Related Experiment Videos
Learning to recover weak signals: A two-stage graph-based framework for all-weather RGB-T object detection
Ruoyan Pei1, Pengge Ma1, Jinwang Qian1
1School of Electronics and Information, Zhengzhou University of Aeronautics, Zhengzhou, China.
Plos One
|August 5, 2026
Summary
This study introduces a novel two-stage multimodal object detection framework using attention fusion and graph neural networks. The method significantly enhances perception in complex traffic, improving day-night performance and small object detection for autonomous driving.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Robotics
Background:
- Object detection in traffic is hindered by illumination changes and difficulty detecting small, distant objects.
- Current methods struggle with degraded features and weak object cues in complex environments.
Purpose of the Study:
- To develop a robust multimodal object detection framework for challenging traffic scenarios.
- To improve perception accuracy and efficiency for autonomous driving systems.
Main Methods:
- A two-stage framework integrating RGB and thermal data using a Dual-Path Attention Fusion Module (DPAFM).
- A Hierarchical Attention Graph Neural Network (HA-GNN) for relational reasoning on detection candidates.
- Utilized KAIST and R-LiViT datasets for comprehensive evaluation.
Main Results:
- Significantly reduced the day-night performance gap from 72.4% to 7.4%.
- Substantially improved detection of small objects in complex traffic conditions.
- Achieved real-time inference speed of 27.4 FPS.
Conclusions:
- The proposed framework offers enhanced robustness and accuracy for object detection in adverse traffic conditions.
- The method provides a strong balance between performance and computational efficiency for autonomous driving.
- The approach effectively mitigates illumination-induced degradation and improves small object recognition.
