Unsupervised habitual activity detection in accelerometer data

• Main Author: Domingo, Carolyn C.
• Format: text
• Language: English
• Published: Animo Repository 2015
• Subjects: Human activity recognition; Accelerometers; Computer Sciences
• Online Access: https://animorepository.dlsu.edu.ph/etd_masteral/5058
• Institution: De La Salle University

Activity recognition is an active area of research that involves recognizing the actions and goals of one or more agents from a series of observations. Previous researches have resulted in various successful approaches capable of recognizing common basic activities such as walking, sitting, standing and lying, mostly through supervised learning. However, supervised learning approach would be limited in that it requires labeled data for prior learning. It would be difficult to provide sufficient amounts of labeled data that is representative of freeliving activities. To address these limitations, this research proposed using motif discovery as an unsupervised activity recognition approach. A 3D accelerometer sensor worn on the dominant arm was used to record the user’s movements as they perform activities of daily living (ADL). Three sets of time diaries were then built by different annotators by watching video recording of the session. The collected accelerometer data was processed and discretized in order to perform motif discovery. Habitual activities would be detected by finding motifs, similar repeating subsequences within the discretized sequence. Evaluating the result against the time diaries using average clustering event purity, a score of 44% was reached. Video analysis shows that the activities being detected were simple and low-level, consisting of only a few movements, and were heavily focused on the arm’s movement.