DU-MD: An Open-Source Human Action Dataset for Ubiquitous Wearable Sensors

• Published in: 2018 Joint 7th International Conference on Informatics, Electronics & Vision (ICIEV) and 2018 2nd International Conference on Imaging, Vision & Pattern Recognition (icIVPR) pp. 567 - 572
• Main Authors: Saha, Swapnil Sayan, Rahman, Shafizur, Rasna, Miftahul Jannat, Mahfuzul Islam, A.K.M., Rahman Ahad, Md. Atiqur
• Format: Conference Proceeding
• Language: English; Japanese
• Published: IEEE 01.06.2018
• Subjects: Accelerometer; Accelerometers; Activities of Daily Living (ADL); ANOVA; Human Action Recognition (HAR); Medical services; Senior citizens; Training; Wearable; Wearable sensors; Wrist
• DOI: 10.1109/ICIEV.2018.8641051

Human Action Recognition (HAR) in healthcare amongst senior citizens focuses on remote surveillance, healthcare monitoring and fall detection. The wearable approach, in particular, wrist-mounted sensors for HAR is most favorable when qualitative characteristics, parameter complexities and market projections are considered. Machine learning models for Activities of Daily Living (ADL) / fall detection require large, hardware-independent and comprehensive ADL datasets exhibiting statistical variance and closeness to real life cases. However, there is a lack of public motion traces filling in all necessary obligations. In this context, the University of Dhaka (DU) Mobility Dataset (MD) was built using 25 subjects (out of 50) with 10 ADL (7 basic ADL and 3 falls), amounting to 2500 (out of a final 5000) training sets using a single wrist-mounted wearable sensor. Some existing public databases have been compared extensively and assembly of the wearable sensor using the recently developed UTokyo Trillion Node Engine Project is illustrated. Statistical tests have been carried out to ensure diversity whilst accuracy of the dataset using existing statistical mechanisms have been acknowledged. Promising diversity and accuracy make this dataset suitable for use in wrist-mounted healthcare monitoring systems.