Young-nam Song, Seung-gyu Kim, Tae-Ho Im, Journal of Internet Computing and Services, Vol. 26, No. 5, pp. 29-38, Oct. 2025
10.7472/jksii.2025.26.5.29, Full Text:  HTML
Keywords: Life-jacket Detection, YOLO11n, Dual Detection Method, Marine Safety, small-object detection

Abstract

This study addresses the critical need for enhancing maritime safety by developing an automatic detection method for verifying life jacket usage, overcoming the limitations of manual surveillance and conventional object detection techniques. To this end, we propose a lightweight two-stage detection framework based on YOLO11n. In the first stage, human objects are detected to define the region of interest (ROI), and in the second stage, life jacket usage is precisely identified within the ROI. A dedicated dataset reflecting real maritime CCTV conditions with blur augmentation was constructed for training and evaluation. Experimental results demonstrate that the proposed method achieves superior performance, attaining mAP@50 of 0.93 and mAP@50:95 of 0.90, outperforming YOLO11s (0.92 / 0.83) and Faster R-CNN (0.92 / 0.82). In terms of efficiency, the proposed approach achieves an average inference latency of 36.65 ms (≈27.3 FPS), comparable to YOLO11s (32.58 ms, ≈30.7 FPS), while being approximately 27 times faster than Faster R-CNN (925.86 ms, ≈1.1 FPS). These findings validate that the proposed sequential lightweight model offers a practical and efficient solution for maritime safety monitoring by balancing accuracy and real-time capability.

Statistics
Cite this article