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

## TL;DR

This paper introduces FID-YOLO, a new pedestrian detection model that improves accuracy in complex environments by combining visible and infrared images.

## Contribution

The novel FID-YOLO model integrates multispectral data using an illumination-aware fusion module and a scale-adaptive detection head.

## Key findings

- FID-YOLO outperforms benchmark models on the M3FD and LLVIP datasets.
- Ablation experiments confirm the importance of each proposed module.
- The model improves detection accuracy in adverse conditions like occlusions and scale variations.

## Abstract

The advancement of pedestrian detection technology is of great importance for various applications such as intelligent driving, object tracking, and robot navigation. Many studies in this field have demonstrated that image quality significantly contributes to the precision of detection. However, unexpected factors such as adverse weather, occlusions, and scale variations, which extremely weaken the main features of the detected objects, leading to a decrease in detection accuracy. To address these problems, we propose a Feature-enriched Image Detection-YOLO (FID-YOLO), to improve pedestrian detection performance in complex environments by integrating visible and infrared light information. Specifically, we design an illumination-aware image fusion module for visible and infrared image information fusion to generate a new image within more information to enrich pedestrian features. Then, a cascaded feature aggregation module using reparameterization and channel shuffle is introduced to enhance the model’s understanding and generalization capabilities for complex scenes. Furthermore, we exploit a scale-adaptive feature detection head for YOLO detector, which solves the problem of detecting small objects at varying object scales. Experiments on M3FD and LLVIP datasets demonstrate that FID-YOLO outperforms the benchmark models in pedestrian detection. Additionally, we validate the indispensability of each proposed module through ablation experiments.