Bento Packaging Activity Recognition from Motion Capture Data

Human activity recognition (HAR) has been an important research field for more than a decade due to its versatile applications in different area. It has gained significant attention in the health care domain. Although it has similarity with other form of activity recognition, it offers a unique set of challenges. Body movements in a food preparation environment are considerably less than many other activities of interest in real world. In this paper, a comprehensive solution has been demonstrated for the Bento Box Packaging Challenge activity recognition. In this case, we present a well-planned approach to recognize activities during packaging tasks from motion capture data. We use dataset obtained from motion capture system where subjects have 13 markers on their upper-body area and by special use of cameras and body suit. We obtain around 50,000 sample for each of the activities. We reduce the data dimensionality and make the data suitable for the classification purpose by extracting reliable and efficient features. After feature extraction process, three different classifiers, e.g., random forest classifier, extra trees classifier, and gradient boosting classifier are compared to check the result. We conclude that this challenging dataset has been observed to work most efficiently for random forest classifier using hyperparameter tuning.