Wenfeng Guo
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Wenfeng Guo has a total of 39 co-authors. Here are the top co-authors by number of shared publications:
- Weiwei Zhang
Tongji University
5 shared publications - Menghao Li
School of mechanical and automotive engineering, Shanghai University of Engineering Science, Shanghai 201620, China
4 shared publications - Wangpengfei Yu
School of mechanical and automotive engineering, Shanghai University of Engineering Science, Shanghai 201620, China
4 shared publications - Jun Li
School of Vehicle and Mobility, Tsinghua University, Beijing, China
3 shared publications - Weiwei Zhang
Shanghai Smart Vehicle Cooperating Innovation Center Co., Ltd., Shanghai 201805, China
3 shared publications Xiaolin Song
State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body, Hunan University, Changsha, China
3 shared publications
Latest Publications
- 9 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 21 (15566013) , 2905-2917
RFFRDet: A Refined Feature Fusion Rotation Detector for Prohibited Item Recognition in X-Ray Images
With increasing security threats in critical infrastructure, intelligent X-ray inspection systems have become essential for modern security frameworks. Contraband detection faces dual challenges: semantic dilution across cross-scale features and degraded directional sensitivity. Existing detection paradigms rely on Euclidean spatial assumptions while ignoring X-ray projection geometry characteristics, leading to feature representation ambiguity and insufficient spatial relationship modeling in complex occlusion scenarios. To address these challenges, we propose RFFRDet, a rotation detector based on refined feature decoupling. First, an Integrated Local Attention (ILA) module is constructed that enables material-aware and geometry-aware feature enhancement through channel-spatial decoupling. Second, a Multi-Resolution Feature Fusion (MRFF) network is designed that achieves optimal coupling between fine-grained spatial positioning and high-level semantic understanding through parallel multi-scale aggregation. Finally, we construct the first multi-directional X-ray security benchmark dataset with rotation annotations for 27,000 images. Extensive experiments demonstrate that RFFRDet achieves 97.4% mAR and 96.8% mAP, representing improvements of 1.6% and 1.5% over state-of-the-art methods, respectively.
- 9 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 21 (15566013) , 2905-2917
RFFRDet: A Refined Feature Fusion Rotation Detector for Prohibited Item Recognition in X-Ray Images
With increasing security threats in critical infrastructure, intelligent X-ray inspection systems have become essential for modern security frameworks. Contraband detection faces dual challenges: semantic dilution across cross-scale features and degraded directional sensitivity. Existing detection paradigms rely on Euclidean spatial assumptions while ignoring X-ray projection geometry characteristics, leading to feature representation ambiguity and insufficient spatial relationship modeling in complex occlusion scenarios. To address these challenges, we propose RFFRDet, a rotation detector based on refined feature decoupling. First, an Integrated Local Attention (ILA) module is constructed that enables material-aware and geometry-aware feature enhancement through channel-spatial decoupling. Second, a Multi-Resolution Feature Fusion (MRFF) network is designed that achieves optimal coupling between fine-grained spatial positioning and high-level semantic understanding through parallel multi-scale aggregation. Finally, we construct the first multi-directional X-ray security benchmark dataset with rotation annotations for 27,000 images. Extensive experiments demonstrate that RFFRDet achieves 97.4% mAR and 96.8% mAP, representing improvements of 1.6% and 1.5% over state-of-the-art methods, respectively.
