Dense Labeling of Human Activity Recognition Using a U-Net++ Network Based on Inertial Sensor Data
Hanchi Wen, Hanzhi Zhang, Zhanzhi Lei, Xiao Liang · 2022 15th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI) · 2022
Human activity recognition (HAR) is used to monitor and identify human activities intelligently and shows a good perspective in many applications. Traditional HAR methods in previous works confront with the multi-class-window problem, so it is difficult to achieve a high time-resolution in activity recognition. Herein, we present a dense labeling method based on a U-Net++ network to overcome the above problem. Inertial sensor data is fed into the network directly to realize activity classification of each sampling point. The architecture of U-Net++ adopts dense skip pathway and deep supervision, which makes the network pruning flexible. We tested our method on the dataset collected in our laboratory and a public dataset, WISDM. The results indicated that our method had a better performance over compared methods. Moreover, network pruning experiments were carried out. The results showed that in the laboratory dataset, the accuracy of a pruned network with 0.04 M parameters was more than 99%, and in WISDM, the accuracy of a pruned network with 0.18 M parameters could reach 97.3%.