OpenPack Challenge 2022 Report: Impact of Data Cleaning and Time Alignment on Activity Recognition

• Published in: IEEE International Conference on Pervasive Computing and Communications workshops (Online) pp. 257 - 258
• Main Author: Matsubayashi, Yusuke
• Format: Conference Proceeding
• Language: English; Japanese
• Published: IEEE 13.03.2023
• Subjects: Cleaning; Conferences; Data collection; Human activity recognition; neural network; Neural networks; Pervasive computing; Training; Training data
• ISSN: 2766-8576
• DOI: 10.1109/PerComWorkshops56833.2023.10150299

This report describes a solution for the OpenPack Challenge 2022 developed by team Malton. Our team mainly focuses on preprocessing of training data for human activity recognition. Because of errors in data collection processes, collected data from subjects can have variety of noises, indicating the importance of data cleaning for the collected data when they are used for training a human activity recognition model. In our solution, we applied two main approaches for the data cleaning. The first approach attempts to remove training data sessions with erroneous data collection procedures. The second approach attempts to align timestamps of different sensors by using binary search. Our experiment revealed that the importance of the preprocessing approaches when we used a standard neural network architecture for activity recognition (DeepConvLSTM). As a result, our solution achieved the F1-score of 91.7%.