- 24 November 2025
- Published by Elsevier in Control Engineering Practice
- Vol. 167
Incorporating accident risk evolution processes into vehicle motion control for defensive driving
- 24 November 2025
- Published by Elsevier in Control Engineering Practice
- Vol. 167
Incorporating accident risk evolution processes into vehicle motion control for defensive driving
- 11 September 2025
- Published by Elsevier in Accident Analysis & Prevention
- Vol. 222, 108224
A crash-injury-aware driving safety field for real-time risk assessment and its application in autonomous vehicle motion planning
- 11 September 2025
- Published by Elsevier in Accident Analysis & Prevention
- Vol. 222, 108224
A crash-injury-aware driving safety field for real-time risk assessment and its application in autonomous vehicle motion planning
- 3 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. PP (00983063) , 1
MMCF-Net: A Micro-Motion Context Fusion Framework for Pedestrian Crossing Intention Recognition in Urban Scenes
As advanced vision sensors and edge computing platforms become increasingly integrated into modern vehicles as consumer technologies and electronics, how to accurately recognize pedestrian crossing intentions in urban environments is a noteworthy consideration. Existing research largely relies on skeletal features or trajectory prediction, which demonstrates limited performance in complex scenarios. A Micro-Motion Context Fusion Network (MMCF-Net) is proposed to address these limitations. First, a high-resolution spatiotemporal attention module is constructed that captures sub-pixel-level micro-motion features such as head direction shifts, center of gravity adjustments, and limb pre-activation through deformable convolution and dual attention mechanisms. Second, a hierarchical multi-modal context encoder is designed that adaptively fuses traffic element semantics, vehicle dynamics, and perspective-corrected spatial relationships. Finally, a dynamic weight allocation network is introduced that implements optimal combination of micro-motion features and contextual information based on scene complexity. Experiments demonstrate that MMCF-Net achieves an F1 score of 92.3%, an 7.0% improvement over existing methods like PIE, with enhanced robustness in low-light and occluded scenarios, providing a novel approach to pedestrian safety in complex urban environments.
- 3 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. PP (00983063) , 1
MMCF-Net: A Micro-Motion Context Fusion Framework for Pedestrian Crossing Intention Recognition in Urban Scenes
As advanced vision sensors and edge computing platforms become increasingly integrated into modern vehicles as consumer technologies and electronics, how to accurately recognize pedestrian crossing intentions in urban environments is a noteworthy consideration. Existing research largely relies on skeletal features or trajectory prediction, which demonstrates limited performance in complex scenarios. A Micro-Motion Context Fusion Network (MMCF-Net) is proposed to address these limitations. First, a high-resolution spatiotemporal attention module is constructed that captures sub-pixel-level micro-motion features such as head direction shifts, center of gravity adjustments, and limb pre-activation through deformable convolution and dual attention mechanisms. Second, a hierarchical multi-modal context encoder is designed that adaptively fuses traffic element semantics, vehicle dynamics, and perspective-corrected spatial relationships. Finally, a dynamic weight allocation network is introduced that implements optimal combination of micro-motion features and contextual information based on scene complexity. Experiments demonstrate that MMCF-Net achieves an F1 score of 92.3%, an 7.0% improvement over existing methods like PIE, with enhanced robustness in low-light and occluded scenarios, providing a novel approach to pedestrian safety in complex urban environments.
- 23 July 2025
- Published by Springer Nature in Signal, Image and Video Processing
- Vol. 19 (11) , 1-11
DualStream-TGCN: A Parallel Transformer-GCN Framework for Pedestrian Trajectory Prediction in Dynamic Environments
- 23 July 2025
- Published by Springer Nature in Signal, Image and Video Processing
- Vol. 19 (11) , 1-11
DualStream-TGCN: A Parallel Transformer-GCN Framework for Pedestrian Trajectory Prediction in Dynamic Environments
- 15 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (19) , 40288-40304
A Crash-Injury-Aware Defensive Motion Planning Framework With Accident Knowledge Extraction for Intelligent Connected Vehicles
Enabled by vehicle-to-vehicle (V2V) communication, the real-time sharing of motion states offers new opportunities to enhance the safety of intelligent connected vehicles (ICVs). This paper introduces a crash-injury-aware defensive motion planning framework that leverages real-time V2V information with crash injury severity prediction for ICVs safety enhancement. First, an interpretable machine learning model is developed using the eXtreme Gradient Boosting (XGBoost) algorithm combined with SHapley Additive exPlanations (SHAP), which identifies key features and qualitatively analyzes the complex interactions that influence crash injury severity. Then, an enhanced kinetic energy model of crash injury severity is established based on the key features, in which the impact area tuning factors are innovatively formulated to account for the injury distribution caused by vehicle geometry. Next, we develop a proactive risk assessment module that conceptualizes driving risk as the spatial projection of injury severity from a hypothetical collision, enabling early detection of potential hazards with V2V information. This module is integrated into a model predictive control (MPC) framework to create a motion planner that enables ICVs to maintain a consistently defensive posture and proactively prevent exposure to high-risk conditions. Finally, case studies validate the effectiveness and superiority of the proposed approach, demonstrating that under normal driving conditions, it proactively intercept the propagation of potential risk at early stages, while under critical conditions, it effectively mitigates crashes by explicitly accounting for key factors influencing crash injury severity.
