Exploiting temporal dependency of RSS data with deep learning for IoT-oriented wireless indoor localization

With the ubiquitous demand for indoor location-based services and the pervasive deployment of Wi-Fi hotspots, wireless indoor localization has been widely studied by utilizing various Wi-F1 signal measurements. Most existing schemes leverage the Received Signal Strength (RSS) of Wi-Fi to conduct cost-efficient indoor localization. However, the RSS data are not only prone to multi-path effects, but also sensitive to time-varying environmental dynamics, making it quite daunting to achieve robust indoor localization. In contrast to existing solutions that focus on spatial features of RSS, in this article, we exploit the temporal dependency of RSS time-series data by integrating the Kalman filter with deep neural networks. In particular, to tame time-varying noises and preserve valuable temporal features in RSS measurements, we propose a time-varying RSS filtering algorithm based on the Kalman filter and a refined post-processing module. Moreover, a deep learning model based on deep neural network (DNN) is further utilized for effective feature extraction on one-dimension RSS fingerprints. The experiment results show that the proposed Kalman-DNN model improves at least 25% localization accuracy in comparison with conventional DNN model. Furthermore, with the localization time as 0.02 ms, the Kalman-DNN model outperforms the Kalman-CNN model in localization accuracy by at least 10%.