TFDet: Target-Aware Fusion for RGB-T Pedestrian Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Pedestrian detection is crucial for traffic safety.
• RGB-based methods struggle in low-light conditions.
• Multispectral approaches using thermal data improve performance but face false positive issues.

Purpose of the Study:

• To develop a novel target-aware fusion strategy for multispectral pedestrian detection.
• To address the negative impact of false positives in fused feature maps.
• To enhance feature contrast for reduced false positives and improved detection.

Main Methods:

• Proposed TFDet, a target-aware fusion strategy with a fusion-refinement paradigm.
• Utilized an adaptive receptive field to integrate RGB and thermal features.
• Employed a segmentation branch and a correlation-maximum loss function to enhance feature contrast.

Main Results:

• TFDet achieved state-of-the-art performance on KAIST and LLVIP pedestrian detection benchmarks.
• Demonstrated significant absolute gains over previous methods (0.65% and 4.1%).
• Showcased strong performance in multiclass object detection and maintained efficiency.

Conclusions:

• TFDet effectively reduces false positives by enhancing feature contrast.
• The method significantly improves multispectral pedestrian detection, particularly in low-light scenarios.
• TFDet offers a robust and efficient solution for advancing road safety through computer vision.