Efficient Skeleton-Based Human Assembly Action Recognition Optimized by Data Augmentation

摘要

As a new assembly model, human-robot collaborative assembly has attracted wide attention due to the diverse requirements of assembly tasks, complex shapes of assembly parts, and limited space in the current manufacturing industry. In the human-robot collaborative assembly environment of complex products, how to recognize human actions has become a key issue to better promote human-robot interaction. Firstly, considering the requirements of safety, efficiency, and accuracy in the human-robot collaborative assembly environment, we realize skeleton-based assembly action recognition. The channel-wise topology refinement graph convolution network (CTR-GCN) is utilized to extract the 3D skeletal features of the operator. In addition, considering the complexity of the assembly action dataset collection, we improve the recognition accuracy by data augmentation and find an appropriate training dataset size. Finally, in this paper, we focus on three assembly actions of the reducer and make an assembly action dataset in NTU RGB+D format. The optimization model based on data augmentation achieved high accuracy on the self-made assembly action dataset.