Ensemble Classifier Managing Uncertainty in Accelerometer Data within Human Activity Recognition Systems

Human activity recognition (HAR) is a field of study that aims to recognize activities from data acquired by video or wearable sensors. The biggest health study in Norway, HUNT, has recently ended it’s fourth study where 38 756 participants have recorded activity data while wearing three-axis accelerometer on their thigh and back. HAR systems often require all sensors to be operative and attached to the participant at all times, and shows weaknesses when performing activity recognition, as a lot of misclassifications occur due to sensors lying still after being detached from the subject’s body during activity recording. To make HAR systems more robust against this issue, this thesis researches on a new type of ensemble classifier where a meta classifier predicts sensor no-wear time, eliminates faulty sensor streams and dynamically adjust the LSTM-RNN sensor position specific classification models used, depending on the data available. The developed meta classifier is trained on a new ”Sensor No-Wear Time” dataset that consists of real-world data, and is able to predict sensor no-wear time with 97.2% accuracy and shows promising results towards making more valid contributions towards public health research, as it eliminates up to several days of misclassifications where sensors have been detached. Research done in this thesis shows that individual models for thigh and back are struggling to classify certain static activities. A model for both sensors combined is therefore the best option for activity classification as it achieves an accuracy of 85.1% compared to the existing HAR system’s 76.5%, and outperforms individual models when classifying static activities. Storing classification results for all participants in HUNT requires huge amounts of storage space, and Feather is proving to be the file format that is best suited for storing activity classification results, as the result file size for each participant is reduced from 2.5 GB to 941 KB with a new compression algorithm. This results in a total reduction of 99.96%, as necessary storage space is reduced from 96.89 TB to 0.036469396 TB for all HUNT4 participants.