Lunch-Box Preparation Activity Understanding from Motion Capture Data Using Handcrafted Features

Japanese packed-meal or Bento preparation or packaging automated by a robot is a recent challenge. It is one of the latest cementations in human activity recognition systems. It shows contingency on physical gesture recognition techniques using some sensor-based datasets, especially some motion capture data or skeleton data. To get a firm grip on the recognition, we (Team Boson Kona) implement a method dealing with some handcrafted features gained from the filtered data. We employ several pre-processing steps to handle any inconsistency of the dataset: after cleaning the data, we use filtering and windowing techniques to process the raw data. Then we figure out multiple statistical and geometrical features from joints-based handcrafted features. After considering the most significant features for model training, we try out support vector machine classifier, random forest classifier and extra tree classifier. After applying all these methodologies, we obtain maximum accuracy of 64.6% using the extra tree classifier.