Wenjing Zhao
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
Wenjing Zhao has a total of 82 co-authors. Here are the top co-authors by number of shared publications:
- Dezong Zhao
University of Glasgow, Glasgow, United Kingdom
13 shared publications - David Flynn
James Watt School of Engineering, University of Glasgow, Glasgow, UK
8 shared publications - Daxin Tian
Beijing Advanced Innovation Center for Big Data and Brain Computing, Beijing Key Laboratory for Cooperative Vehicle Infrastructure Systems and Safety Control, School of Transportation Science and Engineering, Beihang University, Beijing, China
4 shared publications - Zhen Tian
James Watt School of Engineering, University of Glasgow, Glasgow G12 8QQ, UK
4 shared publications - Zhihao Lin
University of Glasgow
4 shared publications Fenglin Liu
School of Information Engineering, Chang'an University, Xi'an 710064, China
3 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) , 12687-12700
Dual-Mode SPL-SLAM: A Robust Semantic Point-Line SLAM System With Monocular Neural Depth and Sensor Fusion for Intelligent Transportation
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 27 (10) , 12687-12700
Dual-Mode SPL-SLAM: A Robust Semantic Point-Line SLAM System With Monocular Neural Depth and Sensor Fusion for Intelligent Transportation
- Open Access
- 25 September 2026
- Published by Elsevier in SSRN Electronic Journal
CWAI: Development and Initial Validation of a Microprompting Generative AI approach to performing Cognitive Work Analysis
Human Factors (Ergonomics, E/HF) is facing increasing calls for a step change in how analytical methods are conducted, with an emphasis on improving robustness and analytical power through the integration of modern data driven and programmatic techniques. The Cognitive Work Analysis (CWA), while of value to the analyst conducting the method, and with the potential for design and ergonomics insights, it is cumbersome to perform and largely confined to outdated tools or generic software. Artificial Intelligence (AI) offers a breakthrough. This paper describes the development and application of a software tool, called CWAI, which uses a novel approach to programmatic AI prompt design. We present an initial validation against an existing human-made CWA and a Microsoft Copilot analysis. Results show the CWAI generally outperforms the Copilot offering in our initial testing and provides a viable alternative to purely human analysis. The paper serves first step towards embracing AI as part of E/HF.
- 25 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCWAI: Development and Initial Validation of a Microprompting Generative AI approach to performing Cognitive Work Analysis
Human Factors (Ergonomics, E/HF) is facing increasing calls for a step change in how analytical methods are conducted, with an emphasis on improving robustness and analytical power through the integration of modern data driven and programmatic techniques. The Cognitive Work Analysis (CWA), while of value to the analyst conducting the method, and with the potential for design and ergonomics insights, it is cumbersome to perform and largely confined to outdated tools or generic software. Artificial Intelligence (AI) offers a breakthrough. This paper describes the development and application of a software tool, called CWAI, which uses a novel approach to programmatic AI prompt design. We present an initial validation against an existing human-made CWA and a Microsoft Copilot analysis. Results show the CWAI generally outperforms the Copilot offering in our initial testing and provides a viable alternative to purely human analysis. The paper serves first step towards embracing AI as part of E/HF.
- 21 July 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 351
DCGM: Synergizing Differential Cross-Modal Fusion and Graph-Mamba for Referring Multi-Object Tracking
- 21 July 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 351
DCGM: Synergizing Differential Cross-Modal Fusion and Graph-Mamba for Referring Multi-Object Tracking
- Open Access
- 14 May 2026
- Published by Elsevier in SSRN Electronic Journal
Towards a Strategic Human Factors Digital Twinning Framework for Transport Decarbonisation
The TransiT project is pioneering the use of national-scale digital twins for the purpose of rapid decarbonisation of the UK’s transport system. The transport system is designed to service humans, and therefore without adequate representation of human factors in the transport system, any digital twins are likely to be inaccurate. This paper describes a strand of research aiming to approach Human Factors digital twinning from a novel perspective. Instead of starting with computational digital twinning platforms, the approach begins the Abstraction Hierarchy (AH). The AH is then systematically connected to data sources through a novel cyber physical systems audit. This audit provides a bridge to a programmatic solution. The AH combined with the mapping of means-ends links to actual or potential cyber physical information sources. This work points towards a computationally parsimonious yet powerful and strategic Human Factors digital twinning approach. The work provides a framework to describe this approach.
