This research focuses on leveraging the YOLOv5s lightweight model to achieve real-time recognition of indoor natural gas work behavior using wearable devices. The approach involves utilizing the YOLOv5s model to process video captured by wearable devices with cameras, detecting and tracking hands, gas leakage detectors, and scenes. Subsequently, the WIoU calculation method is employed to determine the temporal characteristics of containment relationships between different categories, thus ascertaining the completion of specific work actions. The model is trained and tested using a dataset collected through on-site experiments, and its recognition accuracy and real-time performance are evaluated. The experimental results demonstrate significant progress in indoor gas operation behavior recognition, achieving real-time and accurate identification of indoor gas operators’ behaviors.