An Excavator Action Recognition Method Using Skeletal Keypoints and Monocular Depth Estimation Network

Excavator action recognition plays a vital role in productivity assessment and cost analysis, contributing to improved construction efficiency, reduced project duration, and lower costs. However, due to the continuous transitions among excavator actions, recognition errors frequently occur, and manual annotation is prone to inaccuracy without frame-by-frame comparison. To address these issues, this study proposes an excavator action recognition method based on skeletal keypoints and monocular depth estimation. Specifically, this study first uses the YOLOv8 algorithm to detect excavator keypoints, and then utilizes a monocular depth estimation network to obtain the relative depth information of each keypoint, forming three-dimensional spatial coordinate sequences. These 3D keypoint sequences are then input into a Long Short-Term Memory (LSTM) network to learn spatiotemporal variation patterns and classify excavator actions. Experimental results show that the proposed method achieves a precision of 93.44%, a recall of 87.21%, an F1 score of 89.68%, and an accuracy of 91.43%. These findings demonstrate that the proposed method improves recognition accuracy and robustness, providing a reliable approach for excavator action analysis in construction monitoring.