Selected work

Representative Papers

DinoRADE: Full Spectral Radar-Camera Fusion with Vision Foundation Model Features for Multi-class Object Detection in Adverse Weather

Apr 09, 2026

This work addresses the insufficient detection accuracy of existing radar-camera fusion methods for small and vulnerable road users (VRUs) under adverse weather conditions, as well as the lack of fine-grained, multi-class evaluation. To overcome these limitations, the authors propose a radar-centric fusion framework that leverages deformable cross-attention to aggregate features extracted by the DINOv3 vision foundation model around transformed reference points in the camera view. This enables full-spectrum fusion between dense FMCW radar tensors and visual semantics, complemented by a cross-modal feature alignment strategy. The method reports, for the first time, individual detection performance across five object classes on the K-Radar dataset, significantly outperforming state-of-the-art approaches under all-weather conditions and achieving a 12.1% improvement in multi-class detection accuracy.

0 citationsRead paper

DTP-Attack: A decision-based black-box adversarial attack on trajectory prediction

Mar 27, 2026

Trajectory prediction is critical for autonomous driving, yet existing adversarial attacks rely on white-box access or strong physical constraints, limiting their real-world applicability. This work proposes the first black-box adversarial attack framework that requires only binary decision outputs from the target model. By leveraging a boundary-walking algorithm and trajectory proximity constraints, the method generates highly realistic adversarial examples without access to gradients or internal model information. It simultaneously supports both intent misclassification and degradation of prediction accuracy, and is effective against mainstream models such as Trajectron++ and Grip++. Evaluated on the nuScenes and Apolloscape datasets, the approach achieves intent misclassification success rates of 41%–81% with perturbations under 0.45 meters and increases prediction error by 1.9–4.2, significantly outperforming existing black-box methods.

0 citationsRead paper

RADE-Net: Robust Attention Network for Radar-Only Object Detection in Adverse Weather

Feb 23, 2026

This work addresses the limitations of optical sensors in adverse weather and the information loss inherent in existing radar-based detection methods, which often rely on sparse point clouds or 2D projections. To overcome these challenges, the authors propose a novel 3D polar-coordinate projection tailored for 4D radar tensors and introduce RADE-Net, a lightweight network that integrates spatial and channel attention mechanisms to directly predict object centers and regress oriented 3D bounding boxes in the polar domain. The method uniquely preserves Doppler and elevation information while reducing data volume by 91.9%, substantially improving computational efficiency. Evaluated on the K-Radar dataset, the approach achieves a 16.7% mAP gain over the baseline and outperforms the current state-of-the-art radar-only method by 6.5%, notably surpassing certain LiDAR-based solutions under rain, snow, and fog conditions.

0 citationsRead paper

Multi-Agent AI Framework for Road Situation Detection and C-ITS Message Generation

Nov 10, 2025

Traditional road detection methods suffer from poor generalization in unseen scenarios and lack semantic interpretability, undermining the reliability of C-ITS traffic advisories. To address this, we propose a multi-agent collaborative framework integrating Gemini multimodal large language models with visual perception for real-time road situation understanding and standardized C-ITS message generation. Specialized agents perform situational recognition, distance estimation, decision reasoning, and structured message generation, enhancing both system interpretability and semantic reasoning capability. Evaluated on a custom road dataset comprising 103 images, our approach achieves 100% situation detection rate and fully correct syntactic message generation. Experiments demonstrate that Gemini-2.0-Flash outperforms Gemini-2.5-Flash in both accuracy and inference latency. This work establishes a novel, interpretable, and scalable multimodal semantic understanding paradigm for intelligent transportation systems.

0 citationsRead paper

Investigating Traffic Accident Detection Using Multimodal Large Language Models

Sep 23, 2025

To address the heavy reliance of roadside camera-based accident detection on large-scale annotated data, this paper proposes a zero-shot multimodal reasoning framework. First, a synthetic DeepAccident dataset is constructed to mitigate the scarcity of real-world accident samples. Second, structured visual prompts are generated by integrating YOLO (object detection), Deep SORT (multi-object tracking), and SAM (instance segmentation), thereby enhancing the spatiotemporal reasoning capabilities of multimodal large language models (MLLMs). Evaluation on Gemini, Gemma-3, and Pixtral demonstrates that Pixtral achieves an F1-score of 0.71 and recall of 83%; prompt-optimized Gemini attains 90% precision; and Gemma-3 exhibits superior robustness. This work pioneers the synergistic integration of vision foundation models and MLLMs for zero-shot traffic incident understanding—significantly improving accident recognition accuracy, interpretability, and system scalability without requiring task-specific labeled data.