- 14 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessTowards a Strategic Human Factors Digital Twinning Framework for Transport Decarbonisation
The TransiT project is pioneering the use of national-scale digital twins for the purpose of rapid decarbonisation of the UK’s transport system. The transport system is designed to service humans, and therefore without adequate representation of human factors in the transport system, any digital twins are likely to be inaccurate. This paper describes a strand of research aiming to approach Human Factors digital twinning from a novel perspective. Instead of starting with computational digital twinning platforms, the approach begins the Abstraction Hierarchy (AH). The AH is then systematically connected to data sources through a novel cyber physical systems audit. This audit provides a bridge to a programmatic solution. The AH combined with the mapping of means-ends links to actual or potential cyber physical information sources. This work points towards a computationally parsimonious yet powerful and strategic Human Factors digital twinning approach. The work provides a framework to describe this approach.
- Open Access
- 8 April 2026
- Published by Elsevier in SSRN Electronic Journal
An Explainable Digital Twin Framework for Dynamic Optimisation of EV Charging Station Coordination
Electric vehicle (EV) adoption is rapidly growing worldwide, driven by global decarbonisation goals that place growing pressure on charging infrastructure and transport systems. As EV charging demand is spatially distributed and temporally dynamic, allocating vehicles to optimal charging stations (CSs) is a complex, multi-objective decision-making problem shaped by mobility patterns, infrastructure availability, and user behaviour at scale. Existing research models EVs as homogeneous agents or only considers limited combinations of their dynamic factors, including queue time at charging infrastructure, charging cost, or the energy feasibility in guiding to CSs limiting their ability to support scalable and interpretable decision-making. We propose a digital twin (DT) framework that integrates real-world mobility patterns, heterogeneous user behaviour, and diverse CS characteristics into a Multi-Criteria Decision Analysis (MCDA) approach capable of real-time CS selection. The resulting decision-making solution guides EV charging choices based on travel time, queueing time, charging time, charging cost, and energy feasibility. To ensure scalability and support near real-time operation, the proposed DT relies on a machine learning model (LSTM) trained with data generated from simulations of different experimental setups to predict these CSs decisions. The Explainable Artificial Intelligence (XAI) method, SHapley Additive exPlanations (SHAP), is incorporated into the framework to provide real-time explanations, transparency, and interpretability for the prediction of charging decisions. A case study of Orkney, UK, demonstrates the ability of the proposed LSTM that achieved an average accuracy for real-time CS selection of 98.7\% compared to heavier digital models.
- 8 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessAn Explainable Digital Twin Framework for Dynamic Optimisation of EV Charging Station Coordination
Electric vehicle (EV) adoption is rapidly growing worldwide, driven by global decarbonisation goals that place growing pressure on charging infrastructure and transport systems. As EV charging demand is spatially distributed and temporally dynamic, allocating vehicles to optimal charging stations (CSs) is a complex, multi-objective decision-making problem shaped by mobility patterns, infrastructure availability, and user behaviour at scale. Existing research models EVs as homogeneous agents or only considers limited combinations of their dynamic factors, including queue time at charging infrastructure, charging cost, or the energy feasibility in guiding to CSs limiting their ability to support scalable and interpretable decision-making. We propose a digital twin (DT) framework that integrates real-world mobility patterns, heterogeneous user behaviour, and diverse CS characteristics into a Multi-Criteria Decision Analysis (MCDA) approach capable of real-time CS selection. The resulting decision-making solution guides EV charging choices based on travel time, queueing time, charging time, charging cost, and energy feasibility. To ensure scalability and support near real-time operation, the proposed DT relies on a machine learning model (LSTM) trained with data generated from simulations of different experimental setups to predict these CSs decisions. The Explainable Artificial Intelligence (XAI) method, SHapley Additive exPlanations (SHAP), is incorporated into the framework to provide real-time explanations, transparency, and interpretability for the prediction of charging decisions. A case study of Orkney, UK, demonstrates the ability of the proposed LSTM that achieved an average accuracy for real-time CS selection of 98.7\% compared to heavier digital models.
- 20 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (10) , 15456-15468
Pillar-SAN: A Pillar-Based Spatiotemporal Attention Network for 3-D Object Detection From 4-D Radar Point Cloud
3D object detection is a crucial component of the perception system in autonomous driving. With the advent of 4D millimeter-wave radar technology, 3D object detection with 4D radar introduces new challenges and opportunities. 4D radar is recognized as a state-of-the-art sensor owing to its superior all-weather performance and remarkable robustness in complex environmental conditions. However, compared to LiDAR, 4D radar point cloud is sparser and contains more noise. These characteristics hinder the effectiveness of existing LiDAR-based 3D object detection methods when applied to 4D radar, making them less suitable for accurate 3D object detection in this context. To cope with this issue, we propose a Pillar-based Spatiotemporal Attention Network (Pillar-SAN). The network builds a sequential perception attention module to establish intrinsic connections between point cloud across both time and space, thereby enhancing its capability to extract global and local features from sparse point cloud. In addition, the receptive field attention module is built to refine object feature details while effectively suppressing noise. Experimental results on the VOD dataset show our Pillar-SAN achieves significant improvements over previous state-of-the-art methods, with a 7.22pp increase in 3D mAP, a 7.8pp increase in BEV mAP, an 11.42pp increase in BBOX mAP, and a 12.94pp increase in AOS mAP.
