Robust Low-Latency Indoor Localization Using Bluetooth Low Energy

In recent years, indoor positioning technology has been a hot spot, especially Bluetooth Low Energy (BLE)–based fingerprint positioning because it can be implemented on smart phones and has high accuracy. In BLE-based fingerprint positioning, accuracy and delay have always been huge challenges and are both closely related to the sampling time of each position estimate (the window size). In this paper, we analyze how the window size affects the positioning accuracy and the Cramer-Rao lower bound (CRLB) of localization using the received signal strength (RSS). Based on the relation between the window size and the standard deviation of the average RSSs between windows, an enhanced CRLB model is proposed by rewriting the parameters in the existing CRLB model to make it related to the window size. The proposed CRLB model related to window size is verified in two indoor scenarios with different typologies for the beacons and different packet transmission frequency. The enhanced CRLB model provides a theoretical basis for analyzing the relation between delay and accuracy and provides a reference for the decision about the window size in the positioning process. In addition, based on the enhanced CRLB model and our analysis, a scheme for generating a robust fingerprint vector is proposed for RSS instability and uncaught beacons’ RSSs caused by the reduced window size. Because with a small number of samples it is difficult to smooth the fluctuation of the RSS, a low complexity weighted fingerprint construction (WFC) method is proposed. WFC uses the weighted average of the RSS data in the previous window fingerprint vector and the samples in the current window to generate the fingerprint vector of the current window. This weight is scaled by a limit factor related to their difference. The uncaught beacons’ RSSs are estimated by referring to the fingerprints in the fingerprints database at the previous location. In the one-site test, the k-nearest neighbor (KNN) algorithm is used as the classifier. The results show that the proposed scheme outperforms the traditional mean fingerprint vector by about 24.5% in terms of positioning accuracy under low latency (the window size is 0.5 s).