Yang Liu
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
Yang Liu has a total of 90 co-authors. Here are the top co-authors by number of shared publications:
- Xiaobo Qu
Tsinghua University
12 shared publications - Hongyi Lin
School of Vehicle and Mobility, Tsinghua University, Beijing, China
5 shared publications - Liang Wang
School of Vehicle and Mobility, Tsinghua University, Beijing, China
3 shared publications - Yixu He
School of Information Engineering, Changan University, Xian, China
3 shared publications - Kun Gao
Chalmers University of Technology
2 shared publications - Lan Yang
School of Information Engineering, Changan University, Xian, China
2 shared publications
Latest Publications
- 13 August 2026
- Published by Elsevier in Accident Analysis & Prevention
- Vol. 237
Operational efficiency of China’s low-altitude economy: Spatiotemporal patterns, policy drivers, and safety implications
- 13 August 2026
- Published by Elsevier in Accident Analysis & Prevention
- Vol. 237
Operational efficiency of China’s low-altitude economy: Spatiotemporal patterns, policy drivers, and safety implications
- 3 August 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 333
ReVerSeg: Training-free reason-and-verify video segmentation for traffic scenes
- 3 August 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 333
ReVerSeg: Training-free reason-and-verify video segmentation for traffic scenes
- 15 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. PP (15249050) , 1-15
U3A: Uncertainty-Aware Accident Anticipation for Reliable Autonomous Driving
Accident anticipation is essential for autonomous driving, yet existing video-based models often exhibit unstable predictions because they insufficiently account for the uncertainty that accumulates over long temporal sequences. Such instability leads to fluctuating confidence and unreliable alert timing in complex traffic environments. We propose U3A, an uncertainty-aware framework that improves the reliability of accident anticipation by jointly stabilizing representations and regulating decision behavior. U3A introduces a dual diffusion refinement module that enforces temporal coherence at both image and object levels, reducing abrupt feature shifts caused by occlusion and scene variability. To quantify epistemic uncertainty, the framework applies Monte Carlo Dropout over historical states and incorporates these estimates into a time-sensitive actor–critic policy, which learns to delay alerts under unstable confidence and issue warnings only when risk is consistently supported. Experiments on DAD, CCD, A3D, and AoTA show that U3A achieves state-of-the-art accuracy and mean time-to-accident, while producing smoother and more dependable alert trajectories. The results demonstrate that integrating diffusion-based feature refinement with uncertainty-guided policy learning offers a practical path toward more reliable accident anticipation for autonomous vehicles.
- 15 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. PP (15249050) , 1-15
U3A: Uncertainty-Aware Accident Anticipation for Reliable Autonomous Driving
Accident anticipation is essential for autonomous driving, yet existing video-based models often exhibit unstable predictions because they insufficiently account for the uncertainty that accumulates over long temporal sequences. Such instability leads to fluctuating confidence and unreliable alert timing in complex traffic environments. We propose U3A, an uncertainty-aware framework that improves the reliability of accident anticipation by jointly stabilizing representations and regulating decision behavior. U3A introduces a dual diffusion refinement module that enforces temporal coherence at both image and object levels, reducing abrupt feature shifts caused by occlusion and scene variability. To quantify epistemic uncertainty, the framework applies Monte Carlo Dropout over historical states and incorporates these estimates into a time-sensitive actor–critic policy, which learns to delay alerts under unstable confidence and issue warnings only when risk is consistently supported. Experiments on DAD, CCD, A3D, and AoTA show that U3A achieves state-of-the-art accuracy and mean time-to-accident, while producing smoother and more dependable alert trajectories. The results demonstrate that integrating diffusion-based feature refinement with uncertainty-guided policy learning offers a practical path toward more reliable accident anticipation for autonomous vehicles.
- Open Access
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1229-1239
A 2 GBD: Attack-Agnostic Graph Backdoor Defense
Graph Neural Networks (GNNs) are vulnerable to graph backdoor attacks, which poses severe risks to their deployment in safety-critical applications. Existing defenses predominantly focus on specific backdoor triggers, making them brittle and unable to generalize across different backdoor triggers with varying properties. Motivated by this limitation, this work proposes an attack-agnostic graph backdoor defense mechanism A2GBD, which does not require prior knowledge of the specific attack strategies (e.g., edge perturbation, node attribute manipulation) to achieve effective defense. A2GBD consists of suspicious node selection and defense strategy generation. The selection module selects high-suspicion nodes to enhance defense awareness, while the defense agent adaptively determines and executes defense strategies. Extensive experiments on multiple benchmark datasets demonstrate that A2GBD consistently lowers attack success rates while maintaining high clean accuracy, showing strong robustness and generalizability against diverse graph backdoor attack strategies.
