Research On Wi-Fi Acquisition Based On Smartphones And Deep Learning For Activity Recognition

The current main activity recognition system includes contact-based and contactless-based systems.Compared to contact-based approaches like,wearable devices,contactless activity recognition has the advantage of non-intrusiveness,and long-term sensing.When a device is no longer worn on the body,its sensing capabilities are remarkably degraded.Be similar with wearable devices,sensor-based devices are also widely used.However,most existing systems based on sensors can only work normally in some fixed locations where sensors have been installed.There are mainly two contactless activity recognition approaches,which are video-based and Radio Frequency(RF)-based.Video-based activity recognition approaches can directly use large-scale images or videos for data analysis,with limitations on hardware and application conditions.What is more,the cost of hardware deployment and expansion is very high.Methods based on RF include continuous wave radar,ZigBee,Wi-Fi and so on.The Wi-Fi-based contactless activity recognition approaches are cost-effective since they are already widely deployed in the public infrastructures.In this article,the system is proposed that utilizes Wi-Fi signal for the recognition of activity by machine learning algorithms and deep learning model.It has two advantages:one is that it is contactless and the other is that the hardware facilities are more universal.The system uses the classifiers and Convolutional Neural Networks(CNNs)to train the relation model between the Wi-Fi signals and the activities.We took about one month to obtain continuous data of Wi-Fi device in the same conditions and analyzed the challenge and lessons of the case studies of building a model that can address this nonlinear relationship problem.We made groups experiments to analyze RSSI and made comparison tests base on Channel State Information(CSI).The result of experiments shows the potential of utilizing Wi-Fi signals of the environment for ambient stimuli by human via smartphones.As a preliminary exploration for activity classification,the system utilizing RSSI achieved a mean average precision of 95%based on four Machine Learning methods and up to 97.7%based on CNNs.