Action Classification in Human Robot Interaction Cells in Manufacturing: Moving Towards Mutual Performance Monitoring Capacity

Action recognition has become a prerequisite approach to fluent Human-Robot Interaction (HRI) due to a high degree of movement flexibility. With the improvements in machine learning algorithms, robots are gradually transitioning into more human-populated areas. However, HRI systems demand the need for robots to possess enough cognition. The action recognition algorithms require massive training datasets, structural information of objects in the environment, and less expensive models in terms of computational complexity. In addition, many such algorithms are trained on datasets derived from daily activities. The algorithms trained on non-industrial datasets may have an unfavorable impact on implementing models and validating actions in an industrial context. This study proposed a lightweight deep learning model for classifying low-level actions in an assembly setting. The model is based on optical flow feature elicitation and mobilenetV2-SSD action classification and is trained and assessed on an actual industrial activities' dataset. The experimental outcomes show that the presented method is futuristic and does not require extensive preprocessing; therefore, it can be promising in terms of the feasibility of action recognition for mutual performance monitoring in real-world HRI applications. The test result shows 80% accuracy for low-level RGB action classes. The study's primary objective is to generate experimental results that may be used as a reference for future HRI algorithms based on the InHard dataset.