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1229-1239
Open AccessA 2 GBD: Attack-Agnostic Graph Backdoor Defense
Graph Neural Networks (GNNs) are vulnerable to graph backdoor attacks, which poses severe risks to their deployment in safety-critical applications. Existing defenses predominantly focus on specific backdoor triggers, making them brittle and unable to generalize across different backdoor triggers with varying properties. Motivated by this limitation, this work proposes an attack-agnostic graph backdoor defense mechanism A2GBD, which does not require prior knowledge of the specific attack strategies (e.g., edge perturbation, node attribute manipulation) to achieve effective defense. A2GBD consists of suspicious node selection and defense strategy generation. The selection module selects high-suspicion nodes to enhance defense awareness, while the defense agent adaptively determines and executes defense strategies. Extensive experiments on multiple benchmark datasets demonstrate that A2GBD consistently lowers attack success rates while maintaining high clean accuracy, showing strong robustness and generalizability against diverse graph backdoor attack strategies.
- 19 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. PP (15249050) , 1-13
Pretrained Encoder-Based Large-Scale Dynamic Routing for Low-Carbon Electric Demand Responsive Transit
Dynamic routing for large-scale electric Demand Responsive Transit (e-DRT) faces significant challenges due to the stochastic arrival of requests, real-time decision-making requirements, computational complexity, and battery-related constrains. Recent pre-trained large models have emerged as a promising approach for leveraging transferable knowledge to improve optimization efficiency. Thus, this study proposes a pre-trained encoder–based approach to improve both optimization efficiency and generalization capability in dynamic routing tasks. The framework integrates a pre-trained encoder that learns latent representations of requests and vehicles within a shared embedding space, enabling similarity guidance for both clustering and heuristic optimization. The proposed request–vehicle clustering reduces the search space, while the enhanced ALNS performs route optimization within each cluster, subsequently supported by global refinement to correct cross-cluster inconsistencies. In cross-scenario large-scale cases of e-DRT, experimental results show that the pre-trained encoder-based framework achieves higher completion rates, shorter waiting and delay times, and significantly improves objective values compared to baseline algorithms, while maintaining real-time computational efficiency. Ablation studies further verify the contributions of the pre-trained encoder in enhancing solution quality and strong generalization. The proposed framework offers a new perspective on integrating pre-trained representation learning with existing algorithms to address large-scale and real-time e-DRT in intelligent transportation systems.
- 19 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. PP (15249050) , 1-13
Pretrained Encoder-Based Large-Scale Dynamic Routing for Low-Carbon Electric Demand Responsive Transit
Dynamic routing for large-scale electric Demand Responsive Transit (e-DRT) faces significant challenges due to the stochastic arrival of requests, real-time decision-making requirements, computational complexity, and battery-related constrains. Recent pre-trained large models have emerged as a promising approach for leveraging transferable knowledge to improve optimization efficiency. Thus, this study proposes a pre-trained encoder–based approach to improve both optimization efficiency and generalization capability in dynamic routing tasks. The framework integrates a pre-trained encoder that learns latent representations of requests and vehicles within a shared embedding space, enabling similarity guidance for both clustering and heuristic optimization. The proposed request–vehicle clustering reduces the search space, while the enhanced ALNS performs route optimization within each cluster, subsequently supported by global refinement to correct cross-cluster inconsistencies. In cross-scenario large-scale cases of e-DRT, experimental results show that the pre-trained encoder-based framework achieves higher completion rates, shorter waiting and delay times, and significantly improves objective values compared to baseline algorithms, while maintaining real-time computational efficiency. Ablation studies further verify the contributions of the pre-trained encoder in enhancing solution quality and strong generalization. The proposed framework offers a new perspective on integrating pre-trained representation learning with existing algorithms to address large-scale and real-time e-DRT in intelligent transportation systems.
- 16 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 26 (9) , 12819-12833
A Survey and Comprehensive Taxonomy of Tire-Road Adhesion Coefficient Estimation for Intelligent Vehicles
Within autonomous driving research, the intricate variability of the road surface is frequently overlooked, while the tire-road interactions critically impact vehicle stability. This paper comprehensively reviews traditional and emerging tire-road adhesion coefficient (TRAC) estimation methods for intelligent vehicles. We initially categorize traditional methods into cause-based and effect-based approaches, which are founded on vehicle responses and road surface characteristics, respectively. Then, we classify emerging methods into learning-based approaches and hybrid models combining physical principles with data-driven strategies. We eventually point out areas for improvement and future research directions. The proposed systematic taxonomy summarizes the independent and collaborative operations of dynamics analysis and learning methods in TRAC estimation, offering insights for further research.
- 16 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 26 (9) , 12819-12833
A Survey and Comprehensive Taxonomy of Tire-Road Adhesion Coefficient Estimation for Intelligent Vehicles
Within autonomous driving research, the intricate variability of the road surface is frequently overlooked, while the tire-road interactions critically impact vehicle stability. This paper comprehensively reviews traditional and emerging tire-road adhesion coefficient (TRAC) estimation methods for intelligent vehicles. We initially categorize traditional methods into cause-based and effect-based approaches, which are founded on vehicle responses and road surface characteristics, respectively. Then, we classify emerging methods into learning-based approaches and hybrid models combining physical principles with data-driven strategies. We eventually point out areas for improvement and future research directions. The proposed systematic taxonomy summarizes the independent and collaborative operations of dynamics analysis and learning methods in TRAC estimation, offering insights for further research.