- 15 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (19) , 40288-40304
A Crash-Injury-Aware Defensive Motion Planning Framework With Accident Knowledge Extraction for Intelligent Connected Vehicles
Enabled by vehicle-to-vehicle (V2V) communication, the real-time sharing of motion states offers new opportunities to enhance the safety of intelligent connected vehicles (ICVs). This paper introduces a crash-injury-aware defensive motion planning framework that leverages real-time V2V information with crash injury severity prediction for ICVs safety enhancement. First, an interpretable machine learning model is developed using the eXtreme Gradient Boosting (XGBoost) algorithm combined with SHapley Additive exPlanations (SHAP), which identifies key features and qualitatively analyzes the complex interactions that influence crash injury severity. Then, an enhanced kinetic energy model of crash injury severity is established based on the key features, in which the impact area tuning factors are innovatively formulated to account for the injury distribution caused by vehicle geometry. Next, we develop a proactive risk assessment module that conceptualizes driving risk as the spatial projection of injury severity from a hypothetical collision, enabling early detection of potential hazards with V2V information. This module is integrated into a model predictive control (MPC) framework to create a motion planner that enables ICVs to maintain a consistently defensive posture and proactively prevent exposure to high-risk conditions. Finally, case studies validate the effectiveness and superiority of the proposed approach, demonstrating that under normal driving conditions, it proactively intercept the propagation of potential risk at early stages, while under critical conditions, it effectively mitigates crashes by explicitly accounting for key factors influencing crash injury severity.
- Open Access
- 10 July 2025
- Published by Springer Nature in Research Square
WAFAN: Bridging Domain Gaps in X-ray Security Inspection via Weighted Adaptive Feature Adversarial Networks
X-ray security inspection systems face significant challenges when deployed across different scanning devices due to domain shifts that substantially degrade prohibited item recognition performance. Existing deep learning approaches assume consistent label distributions between training and testing domains, which is impractical in real-world security scenarios where devices vary in penetration capabilities and imaging characteristics. To address this critical limitation, we propose a Weighted Adaptive Feature Adversarial Network (WAFAN) that leverages partial domain adaptation (PDA) to achieve robust cross-domain prohibited item recognition. Our approach introduces three key innovations: (1) a similarity-based weight assignment strategy that effectively distinguishes relevant target-domain classes from irrelevant source-domain categories, mitigating negative transfer effects; (2) an adaptive convolution module integrated via residual connections that explicitly captures and compensates for cross-domain feature distribution differences; and (3) a multi-space adversarial learning framework that ensures domain-invariant feature extraction. Comprehensive experiments on PIDray and SIXray datasets demonstrate that WAFAN achieves superior performance over state-of-the-art PDA methods, with classification accuracy exceeding 80% on target domains while maintaining computational efficiency. The proposed method offers a practical solution for deploying AI-powered security inspection systems across heterogeneous X-ray devices without requiring extensive retraining or domain-specific annotations.
- 10 July 2025
- Published by Springer Nature in Research Square
Open AccessWAFAN: Bridging Domain Gaps in X-ray Security Inspection via Weighted Adaptive Feature Adversarial Networks
X-ray security inspection systems face significant challenges when deployed across different scanning devices due to domain shifts that substantially degrade prohibited item recognition performance. Existing deep learning approaches assume consistent label distributions between training and testing domains, which is impractical in real-world security scenarios where devices vary in penetration capabilities and imaging characteristics. To address this critical limitation, we propose a Weighted Adaptive Feature Adversarial Network (WAFAN) that leverages partial domain adaptation (PDA) to achieve robust cross-domain prohibited item recognition. Our approach introduces three key innovations: (1) a similarity-based weight assignment strategy that effectively distinguishes relevant target-domain classes from irrelevant source-domain categories, mitigating negative transfer effects; (2) an adaptive convolution module integrated via residual connections that explicitly captures and compensates for cross-domain feature distribution differences; and (3) a multi-space adversarial learning framework that ensures domain-invariant feature extraction. Comprehensive experiments on PIDray and SIXray datasets demonstrate that WAFAN achieves superior performance over state-of-the-art PDA methods, with classification accuracy exceeding 80% on target domains while maintaining computational efficiency. The proposed method offers a practical solution for deploying AI-powered security inspection systems across heterogeneous X-ray devices without requiring extensive retraining or domain-specific annotations.
