An improved Wi-Fi sensing-based human activity recognition using multi-stage deep learning model

Abstract

Human activity detection and response in real time is one of the major objectives in many defense applications. Many technologies that include surveillance cameras and wearable devices using sensors were proposed for human activity detection. But, these technologies are not popular due to a lack of convenience, privacy, availability, and affordability. Sensing and recognizing various human activities using wireless signal characteristics have been in the reckoning in recent years. This paper proposes a multi-stage deep learning model consisting of a convolutional neural network and other popular deep neural architectures, namely Alexnet, Googlenet, and Squeezenet, for Wi-Fi sensing-based human activity recognition. Transfer learning is used for recognizing crucial human activities using the channel state information (CSI) of the Wi-Fi signal. The objective is to identify critical human activities and then to minimize the false negatives. In the proposed multi-stage model, a data sample is tested in one network for identifying the critical activity; if classified as negative with predefined confidence, the sample is tested again in the subsequent architectures. A sample is considered negative only if all the stages confirm it as negative. Our proposed multi-stage model can reduce the false negatives with a slight increase in the false positive rate. We have considered the dataset consisting of measured CSI values of three transmitter–receiver pairs for six different activities: fall, wave, jump, walk, clap, and sit. We report the results of the experiments with a focus on reducing false negatives of individual activities during the testing phase.