紀要論文

ウェアラブルデバイスにおけるデータ品質が行動認識精度へ及ぼす影響

The Impact of Wearable Data Quality on Activity Recognition Accuracy

jpn

24321192

https://doi.org/10.15002/00021950

法政大学大学院情報科学研究科

法政大学大学院紀要. 情報科学研究科編

14

1

6

2019-03-31

Version of Record

This research is to examine the data quality from wearable devices and the possible impact on activity recognition accuracy. In our experiments, multiple wearables were wearied in different parts of a participant for capturing motion data in daily activities such as walking, jogging, stair climbing, stair descending, and sitting. The captured data has been brought about some quality problem when a wearable is misplaced in an improper position, and temporal problems in captured data from one and multiple devices. For a data stream from a single device, the time intervals between samples change greatly around a sampling frequency specified in a sensing program. When multiple devices are used, there exist temporal differences among multi data streams. After raw data pre-processing for certain data quality improvement by means of filtration, interpolation and normalization, we conduct a set of activity recognition with typical machine learning algorithms such as SVM, K-NN, and decision tree. The results have shown that the temporal deviations in sensed data affect the recognition accuracy. The position shift of a device can affect much on the recognition accuracy, but this affect may be mitigated when using multiple devices together for recognition. Furthermore, it is found that the proper selection of feature quantity can improve the recognition accuracy. Among the three methods, SVM doesn’t show enough flexible for testing various data situations.

Article

AA12746425