Scarce samples learning in traffic object detection based on improved YOLOv8s

Abstract

Aiming at poor detection performance of YOLOv8s for scarce traffic samples in domestic road scenes, this paper proposes FCC-YOLOv8s with three optimizations. A Feature Similarity Fusion Module (FSFM) adopts dual-distance metric learning to reduce data dependence via a feature reference library. Improved CCIoU loss supplements diagonal size constraint to boost bounding box regression accuracy. Lightweight CBAM is embedded after backbone P3/P4 C2f and P5 SPPF to suppress background noise. Experiments on a self-built Chinese traffic dataset show the proposed method reaches 86.75% mAP@0.5, 13.97% higher than original YOLOv8s, with rare-class AP increased by 21.33%, effectively improving scarce-object detection.