YOLO-FFRD: Dynamic Small-Scale Pedestrian Detection Algorithm Based on Feature Fusion and Rediffusion Structure

Abstract

To address the challenges of detecting dynamic small targets such as pedestrians in complex dynamic environments for mobile robots, this paper proposes a dynamic small-target detection algorithm based on feature fusion and rediffusion structure, which is suitable for deployment on mobile robot platforms. Mobile robots can utilize depth camera information to identify and avoid small targets like pedestrians and vehicles in complex environments. Traditional deep learning-based object detection algorithms perform poorly when applied to the field of mobile robotics, especially in detecting dynamic small targets. To improve this, we apply an enhanced object recognition algorithm to mobile robot platforms. To verify the effectiveness of the proposed algorithm, we conduct relevant tests and ablation studies in various environments and perform multi-class small-target detection on the public VisDrone2019 dataset. Compared with the original YOLOv8 algorithm, our proposed method improves accuracy by 5% and increases mAP0.5 and mAP0.5-0.95 by approximately 3%. Overall, the experimental results show that the high-performance small-target detection algorithm based on feature fusion and rediffusion structure significantly reduces the miss detection rate and exhibits good generalization ability, which can be extended to multi-class small-target detection. This is of great significance for improving the environmental perception ability of robots.