Research on Feature Selection of Human Physical Activity Recognition for IOT Wearable Devices

Using wearable devices to obtain daily physical activity data makes great contributions to health monitoring and evaluation of human. Wearable devices such as bracelets and watches have also developed rapidly in recent years. However, wearable devices cannot perform complex calculations or transmit large amounts of data on a chip because of limited computing power and strict power control. Therefore, it is necessary to extract the most effective features with the least amount of computation. In this paper, physical activity data is obtained by attaching sensor nodes to multiple parts of the human body, then time-domain features and frequency-domain features are extracted from sliding windows of different lengths to study the implementation scheme of efficient recognition of human activity. It is found that the recognition effect is better when two sensor nodes are tied to trunk and left leg respectively and 250 sampling points are selected as the sliding window length. Experimental results show that compared with using all features directly, the time-domain feature selected by Genetic Algorithm is the most effective feature. The energy consumption of data transmission is reduced by 40.5%, the accuracy increased from 95.2% to 96.67%, and it also has superior real-time performance.