FID-YOLO: A pedestrian detection model integrating multispectral information in complex environments

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Pedestrian detection is crucial for intelligent driving, object tracking, and robot navigation.
• Image quality significantly impacts detection precision.
• Adverse weather, occlusions, and scale variations degrade pedestrian detection accuracy by weakening object features.

Purpose of the Study:

• To improve pedestrian detection performance in complex environments.
• To address challenges posed by poor image quality and object variations.

Main Methods:

• Proposed Feature-enriched Image Detection-YOLO (FID-YOLO) model.
• Integrated visible and infrared light information using an illumination-aware image fusion module.
• Introduced a cascaded feature aggregation module with reparameterization and channel shuffle.
• Developed a scale-adaptive feature detection head for YOLO.

Main Results:

• FID-YOLO demonstrated superior performance compared to benchmark models on M3FD and LLVIP datasets.
• Experiments validated the effectiveness of each proposed module through ablation studies.
• The system successfully enhanced pedestrian features and improved detection in complex scenarios.

Conclusions:

• FID-YOLO effectively improves pedestrian detection accuracy in challenging environmental conditions.
• The integration of multi-modal information and advanced feature processing modules enhances model generalization.
• The proposed methods offer a significant advancement for safety-critical applications like autonomous driving.