Continuous Action Counting And Recognition Based On CSI

Human continuous activity recognition plays an increasingly important role in many fields,such as smart home,somatic games,and health care.Nowadays wireless sensing technology is widely applied in passive(unobtrusive)human continuous activity recognition.Compared with positive detection,passive detection doesn't require users to carry devices and is of less Obstruction.Compared with the vision-based and optics-based activity recognition technology,Wi Fi-based recognition technology isn't constrained by light conditions,and it will not cause privacy issues.Wi Fi-based activity recognition has been proposed,which mostly relies on Received Signal Strength(RSS)? software definition radio(SDR)and Software Defined Radio(SDR).The recognition effect of RSS is highly sensitive to environment variation,while SDR requires dedicated hardware to transmit and receive modulated signals.Compared with RSS and SDR,CSI has gradually become the main trend of the study of Wi Fi-based activity recognition,which contains fine-grained information of small-scale fading and multipath effects caused by activity.In particular,most of the proposed CSI activity recognition methods focus on solving the problems of how to recognize more activities,and how to improve the accuracy of activity recognition.However,there is no work attempting to achieve both the goals of activity recognition and activity counting simultaneously.Most existing CSI activity recognition methods only output the recognition results at present without any temporal/evolutionary details.Aiming to realize a method which can count and recognize human action simultaneously,we conduct research on CSI data sampling and preprocessing,action detection and counting,action recognition.The main contributions are described as follows:(1)In the data sampling and preprocessing part,Wi-ACR uses commercial Wi Fi equipment and an experimental platform built on Linux csitool based on 802.11 n to collect CSI data that changes due to human actions.Then,Wi-ACR applies Hampel filter to removes the impracticable outliers of each subcarrier.Squat and walk have a relatively large range of action.The frequency of CSI amplitude fluctuation is within the 30 Hz ? 60 Hz band.Wi-ACR passes signal to a Butterworth low-pass filter to remove the high-frequency noise,and finally standardizes the data.(2)In the action counting part,we propose a novel action detection method and action counting method,which can detect the start time and end time of each actions and count the number of actions.First of all,Wi-ACR eliminates the “silent” part of CSI stream,which does not include any action,and determine the start time and end time of each period of actions,which is referred to as the coarse-grained detection of activity.It involves sliding window,PCA algorithm and calculation of activity indicators.We further employ a peak-finding algorithm to determine the start time and end time of each action,which is referred to as the fine-grained detection of activity.This algorithm works under the assumption that the beginning of a new action(possibly while other actions are still underway)will cause a spike in the indicator of activity.Wi-ACR counts the number of peaks,which also represents the number of actions.(3)In the action recognition part,Wi-ACR takes two different action recognition models to recognize action.One is the Waveform-feature-based action recognition model,which employs Discrete Wavelet Transformation(DWT)to extract waveform features to analyze correlation of action waveforms and perform best-fit matching based on dynamic time warping(DTW).Finally,it recognizes the action of each action period by KNN.The other is the Statistical-feature-based action recognition model,which extracts statistical features from waveform,then inputs them to KNN classifiers and SVM classifiers to obtain action recognition results.All the above work contributes to a CSI based human action counting and recognition method named Wi-ACR.The experimental results show that Wi-ACR can achieve the average action counting accuracy of 95% and recognition accuracy of 90% with these two recognition models,for5 volunteers in the indoor scenarios with two types of actions(i.e.squat and walk)occurring simultaneously.Compared with existing action counting and recognition methods,Wi-ACR can count and recognize action simultaneously,which achieves high counting and recognition accuracy.