MarkYolo: An Enhanced YOLOv10 Network with Dynamic Convolution and Attention Mechanism for Circular Marker Detection in High-Speed Video Measurement

Abstract

In high-speed video measurement, accurate detection of circular markers is critical for applications in structural analysis, motion tracking, and industrial automation. Traditional marker detection methods often struggle with challenges such as dynamic occlusion, complex backgrounds, and scale variations. To address these issues, this paper proposes MarkYolo, an enhanced object detection framework based on YOLOv10 tailored for robust circular marker detection. Key innovations include: (1) Omni-dimensional Dynamic Convolution (ODConv) integrated into a novel COD module to capture multi-dimensional contextual features while reducing computational complexity; (2) an Adaptive Fine-Grained Channel Attention (AFGCAttention) mechanism to enhance small object localization by adaptively fusing global and local information; and (3) Normalized Wasserstein Distance (NWD) loss to improve robustness against positional shifts and scale variations by modeling bounding boxes as Gaussian distributions. Experiments on the CME dataset demonstrate that MarkYolo achieves a state-of-the-art AP50-95 of 75.4%, outperforming the baseline YOLOv10 by 4.9% while maintaining real-time efficiency. The model also reduces false positives and missed detections in complex scenarios, offering significant advancements for high-speed photogrammetry applications. Further ablation studies validate the synergistic contributions of each proposed module, highlighting improvements in recall (96.4%), precision (98.3%), and computational efficiency (8.6 GFLOPs). This work provides a practical solution for enhancing marker detection accuracy in dynamic environments and lays a foundation for future lightweight deployments on edge devices.