HandFi: WiFi Sensing based Hand Gesture Recognition using Channel State Information

Abstract

Gesture recognition using WiFi sensing is an emerging and innovative approach that leverages the variations in WiFi signals caused by human hand movements to enable non-contact and intuitive human-computer interaction. The identification of the hand used to perform a gesture significantly impacts the channel state information (CSI) characteristics. It plays a crucial role in data analysis and interpretation. This paper presents a model for classifying the gesture and the hand with which it is performed. We use an Intel 5300 network interface card as a receiver and a commodity WiFi router to collect the CSI values. The CSI values are collected for six left-hand and six right-hand gestures (push left, push right, increase, decrease, stop, and next). The collected CSI values are pre-processed to find the CSI ratio among the sub-carriers. Then four different methods including proposed method are used for classifying the hand and the gesture. The four methods are (1) use of the Hampel filter, (2) use of CSI ratio only, (3) use of correlation matrix along with CSI ratio, and (4) proposed CSI ratio and interquartile distance based method. Then various machine learning techniques such as KNN, Decision trees, and SVM are used to classify the hand (right or left) used for signaling the gestures and the type of gestures. It is found that using the top 30 inter-quartile distance of the CSI-ratios corresponding to the CSI values of the sub-carriers while preserving their positions as the feature vector, the KNN machine learning model can classify the hand and gesture with approximately 99% accuracy for test data. Similar performance is also observed for other performance evaluation parameters like precision and recall.