Hand gesture recognition for safe operation of tower cranes: a thermal image-based method using enhanced YOLOv8

Purpose Effective communication between operators and signalmen is critical for safe tower crane operations, particularly during blind lifts and under poor lighting conditions. Existing vision-based hand gesture recognition methods predominantly rely on RGB images, which are highly sensitive to illumination variations, while alternative sensing modalities often lack sufficient interpretability for safety-critical construction applications. This study aims to develop a robust and interpretable hand gesture recognition approach to support safer tower crane operations under varying lighting conditions. Design/methodology/approach A thermal dataset containing 25 standardized hand gestures was developed from real construction sites. In addition, an enhanced YOLOv8 model was designed by incorporating attention mechanisms and a dynamic detection head to improve recognition performance on thermal images. Findings The results show that thermal images enable reliable gesture recognition under low-light conditions while retaining the interpretability of vision-based solutions. The proposed model achieved a precision of 97.26% and an mAP@0.5 of 99.05%, outperforming the baseline model and several recent YOLO variants. These findings demonstrate the feasibility of thermal imaging for supporting safer tower crane operations. Originality/value This study firstly introduces thermal imagery as a promising visual modality for hand gesture recognition in tower crane operations and provides a tailored deep learning solution for thermal data, contributing to intelligent assistance systems for safe tower crane operations.