0 citationsRead paper

Recent publications

Latest Papers

DinoRADE: Full Spectral Radar-Camera Fusion with Vision Foundation Model Features for Multi-class Object Detection in Adverse Weather

Apr 09, 2026

This work addresses the insufficient detection accuracy of existing radar-camera fusion methods for small and vulnerable road users (VRUs) under adverse weather conditions, as well as the lack of fine-grained, multi-class evaluation. To overcome these limitations, the authors propose a radar-centric fusion framework that leverages deformable cross-attention to aggregate features extracted by the DINOv3 vision foundation model around transformed reference points in the camera view. This enables full-spectrum fusion between dense FMCW radar tensors and visual semantics, complemented by a cross-modal feature alignment strategy. The method reports, for the first time, individual detection performance across five object classes on the K-Radar dataset, significantly outperforming state-of-the-art approaches under all-weather conditions and achieving a 12.1% improvement in multi-class detection accuracy.

0 citationsRead paper

DTP-Attack: A decision-based black-box adversarial attack on trajectory prediction

Mar 27, 2026

Trajectory prediction is critical for autonomous driving, yet existing adversarial attacks rely on white-box access or strong physical constraints, limiting their real-world applicability. This work proposes the first black-box adversarial attack framework that requires only binary decision outputs from the target model. By leveraging a boundary-walking algorithm and trajectory proximity constraints, the method generates highly realistic adversarial examples without access to gradients or internal model information. It simultaneously supports both intent misclassification and degradation of prediction accuracy, and is effective against mainstream models such as Trajectron++ and Grip++. Evaluated on the nuScenes and Apolloscape datasets, the approach achieves intent misclassification success rates of 41%–81% with perturbations under 0.45 meters and increases prediction error by 1.9–4.2, significantly outperforming existing black-box methods.

0 citationsRead paper

RADE-Net: Robust Attention Network for Radar-Only Object Detection in Adverse Weather

Feb 23, 2026

This work addresses the limitations of optical sensors in adverse weather and the information loss inherent in existing radar-based detection methods, which often rely on sparse point clouds or 2D projections. To overcome these challenges, the authors propose a novel 3D polar-coordinate projection tailored for 4D radar tensors and introduce RADE-Net, a lightweight network that integrates spatial and channel attention mechanisms to directly predict object centers and regress oriented 3D bounding boxes in the polar domain. The method uniquely preserves Doppler and elevation information while reducing data volume by 91.9%, substantially improving computational efficiency. Evaluated on the K-Radar dataset, the approach achieves a 16.7% mAP gain over the baseline and outperforms the current state-of-the-art radar-only method by 6.5%, notably surpassing certain LiDAR-based solutions under rain, snow, and fog conditions.

0 citationsRead paper

Multi-Agent AI Framework for Road Situation Detection and C-ITS Message Generation

Nov 10, 2025

Traditional road detection methods suffer from poor generalization in unseen scenarios and lack semantic interpretability, undermining the reliability of C-ITS traffic advisories. To address this, we propose a multi-agent collaborative framework integrating Gemini multimodal large language models with visual perception for real-time road situation understanding and standardized C-ITS message generation. Specialized agents perform situational recognition, distance estimation, decision reasoning, and structured message generation, enhancing both system interpretability and semantic reasoning capability. Evaluated on a custom road dataset comprising 103 images, our approach achieves 100% situation detection rate and fully correct syntactic message generation. Experiments demonstrate that Gemini-2.0-Flash outperforms Gemini-2.5-Flash in both accuracy and inference latency. This work establishes a novel, interpretable, and scalable multimodal semantic understanding paradigm for intelligent transportation systems.

0 citationsRead paper

Investigating Traffic Accident Detection Using Multimodal Large Language Models

Sep 23, 2025

To address the heavy reliance of roadside camera-based accident detection on large-scale annotated data, this paper proposes a zero-shot multimodal reasoning framework. First, a synthetic DeepAccident dataset is constructed to mitigate the scarcity of real-world accident samples. Second, structured visual prompts are generated by integrating YOLO (object detection), Deep SORT (multi-object tracking), and SAM (instance segmentation), thereby enhancing the spatiotemporal reasoning capabilities of multimodal large language models (MLLMs). Evaluation on Gemini, Gemma-3, and Pixtral demonstrates that Pixtral achieves an F1-score of 0.71 and recall of 83%; prompt-optimized Gemini attains 90% precision; and Gemma-3 exhibits superior robustness. This work pioneers the synergistic integration of vision foundation models and MLLMs for zero-shot traffic incident understanding—significantly improving accident recognition accuracy, interpretability, and system scalability without requiring task-specific labeled data.

0 citationsRead paper