A Multiscale Windows Deep Learning Aproach For Sensor-Based Human Activity Recogniton
Furong Duan, Tao Zhu, Jinqiang Wang, Zhenyu Liu · 2022
Deep learning for sensor-based Human activity recognition (HAR) is a popular research topic and can be widely used in many real-life applications. By far, most HAR deep learning researches use a fixed-size window to segment sensor data streams into samples of the same length, then perform activity recognition. However, in the real world, the length of activity sequences could be different, fixing window size can not segment sensor data stream very well and could impair recognition performance. In this paper, we propose a multi-scale window deep learning approach for HAR, which generates multiple windows of different lengths centered on each unit of the feature sequence. We conduct experiments on three popular benchmark datasets. The results show that our proposed method outperforms the state-of-the-art performance.