- 20 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 74 (5) , 7404-7415
A High-Precision Calibration and Evaluation Method Based on Binocular Cameras and LiDAR for Intelligent Vehicles
Accurate positioning is crucial for intelligent vehicles, especially in scenarios with spatial constraints such as close vehicle proximities, tight parking spaces, and the docking process of autonomous modular buses (AMBs). Binocular cameras and Light Detection and Ranging (LiDAR) have shown great potential in intelligent vehicle applications. However, existing methods mostly rely on comparing the inter-camera extrinsic matrices and the results of the calibration between a single camera and LiDAR. This not only leads to the accumulation of errors in each process but also fails to accurately determine the source of errors when calibration results are suboptimal. To overcome these problems, this paper proposes a high-precision, phased joint calibration method based on binocular cameras and LiDAR, along with a combined global and local evaluation approach, and introduces a visualization scheme to enhance the reliability and intuitiveness of the joint calibration process. Experimental results on AMBs demonstrate that our methodology and selection of intrinsic and extrinsic parameters significantly improve performance compared to other mainstream methods.
- 20 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 74 (5) , 7404-7415
A High-Precision Calibration and Evaluation Method Based on Binocular Cameras and LiDAR for Intelligent Vehicles
Accurate positioning is crucial for intelligent vehicles, especially in scenarios with spatial constraints such as close vehicle proximities, tight parking spaces, and the docking process of autonomous modular buses (AMBs). Binocular cameras and Light Detection and Ranging (LiDAR) have shown great potential in intelligent vehicle applications. However, existing methods mostly rely on comparing the inter-camera extrinsic matrices and the results of the calibration between a single camera and LiDAR. This not only leads to the accumulation of errors in each process but also fails to accurately determine the source of errors when calibration results are suboptimal. To overcome these problems, this paper proposes a high-precision, phased joint calibration method based on binocular cameras and LiDAR, along with a combined global and local evaluation approach, and introduces a visualization scheme to enhance the reliability and intuitiveness of the joint calibration process. Experimental results on AMBs demonstrate that our methodology and selection of intrinsic and extrinsic parameters significantly improve performance compared to other mainstream methods.
- 8 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Big Data
- Vol. 11 (3) , 1568-1587
Big Data-Driven Advancements and Future Directions in Vehicle Perception Technologies: From Autonomous Driving to Modular Buses
The rapid development of big data and artificial intelligence (AI) is revolutionizing the automotive and transportation industries, leading to the creation of the Autonomous Modular Bus (AMB). Designed to address the key challenges of modern public transportation systems, the AMB adopts a modular dynamic assembly approach. However, existing research on the AMB predominantly focuses on operational aspects, whereas in-transit docking remains the primary obstacle to its commercial deployment. This challenge stems from the fact that current perception accuracy in autonomous vehicles is limited to the decimeter level, with insufficient capability to manage adverse weather and complex traffic conditions. To enable AMBs to achieve full-scenario autonomous driving capabilities, this paper reviews current perception technologies from three perspectives: single-vehicle single-sensor perception, multi-sensor fusion perception, and cooperative perception. It examines the characteristics of existing perception solutions and evaluates their applicability to AMB-specific requirements. Furthermore, considering the unique challenges of in-transit docking, this paper identifies and proposes four future research directions for advancing AMB perception systems as well as general autonomous driving technologies.
- 8 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Big Data
- Vol. 11 (3) , 1568-1587
Big Data-Driven Advancements and Future Directions in Vehicle Perception Technologies: From Autonomous Driving to Modular Buses
The rapid development of big data and artificial intelligence (AI) is revolutionizing the automotive and transportation industries, leading to the creation of the Autonomous Modular Bus (AMB). Designed to address the key challenges of modern public transportation systems, the AMB adopts a modular dynamic assembly approach. However, existing research on the AMB predominantly focuses on operational aspects, whereas in-transit docking remains the primary obstacle to its commercial deployment. This challenge stems from the fact that current perception accuracy in autonomous vehicles is limited to the decimeter level, with insufficient capability to manage adverse weather and complex traffic conditions. To enable AMBs to achieve full-scenario autonomous driving capabilities, this paper reviews current perception technologies from three perspectives: single-vehicle single-sensor perception, multi-sensor fusion perception, and cooperative perception. It examines the characteristics of existing perception solutions and evaluates their applicability to AMB-specific requirements. Furthermore, considering the unique challenges of in-transit docking, this paper identifies and proposes four future research directions for advancing AMB perception systems as well as general autonomous driving technologies.