A Self-learning Fingerprinting Matching Algorithm for Indoor Wi-Fi Positioning
Guenther Retscher and Anton Stangl, Vienna University of Technology, Austria
The use of Wi-Fi for positioning has become popular as it is the most prominent signal-of-opportunity that is available nowadays in many buildings and public spaces. In general, Wi-Fi positioning is mainly based on the measurement of the received signal strength indicator (RSSI) to the surrounding wireless access points (APs). Common positioning methods include location fingerprinting and trilateration. The main disadvantage of trilateration is, on the one hand, that theoretical path loss models for RSSI to range conversion have to be employed to be able to calculate a position fix. Fingerprinting, on the other hand, requires usually high workloads in system training to achieve acceptable positioning accuracies as RSSI scans have to be carried out in the training phase on a large number of known reference points placed throughout the area of interest. Actually, continuous system training would be required to accommodate for the occurring high spatial and temporal RSSI fluctuations as they are degrading the achievable positioning accuracies significantly. In the new approach a self-learning strategy for training is developed which aims to reduce the effect of Wi-Fi signal variations and noise on the positioning result while reducing the required workload for system training at the beginning. The core of this straightforward approach is that only differences of the absolute RSSI values between measurements from different APs are used in a simple matching algorithm. For that purpose, the highest and most stable RSSI measurements from six APs are selected from all visible APs in the surrounding environment. As usually in fingerprinting the current RSSI observations are then compared to the values in a fingerprinting database obtained in the continuous system training. The minimum sum of the six obtained RSSI differences between the database entries and the current measurements results then in the most likely match. In other words, the total sum yields the location of the mobile device with the highest probability. If for an AP out of the six no RSSI above a certain threshold set to -75dBm is obtained then another AP with the next highest RSSI value is included into the database search. Furthermore, RSSI jumps and large signal variations between consecutive measurement epochs above a threshold of 10dBm are discarded in the matching approach. Thus, their influence is significantly reduced as only Wi-Fi scans are used which have been already verified. Current verified RSSI observations are then always used to update the fingerprinting database to be able to achieve a continuous training. The approach also does not require previous knowledge of the building, such as a building model, as due to the continuous database update self-learning of the environment is achieved. A client processing solution is employed in our approach where the functionality of a previously developed server-client solution is embedded into the smartphone. This has the advantage that data protection issues and the privacy of the user are always guaranteed.
In the practical evaluation, the suitability and performance of the self-learning matching approach is investigated. Experiments were conducted in a multi-storey office building environment and a two-storey flat with roof top terrace in a residential building. The results showed that a correct match of over 98% is obtained to locate the mobile device in a certain room. Further improvement of the matching approach is expected if an additional weighting of the measured RSSI values depending on their significance is applied. This is currently the task of further investigations. Furthermore, the integration of the inertial smartphone sensors, such as accelerometers, gyroscope, digital compass and barometric pressure sensor, has the potential for further improvement of the model due to the possibility of movement pattern recognition of the user. Sudden position changes between neighbouring rooms are then only considered in the solution if steps are detected from the accelerometer measurements while the user walks with his smartphone. In this context, a resulting change between rooms is also only accepted if they are inter-connected. In addition, the change of floor can be verified using the observations of the barometric pressure sensor. Moreover, different locations and orientations of the smartphone are considered in the solution, such as held in the hand by the user or put into the trousers pocket.