Currently, WiFi-based user continuous action counting and recognition is limited to a single person. Being able to continuously analyze and record the different actions of multiple users in a device-free scene is one of the most challenging job to date. In this paper, we present a new WiFi-based multi-user action recognition system, called IMar, which can achieve decomposition of multi-user action information and retain action features as much as possible, i.e., simultaneously recognize and count the continuous and different actions of multiple people. Our main technical route is to design a Dynamic Propagation Delay Threshold Sanitization (DPDTS) algorithm to retain the path information that only passes through the target user body, in order to reduce the multipath effect and make the data as pure as possible, and then model the amplitude relationship of the multi-person action scene. After acquiring the individual data according to the model and tensor decomposition, we propose a Multiplayer Action Amplitude Decomposition and Completion (MAADC) algorithm to obtain more informative data for individual continuous action. Moreover, the single-person data of subcarrier-level obtained by tensor decomposition is extended to the data of stream-level, which brings great convenience to the single-person action recognition. Experimental results show that IMar can work with up to 6 people. The average recognition accuracy and counting accuracy are 78% and 91% respectively in the experimental group of continuous and natural actions by multiple users, and the average recognition accuracy of the experiments in all cases is 83.7%.
Read full abstract