MCSM-Wri: A Small-Scale Motion RecognitionMethod Using WiFi Based on Multi-ScaleConvolutional Neural Network.

Abstract

Small-scale motion recognition has received wide attention recently with the development of environmental perception technology based on WiFi, and some state-of-the-art techniques have emerged. The wide application of small-scale motion recognition has aroused people’s concern. Handwritten letter is a kind of small scale motion, and the recognition for small-scale motion based on WiFi has two characteristics. Small-scale action has little impact on WiFi signals changes in the environment. The writing trajectories of certain uppercase letters are the same as the writing trajectories of their corresponding lowercase letters, but they are different in size. These characteristics bring challenges to small-scale motion recognition. The system for recognizing small-scale motion in multiple classes with high accuracy urgently needs to be studied. Therefore, we propose MCSM-Wri, a device-free handwritten letter recognition system using WiFi, which leverages channel state information (CSI) values extracted from WiFi packets to recognize handwritten letters, including uppercase letters and lowercase letters. Firstly, we conducted data preproccessing to provide more abundant information for recognition. Secondly, we proposed a ten-layers convolutional neural network (CNN) to solve the problem of the poor recognition due to small impact of small-scale actions on environmental changes, and it also can solve the problem of identifying actions with the same trajectory and different sizes by virtue of its multi-scale characteristics. Finally, we collected 6240 instances for 52 kinds of handwritten letters from 6 volunteers. There are 3120 instances from the lab and 3120 instances are from the utility room. Using 10-fold cross-validation, the accuracy of MCSM-Wri is 95.31%, 96.68%, and 97.70% for the lab, the utility room, and the lab+utility room, respectively. Compared with Wi-Wri and SignFi, we increased the accuracy from 8.96% to 18.13% for recognizing handwritten letters.