PF-YOLO: An Improved YOLOv8 for Small Object Detection in Fisheye Images

Abstract

Top-view fisheye cameras are widely used in personnel surveillance for their broad field of view, but their unique imaging characteristics pose challenges like distortion, complex scenes, scale variations, and small objects near image edges. To tackle these, we proposed peripheral focus you only look once (PF-YOLO), an enhanced YOLOv8n-based method. Firstly, we introduced a cutting-patch data augmentation strategy to mitigate the problem of insufficient small-object samples in various scenes. Secondly, to enhance the model’s focus on small objects near the edges, we designed the peripheral focus loss, which uses dynamic focus coefficients to provide greater gradient gains for these objects, improving their regression accuracy. Finally, we designed the three dimensional (3D) spatial-channel coordinate attention C2f module, enhancing spatial and channel perception, suppressing noise, and improving personnel detection. Experimental results demonstrate that PF-YOLO achieves strong performance on the challenging events for person detection from overhead fisheye images (CEPDTOF) and in-the-wild events for people detection and tracking from overhead fisheye cameras (WEPDTOF) datasets. Compared to the original YOLOv8n model, PF-YOLO achieves improvements on CEPDTOF with increases of 2.1%, 1.7% and 2.9% in mean average precision 50 ($ \mathrm{mAP\ 50} $), $ \mathrm{mAP\ 50-95} $, and $ \mathrm{recall} $, reaching 95.7%, 65.8% and 95.5%, respectively. On WEPDTOF, PF-YOLO achieves substantial improvements with increases of 31.4%, 14.9%, 61.1% and 21.0% in $ \mathrm{mAP\ 50} $, $ \mathrm{mAP\ 50-95} $, $ \mathrm{precision} $ and $ \mathrm{recall} $ reaching 53.1%, 22.8%, 91.2% and 57.2%, respectively.

Keywords

fisheye / object detection and recognition / small object detection / deep learning