- Open Access
Driver Injury Prediction and Factor Analysis in Passenger Vehicle-to-Passenger Vehicle Collision Accidents Using Explainable Machine Learning
Vehicle accidents, particularly PV-PV collisions, result in significant property damage and driver injuries, causing substantial economic losses and health risks. Most existing studies focus on macro-level predictions, such as accident frequency, but lack detailed collision-level analysis, which limits the precision of severity prediction. This study investigates various accident-related factors, including environmental conditions, vehicle attributes, driver characteristics, pre-crash scenarios, and collision dynamics. Data from NHTSA’s CRSS and FARS datasets were integrated and balanced using random over-sampling and under-sampling techniques to address severity-level data imbalances. The mRMR algorithm was employed for feature selection to minimize redundancy and identify key features. Five advanced machine learning models were evaluated for severity prediction, with XGBoost achieving the best performance: 84.9% accuracy, 84.85% precision, 84.90% recall, and an F1-score of 84.87%. SHAP analysis was utilized to interpret the model and conduct a comprehensive analysis of accident features, including their importance, dependencies, and combined effects on severity prediction. This study achieved high accuracy in predicting accident severity across all levels in PV-PV collisions. Moreover, by integrating the SHAP model interpretation method, we conducted detailed feature analysis at global, local, and individual case levels, thereby filling the gap in PV-PV accident severity prediction and feature analysis.
Open AccessDriver Injury Prediction and Factor Analysis in Passenger Vehicle-to-Passenger Vehicle Collision Accidents Using Explainable Machine Learning
Vehicle accidents, particularly PV-PV collisions, result in significant property damage and driver injuries, causing substantial economic losses and health risks. Most existing studies focus on macro-level predictions, such as accident frequency, but lack detailed collision-level analysis, which limits the precision of severity prediction. This study investigates various accident-related factors, including environmental conditions, vehicle attributes, driver characteristics, pre-crash scenarios, and collision dynamics. Data from NHTSA’s CRSS and FARS datasets were integrated and balanced using random over-sampling and under-sampling techniques to address severity-level data imbalances. The mRMR algorithm was employed for feature selection to minimize redundancy and identify key features. Five advanced machine learning models were evaluated for severity prediction, with XGBoost achieving the best performance: 84.9% accuracy, 84.85% precision, 84.90% recall, and an F1-score of 84.87%. SHAP analysis was utilized to interpret the model and conduct a comprehensive analysis of accident features, including their importance, dependencies, and combined effects on severity prediction. This study achieved high accuracy in predicting accident severity across all levels in PV-PV collisions. Moreover, by integrating the SHAP model interpretation method, we conducted detailed feature analysis at global, local, and individual case levels, thereby filling the gap in PV-PV accident severity prediction and feature analysis.
- Open Access
- 16 April 2025
- Published by Springer Nature in Research Square
A Real-Time Context-Dependent Driver Attention AssessmentSystem Using Pupil Orientation and Head Posture
With the popularization of social media andInternet technology, people are more and more easilydistracted when driving, which has a serious adverseimpact on the safety of expected functions, and reducesthe efficiency of control transmission between humanand computer. This article proposes a real-time pupil re-construction (PURE) algorithm for real-time detectionof pupil direction, and combines it with a 3D eye modelto obtain the gaze direction of the pupil in real time.Subsequently, the Posit algorithm was used to obtain real-time head posture, and the head posture and pupil gaze direction were used as improved input features toestimate the driver’s gaze area in real-time using a random forest model. Finally, this article comprehensivelyconsiders the impact of multiple real-time indicatorson driving attention, such as the number of eye gazeareas, eye scanning areas, road conditions, brain latency,and driving tasks, and proposes a prototype framework for real-time driver multi indicator attention evaluation.Through various real-time evaluation experiments,we have demonstrated that: 1) these context relatedattention features are universal when tested in real-timeon 16 drivers in a driving simulator; 2) The proposeddriver attention evaluation model performs well underreal-time conditions and can produce reliable results invarious driving scenarios.
