Bento Packaging Activity Recognition Based on Statistical Features

Due to the fast advancements of low-cost micro-embedded sensors and MoCap sensors, human action recognition has become an essential study topic and is garnering a lot of interest in different sectors. Recently, it is drawing a lot of attention in human-robot collaboration to assist human to preform regular tasks step-wise because it is difficult to obtain human labor at a lower wage to monitor industrial works. In this work, we have presented a straightforward machine learning paradigm to recognize ten different Bento (lunch-box) packaging activities in real-time world. Unlike other skeleton-based human activity recognition domain, it is a very challenging task due to the absence of lower-body marker information. Under these circumstances, we have provided an in-depth statistical analysis of different Bento packaging activities. After feature extraction process, we have used several machine algorithms and obtained best results in random forest classifier using hyperparameter tuning. We have achieved 64.9% validation accuracy using leave-one-out method.