Skeleton-based action recognition for industrial packing process

The applications of action recognition in real-world scenarios are challenging. Although state-of-the-art methods have demonstrated good performance on large scale datasets, we still face complex practical problems and inappropriate models. In this work, we propose a novel local image directed graph neural network (LI-DGNN) to solve a real-world production scenario problem which is the completeness identification of accessories during the range hood packing process in a kitchen appliance manufacturing workshop. LI-DGNN integrates skeleton-based action recognition and local image classification to make good use of both human skeleton data and appearance information for action recognition. The experimental results demonstrate the high recognition accuracy and good generalization ability on the range hood packing dataset (RHPD) which is generated in the industrial packing process. The results can meet the recognition requirements in the actual industrial production process.