Abstract:
The development of location-based service (LBS) in outdoor environments relies on Global Positioning System (GPS) technology to determine the location. However, for indoor LBS, GPS is not reliable because it has low accuracy for indoor environments. Therefore, it is necessary to have an indoor positioning system (IPS) for indoor LBS. Wi-Fi-based IPS research is growing along with the many uses and availability of Wi-Fi. For static indoor environments, fingerprinting techniques have better accuracy than distance-based approaches such as trilateration and min-max. The fingerprint technique can also be applied by utilizing simple and straightforward parameters, i.e., received signal strength indicator (RSSI). Some of the fingerprint technique's challenges are the length of time, effort, and high cost of collecting the database. In this study, we apply the interpolation techniques to reduce time and effort in collecting fingerprint data. The interpolation used is the Neville interpolation and Bilinear interpolation. Comparing the positioning results between the classical pattern matching algorithm, minimum Euclidean distance (MED), and modern machine learning-based, the Random Forest algorithm, is discussed in detail. We conducted a measurement campaign in an unideal indoor environment to see how far our proposed method can still handle the fluctuated values of RSSI. From several measurements and scenarios presented in this study, the MED algorithm is still better in accuracy and precision than the random forest algorithm. However, in almost all scenarios, the Random Forest can better perform MED in terms of decreasing the maximum estimated error. The accuracy and precision between MED and random forest are up to 0.5 meters, and the precision difference is up to 20%. Performance improvements due to using database interpolation range from 3% to 30%. The database from the interpolation results is also acceptable in the performance metric of the positioning system. However, using the database from the actual position is better and outperformed the database from interpolation results. The low similarity of actual database and database synthesis from interpolation is due to the non-linear and fluctuated RSSI values in our measurements due to the unstable and time-varying effects.