- 16 April 2025
- Published by Springer Nature in Research Square
Open AccessA Real-Time Context-Dependent Driver Attention AssessmentSystem Using Pupil Orientation and Head Posture
With the popularization of social media andInternet technology, people are more and more easilydistracted when driving, which has a serious adverseimpact on the safety of expected functions, and reducesthe efficiency of control transmission between humanand computer. This article proposes a real-time pupil re-construction (PURE) algorithm for real-time detectionof pupil direction, and combines it with a 3D eye modelto obtain the gaze direction of the pupil in real time.Subsequently, the Posit algorithm was used to obtain real-time head posture, and the head posture and pupil gaze direction were used as improved input features toestimate the driver’s gaze area in real-time using a random forest model. Finally, this article comprehensivelyconsiders the impact of multiple real-time indicatorson driving attention, such as the number of eye gazeareas, eye scanning areas, road conditions, brain latency,and driving tasks, and proposes a prototype framework for real-time driver multi indicator attention evaluation.Through various real-time evaluation experiments,we have demonstrated that: 1) these context relatedattention features are universal when tested in real-timeon 16 drivers in a driving simulator; 2) The proposeddriver attention evaluation model performs well underreal-time conditions and can produce reliable results invarious driving scenarios.
- Open Access
- 26 March 2025
- Published by MDPI AG in Preprints.org
Driver Injury Prediction and Factor Analysis in Passenger Vehicle-to-Passenger Vehicle Collision Accidents Using Explainable Machine Learning
Vehicle accidents, particularly passenger vehicle-to-passenger vehicle (PV-PV) collisions, cause significant property damage and driver injuries, resulting in substantial economic losses and health risks. Most existing studies focus on macro-level predictions, such as accident frequency, but lack detailed collision-level analysis, which limits precise severity prediction. This study investigates various accident-related factors, including environmental conditions, vehicle attributes, driver characteristics, pre-crash scenarios, and collision dynamics.Using data from the National Highway Traffic Safety Administration’s (NHTSA) Crash Report Sampling System (CRSS) and Fatality Analysis Reporting System (FARS), the dataset was balanced by integrating complementary severity-level data with random over-sampling and under-sampling techniques. The Minimum Redundancy Maximum Relevance (mRMR) algorithm was used for feature selection to minimize redundancy and identify key features.Five advanced machine learning models were employed for severity prediction, with Extreme Gradient Boosting (XGBoost) achieving the best performance: 84.9% accuracy, 84.85% precision, 84.90% recall, and an F1-score of 84.87%. SHapley Additive exPlanations (SHAP) analysis was employed to interpret the model and conduct a deep analysis of accident features, including feature importance, dependencies, and their combined effects on severity prediction.The findings highlight the effectiveness of interpretable machine learning in understanding accident severity and identifying key factors influencing driver injuries.
- 26 March 2025
- Published by MDPI AG in Preprints.org
Open AccessDriver Injury Prediction and Factor Analysis in Passenger Vehicle-to-Passenger Vehicle Collision Accidents Using Explainable Machine Learning
Vehicle accidents, particularly passenger vehicle-to-passenger vehicle (PV-PV) collisions, cause significant property damage and driver injuries, resulting in substantial economic losses and health risks. Most existing studies focus on macro-level predictions, such as accident frequency, but lack detailed collision-level analysis, which limits precise severity prediction. This study investigates various accident-related factors, including environmental conditions, vehicle attributes, driver characteristics, pre-crash scenarios, and collision dynamics.Using data from the National Highway Traffic Safety Administration’s (NHTSA) Crash Report Sampling System (CRSS) and Fatality Analysis Reporting System (FARS), the dataset was balanced by integrating complementary severity-level data with random over-sampling and under-sampling techniques. The Minimum Redundancy Maximum Relevance (mRMR) algorithm was used for feature selection to minimize redundancy and identify key features.Five advanced machine learning models were employed for severity prediction, with Extreme Gradient Boosting (XGBoost) achieving the best performance: 84.9% accuracy, 84.85% precision, 84.90% recall, and an F1-score of 84.87%. SHapley Additive exPlanations (SHAP) analysis was employed to interpret the model and conduct a deep analysis of accident features, including feature importance, dependencies, and their combined effects on severity prediction.The findings highlight the effectiveness of interpretable machine learning in understanding accident severity and identifying key factors influencing driver injuries.