Research Article
LightSeg: An Online and Low-Latency Activity Segmentation Method for Wi-Fi Sensing
@INPROCEEDINGS{10.1007/978-3-031-34776-4_13, author={Liming Chen and Xiaolong Zheng and Leiyang Xu and Liang Liu and Huadong Ma}, title={LightSeg: An Online and Low-Latency Activity Segmentation Method for Wi-Fi Sensing}, proceedings={Mobile and Ubiquitous Systems: Computing, Networking and Services. 19th EAI International Conference, MobiQuitous 2022, Pittsburgh, PA, USA, November 14-17, 2022, Proceedings}, proceedings_a={MOBIQUITOUS}, year={2023}, month={6}, keywords={WiFi sensing CSI Activity segmentation}, doi={10.1007/978-3-031-34776-4_13} }- Liming Chen
Xiaolong Zheng
Leiyang Xu
Liang Liu
Huadong Ma
Year: 2023
LightSeg: An Online and Low-Latency Activity Segmentation Method for Wi-Fi Sensing
MOBIQUITOUS
Springer
DOI: 10.1007/978-3-031-34776-4_13
Abstract
WiFi based activity recognition mainly uses the changes of Channel State Information (CSI) to capture motion occurrence. Extracting correct segments that correspond to activities from CSI series is then a prerequisite for activity recognition. Researchers have designed various segmentation methods, including threshold-based and deep learning-based methods. However, threshold-based methods are highly empirical and the threshold is usually dependent on the application and environment. When dealing with mixed-grained activities, the predefined threshold will fail. On the other hand, deep learning-based methods are impractical for online systems with low-latency demand because of their high overhead. In this paper, we propose LightSeg, an online and low-latency segmentation method leveraging an activity granularity-aware threshold that quickly adjusts itself based on the granularity of the activity in the current detecting window. We propose a threshold post-decision mechanism that detects the end of a segment first and then decides the appropriate threshold based on the most recent activity. By this way, LightSeg automatically adapts to different activity granularity in practice. Compared to existing threshold-based methods, LightSeg greatly reduces the dependence on expertise to decide the threshold. Experimental results show that LightSeg improves the segmentation accuracy by up to 14(\%)compared to the existing threshold-based method and reduces the data processing time by 97(\%)compared to the deep learning-based method.
