Identifying Children's Activities

Abstract

Dutch children achieve an insufficient amount of physical activity. For that reason, Delft University of Technology is looking to develop a wearable that contributes to motivating children to move more frequently and more intensely while collecting data on the movements made by its wearer. In this thesis it was investigated whether it would be possible to identify the activities performed by free-living children and what type of data from which body placement should be obtained in order to do so. Due to restrictions resulting from COVID-19, triaxial accelerometer- and gyroscope measurements of typical child play activities were carried out on various body parts of eight adults. A Long-Short Term Memory algorithm was applied to short sequential sections of summarized accelerometer data. This algorithm gives a prediction of the activity executed by the wearer of a triaxial accelerometer for every 10 seconds in time. The effects of the wear-site (left wrist, right wrist, right hip, left ankle or right ankle), type of accelerometer (low noise or wide range) and epoch length (0.25, 0.33, 0.5 or 1 second) were studied. The shortest epoch length was found to result in the most precise predictions per activity and the highest overall accuracy for classifying the activities. The hip and right wrist placement perform better than the other locations. A wrist placement is favored over the hip because a heart rate sensor can be added to the former. Measuring the heart rate in combination with classifying the activities performed gives insight in both intensity and variety of movements made by children. To further increase the performance of these classifications, predictions made with a score below a certain threshold, 0.775 in this thesis, can be excluded. This will decrease the amount of classifications made but it improves the accuracy as well as causing the precision with which each activity is recognized to rise. Without excluding results with a score below this threshold value, classifications of low noise 0.25 second epoch data from the right wrists have an accuracy of 74.8% and a precision per activity of 54.5% - 82.1%. Removing the more uncertain predictions yields an accuracy of 84.0% and a precision between 57.8% - 91.5%. Clustering specific activities, such as sitting and lying down, increases the precision considerably. The accuracy and precisions when applying sitting and lying down as one cluster in combination with removing the uncertain predictions for the low noise sensor become 85.4% and 70.1% - 92.2%. Before this algorithm can be successfully implemented in combination with the intended wearable, the precision with which each individual activity is identified should be <90%. To transfer the collected accelerometer- and heart rate data to a smartphone, it is recommended to use NFC technology. Through a smartphone application it can then be made available for research purposes and the children can get feedback on the variety and quantity of their physical activity of the past two days. A battery that is appropriately sized for a child's wrist wearable will only be able to power the wearable for a few days. This is an insufficient battery life for the intended use, so the battery has to be charged. NFC technology additionally offers energy harvesting capabilities, making it possible to wirelessly charge the wearable via a smartphone. From this exploratory research it can be concluded that it will be challenging to develop a low-cost wearable that can identify activities and measure how frequently free-living children are physically active. This is the case for both software, the algorithm that predicts which activities were performed, and hardware, where the most pressing challenge is the battery life. However, it is believed that with more extensive research it is possible to create a fully operational wearable.