Identification of Food Packaging Activity Using MoCap Sensor Data

The automation system has brought a revolutionary change in our lives. Food packaging activity recognition can add a new dimension to industrial automation systems. However, it is challenging to identify the packaging activities using only skeleton data of the upper body due to the similarities between the activities and subject-dependent results. Bento Packaging Activity Recognition Challenge 2021 provides us with a dataset of ten different activities performed during Bento box packaging in a laboratory using MoCap (motion capture) sensors. Bento box is a single-serving packed meal that is very popular in Japanese cuisine. In this paper, we develop methods using the classical machine learning approach, as the given dataset is small compared to other skeleton datasets. After preprocessing, we extract different hand-crafted features and train different models like extremely randomized trees, random forest, and XGBoost classifiers and select the best model based on cross-validation score. Then, we explore different combinations of features and use the best combination of features for prediction. By applying our methodology, we achieve 64% accuracy and 53.66% average accuracy in tenfold cross-validation and leave-one-subject-out cross-validation, respectively.