主催: The Japan Society of Mechanical Engineers
会議名: スポーツ工学・ヒューマンダイナミクス2016
開催日: 2016/11/09 - 2016/11/11
As the expectations of wearable technologies for personalised health and lifestyle continues to expand, it becomes increasingly important to reduce the power demands placed on the hardware where possible. For applications such as activity recognition, particularly in fitness, there are a number of options available for identifying and classifying these activities but it is not always clear which particular method should be implemented. This paper explores a variety of classification models provided by the MATLAB numerical analysis software and describes a training and testing outline using a previously validated dataset. Each model is trained prior to testing with a consistent training accuracy of > 99% for the given data. The results showed that for the same test data, decision tree models outperformed discriminant analysis, k-nearest-neighbour, support vector machine by a factor of 1.91, 8.37, and 36.15 respectively. Finally, the considerations of appropriate classification model selection are discussed in the context of light-weight, low power wearable devices where processing hardware and battery limitations are key factors.