An Enhanced Feature-Fusion Network for Small-Scale Pedestrian Detection on Edge Devices

Small-scale pedestrian detection is one of the challenges in general object detection. Factors such as complex backgrounds, long distances, and low-light conditions make the image features of small-scale pedestrians less distinct, further increasing the difficulty of detection. To address these challenges, an Enhanced Feature-Fusion YOLO network (EFF-YOLO) for small-scale pedestrian detection is proposed. Specifically, this method employs a backbone based on the FasterNet block within YOLOv8n, which is designed to enhance the extraction of spatial features while reducing redundant operation. Furthermore, the gather-and-distribute (GD) mechanism is integrated into the neck of the network to realize the aggregation and distribution of global information and multi-level features. This not only strengthens the faint features of small-scale pedestrians but also effectively suppresses complex background information, thereby improving the accuracy of small-scale pedestrians. Experimental results indicate that EFF-YOLO achieves detection accuracies of 72.5%, 72.3%, and 91% on the three public datasets COCO-person, CityPersons, and LLVIP, respectively. Moreover, the proposed method reaches a detection speed of 50.7 fps for 1920 x 1080-pixel video streams on the edge device Jetson Orin NX, marking a 15.2% improvement over the baseline network. Thus, the proposed EFF-YOLO method not only boasts high detection accuracy but also demonstrates excellent real-time performance on edge devices.