third prize in the China Intelligent Transportation Innovation Challenge—Track Line Scene Semantic Segmentation Group and earned the runner-up title and the Innovative Solution Award in the Robust Depth Estimation Group at the RoboDrive Challenge, held during the top robotics conference ICRA. The technologies he developed have been deployed on 29 lines across 10 cities, including Beijing Subway Line 17 and Hong Kong's Tsuen Wan Line, with a cumulative operational mileage exceeding 10 million kilometers.

In response to the demand for ultra-long-distance, highly reliable foreign object detection in rail transit scenarios, his research focuses on foreign object detection within the forward clearance space of urban rail trains. Following the approach of "data enhancement–clearance modeling–object detection," he concentrates on the following three research areas:

(1) A depth supervision method based on a pooling guidance module was proposed. Guided by the assumption of depth consistency, a ground truth generation module based on pooling guidance was designed, and a loss function based on the guided depth was developed using the generated ground truth. This addresses the issue of significant depth noise in existing methods when completing depth maps from non-repetitive laser point clouds.

(2) A detection head structure based on feature reconstruction was designed. A 2D-3D implicit spatial association method based on iterative mapping was constructed, and a feature reconstruction module based on multi-scale neighborhood aggregation was designed. Additionally, based on the assumption of a fixed distance between the two rails, a dual-rail coordinate regression method incorporating information exchange was proposed, addressing the problem of missing 3D information in scenarios with slope variations.

(3) A progressive optimization strategy based on sparse cascade octrees was designed. A multi-scale feature fusion method based on cascade optimization was proposed, along with an octree-based occupancy optimization method and a dense pruning mechanism based on sparse convolution. This addresses the issue of high computational cost in occupancy grid networks.

Representative Publications：

1. Songyue Yang, Zhangyu Wang, Guizhen Yu, Bin Zhou, Peng Chen, Sifen Wang and Qiang Zhang. RailDepth: A Self-Supervised Network for Railway Depth Completion Based on a Pooling-Guidance Mechanism [J]. IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1-13, 2023, doi: 10.1109/TIM.2023.3284958. (SCI, IF: 5.6, JCR Q1).

2. Songyue Yang, Guizhen Yu, Zhangyu Wang, Bin Zhou, Peng Chen and Qiang Zhang. A Topology Guided Method for Rail-Track Detection [J]. IEEE Transactions on Vehicular Technology, vol. 71, no. 2, pp. 1426-1438, 2022, doi: 10.1109/TVT.2021.3133327. (SCI, IF: 6.8, JCR Q1).

3. Songyue Yang, Zhangyu Wang, Guizhen Yu and Wentao Liu. Key Point Estimate Network for Rail-Track Detection [J]. IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 5, pp. 4077-4088, 2023, doi: 10.1109/TITS.2023.3327996. (SCI, IF: 8.5, JCR Q1).

4. Songyue Yang, Bin Zhou, Zhangyu Wang, Hongbo Li, Wentao Liu, Guizhen Yu.  CsoOcc: An Occupancy Network for Train Forward Perception Based on Cascade Sparsity Octree [J]. IEEE Sensors Journal, vol. 25, no. 7, pp. 12361-12371, 1 April1, 2025, doi: 10.1109/JSEN.2025.3539626. (SCI, IF: 4.5, JCR Q1).

5. Songyue Yang, Guizhen Yu, Zhijun Meng, Zhangyu Wang, and Han Li. Autonomous obstacle avoidance of UAV based on deep reinforcement learning [J]. Journal of Intelligent & Fuzzy Systems, vol. 42, no.4 , pp. 3323–3335, 2022, doi: 10.3233/JIFS-211192. (SCI, IF: 2.0, JCR Q3).

Patents：

1. Yu Guizhen, Yang Songyue, Liu Wentao, Wang Zhangyu, Zhou Bin. A Track Line Fitting Method in Complex Scenes Based on Image Segmentation [P]. China: CN202110066159.3 (as the first student author, granted).

2. Zhou Bin, Yang Songyue, Wang Zhangyu, Yu Guizhen, Liu Wentao. A Precise Detection Method for Train Platform Stopping Points Based on Lidar and Vision Fusion [P]. China: CN202110939927.1 (as the first student author, granted).

3. Wang Zhangyu, Yang Songyue, Zhang Haojie, Li Hongbo, Yu Guizhen. A Train Positioning Correction Method and Device Based on High-Precision Map Trajectory Priors [P]. China: CN202311825611.5 (as the first student author, granted).

Grants as Principal Investigator：

1. Research on Vision-Based Ultra-Long-Range 3D Generic Obstacle Detection Method for Trains — National Postdoctoral Researcher Support Program (Grade B), 2025.07–2027.07.

2. Research on Intrusion Detection Method for 3D Clearance Obstacles in Urban Rail Trains Based on Super-Resolution Occupancy Grids — National Natural Science Foundation of China (Youth Program), 2026.01–2028.12.

3. Research on Vision-Based Intrusion Detection Method for Obstacles Within Train Clearance Space Under Adverse Weather Conditions — China Postdoctoral Science Foundation General Program, 2025.07–2027.07.