- 20 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (10) , 15456-15468
Pillar-SAN: A Pillar-Based Spatiotemporal Attention Network for 3-D Object Detection From 4-D Radar Point Cloud
3D object detection is a crucial component of the perception system in autonomous driving. With the advent of 4D millimeter-wave radar technology, 3D object detection with 4D radar introduces new challenges and opportunities. 4D radar is recognized as a state-of-the-art sensor owing to its superior all-weather performance and remarkable robustness in complex environmental conditions. However, compared to LiDAR, 4D radar point cloud is sparser and contains more noise. These characteristics hinder the effectiveness of existing LiDAR-based 3D object detection methods when applied to 4D radar, making them less suitable for accurate 3D object detection in this context. To cope with this issue, we propose a Pillar-based Spatiotemporal Attention Network (Pillar-SAN). The network builds a sequential perception attention module to establish intrinsic connections between point cloud across both time and space, thereby enhancing its capability to extract global and local features from sparse point cloud. In addition, the receptive field attention module is built to refine object feature details while effectively suppressing noise. Experimental results on the VOD dataset show our Pillar-SAN achieves significant improvements over previous state-of-the-art methods, with a 7.22pp increase in 3D mAP, a 7.8pp increase in BEV mAP, an 11.42pp increase in BBOX mAP, and a 12.94pp increase in AOS mAP.
- 27 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 22 (5) , 3799-3810
Enhancing Data Efficiency With a Trustworthy Counterfactual Generative Model
Leveraging limited data to synthesize an additional training set is essential for robotic vision, particularly in dynamic environments where collecting large datasets is impractical. Traditional robotic vision systems rely on extensive training data for object recognition and scene understanding but struggle to generalize to real-world variations, such as lighting conditions, occlusions, and sensor noise. This article proposes causal diffuse variational autoencoder (causal DiffuseVAE), a novel method integrating causal inference with high-fidelity image synthesis to generate counterfactual images. By combining the disentanglement properties of variational autoencoders (VAEs) with the generative capabilities of diffusion models, causal DiffuseVAE produces realistic, interpretable simulations of variations, such as shadows and occlusions. This combination enables data-efficient generative modeling by learning from small subsets and synthesizing missing or unseen samples. In addition, causal inference ensures that generated data follow real-world dependencies, making it robust and interpretable for deployment in unpredictable environments. Four baseline approaches are evaluated across six different datasets, demonstrating that causal DiffuseVAE consistently outperforms the four baseline approaches.
- 27 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 22 (5) , 3799-3810
Enhancing Data Efficiency With a Trustworthy Counterfactual Generative Model
Leveraging limited data to synthesize an additional training set is essential for robotic vision, particularly in dynamic environments where collecting large datasets is impractical. Traditional robotic vision systems rely on extensive training data for object recognition and scene understanding but struggle to generalize to real-world variations, such as lighting conditions, occlusions, and sensor noise. This article proposes causal diffuse variational autoencoder (causal DiffuseVAE), a novel method integrating causal inference with high-fidelity image synthesis to generate counterfactual images. By combining the disentanglement properties of variational autoencoders (VAEs) with the generative capabilities of diffusion models, causal DiffuseVAE produces realistic, interpretable simulations of variations, such as shadows and occlusions. This combination enables data-efficient generative modeling by learning from small subsets and synthesizing missing or unseen samples. In addition, causal inference ensures that generated data follow real-world dependencies, making it robust and interpretable for deployment in unpredictable environments. Four baseline approaches are evaluated across six different datasets, demonstrating that causal DiffuseVAE consistently outperforms the four baseline approaches.
