A drill pipe counting method based on skeleton action recognition

Abstract

A drill pipe counting method based on human skeleton sequence recognition is proposed to address the issues of high workload, high environmental requirements, and large counting errors in existing drill pipe counting methods. Firstly, obtain human skeleton information through YOLOv7 and FastPose, and create a human skeleton dataset; Secondly, based on ST-GCN, an ML-ASTGNet action recognition network is designed, which captures more global contextual information through adaptive graph convolution (A-GCN) and improves the spatial modeling ability of the action recognition network; Introducing the Time Motion Excitation Module (TME), which utilizes the idea of time difference to characterize motion information, generates larger weights for frames with larger motion amplitudes to highlight sensitive features of motion, and designs a Multi Scale Multi Fine Granular Time Convolution (DM-TCN) to learn the final time feature information from different scales, improving the time modeling ability of action recognition networks. The experimental results show that the accuracy of ML-ASTGNet on the self built human skeleton dataset reaches 92.1%, which is 8.2% higher than ST-GCN and shows better action recognition performance. Finally, this method has achieved good counting performance in actual drill pipe counting tests with small counting errors.