Regardless of Sensor Location Human Activity Recognition Based on Transformer

In recent years, research on human activity recognition (HAR) using MEMS sensors has been widely studied. However, sensors placed on different parts of the body exhibited significant differences in recognition rates. This research addressed this issue by proposing a robust HAR method that was independent of sensor location. In response to the problems of noise in sensor data, unclear features, and the reliance on traditional feature extraction on experience, this paper used Transformer models to experiment and verify the UCI machine learning repository, daily activities, and sports activity datasets. The experimental results demonstrated that regardless of whether the sensor location was restricted when using a Transformer model for classification in human activity recognition tasks with single-device multi-sensor scenarios, the accuracy rate reached above 98%. Moreover, by comparing the recognition rates of manually extracted features and Transformer self-learned extracted features for human activity recognition classification, it was found that the effect of Transformer self-learned extracted features was better than manually extracted features. This provided a certain reference for human activity recognition based on MEMS sensors.