- 19 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 75 (1) , 84-99
Balanced Exploration and Attention-Inspired Decision Making for Autonomous Driving
Autonomous driving has attracted great interest due to its potential capability in enhancing safety and improving traffic efficiency. Both model-based and learning-based methods are widely used in autonomous driving. Out of which, model-based methods rely on existing events in the dataset but are poor in learning extended situations. As a comparison, the deep Q-learning network (DQN) has a strong capability in learning within interactive driving. However, existing DQN faces challenges in convergence in terms of speed and accuracy, especially in interactive environments. Furthermore, the poor convergence causes high risks of collisions and slow driving speed. Therefore, this paper presents a modified DQN to achieve a lower number of collisions, higher average driving speed, and faster convergence during interactive driving. The modified DQN is developed by introducing a risk-attention mechanism, a balanced reward function, and a collision-supervised mechanism (RBDQN-CS). The proposed risk-attention mechanism enhances the DQN to pay attention to high-frequent interactions. The proposed balanced reward function specifies the weight of the control strategy to handle the interactions with surrounding human driven vehicles. The collision-supervised mechanism detects the collision risks and prevents the collision occurrence during lane-changing. Simulation results demonstrate that RBDQN-CS outperforms DQN and other popular baseline DRL algorithms.
- 19 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 75 (1) , 84-99
Balanced Exploration and Attention-Inspired Decision Making for Autonomous Driving
Autonomous driving has attracted great interest due to its potential capability in enhancing safety and improving traffic efficiency. Both model-based and learning-based methods are widely used in autonomous driving. Out of which, model-based methods rely on existing events in the dataset but are poor in learning extended situations. As a comparison, the deep Q-learning network (DQN) has a strong capability in learning within interactive driving. However, existing DQN faces challenges in convergence in terms of speed and accuracy, especially in interactive environments. Furthermore, the poor convergence causes high risks of collisions and slow driving speed. Therefore, this paper presents a modified DQN to achieve a lower number of collisions, higher average driving speed, and faster convergence during interactive driving. The modified DQN is developed by introducing a risk-attention mechanism, a balanced reward function, and a collision-supervised mechanism (RBDQN-CS). The proposed risk-attention mechanism enhances the DQN to pay attention to high-frequent interactions. The proposed balanced reward function specifies the weight of the control strategy to handle the interactions with surrounding human driven vehicles. The collision-supervised mechanism detects the collision risks and prevents the collision occurrence during lane-changing. Simulation results demonstrate that RBDQN-CS outperforms DQN and other popular baseline DRL algorithms.
- Top Cited Papers
- 15 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 27 (2) , 1709-1730
Evaluating Scenario-Based Decision-Making for Interactive Autonomous Driving Using Rational Criteria: A Survey
Autonomous vehicles (AVs) promise substantial gains in safety, reliability, and decarbonization, yet safe and efficient interaction in dynamic, heterogeneous traffic remains a key barrier to large-scale deployment. Deep reinforcement learning (DRL) has emerged as a data-driven approach for learning adaptive decision policies that handle complex, unpredictable environments better than rule-based methods. However, different scenarios impose distinct requirements, necessitating scenario-specific algorithms. This survey systematically reviews DRL for four typical scenarios (highways, on-ramp merging, roundabouts, and unsignalized intersections), summarizes road features and recent advances, and evaluates methods using five criteria: driving safety, driving efficiency, training efficiency, unselfishness, and interpretability (DDTUI). Each DDTUI criterion is analyzed with respect to the reviewed algorithms. In addition, a dedicated scenario-centric learning transferability analysis is introduced that systematically evaluates whether each reviewed method demonstrates scene-specific learning improvements and assesses how effectively their designs transfer across the four scenarios. Finally, the challenges for future DRL-based decision-making algorithms are summarized.
- 15 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 27 (2) , 1709-1730
Top Cited PapersEvaluating Scenario-Based Decision-Making for Interactive Autonomous Driving Using Rational Criteria: A Survey
Autonomous vehicles (AVs) promise substantial gains in safety, reliability, and decarbonization, yet safe and efficient interaction in dynamic, heterogeneous traffic remains a key barrier to large-scale deployment. Deep reinforcement learning (DRL) has emerged as a data-driven approach for learning adaptive decision policies that handle complex, unpredictable environments better than rule-based methods. However, different scenarios impose distinct requirements, necessitating scenario-specific algorithms. This survey systematically reviews DRL for four typical scenarios (highways, on-ramp merging, roundabouts, and unsignalized intersections), summarizes road features and recent advances, and evaluates methods using five criteria: driving safety, driving efficiency, training efficiency, unselfishness, and interpretability (DDTUI). Each DDTUI criterion is analyzed with respect to the reviewed algorithms. In addition, a dedicated scenario-centric learning transferability analysis is introduced that systematically evaluates whether each reviewed method demonstrates scene-specific learning improvements and assesses how effectively their designs transfer across the four scenarios. Finally, the challenges for future DRL-based decision-making algorithms are summarized.