Empirical Performance Evaluation of WIFI Fingerprinting Algorithms for Indoor Localization

In this paper, we empirically evaluate the performance of WIFI fingerprinting technique in a realistic indoor environment with a total coverage area of 740m2. We also develop two localization algorithms, called PA-1 and PA-2. These algorithms are based on probabilistic framework assuming Gaussian PDF of error. PA-1 is more complex and is designed for the case when direction information is not measured in the online phase, while PA-2 is less complex and it can leverage direction information. Through manual calibration, a detailed fingerprinting database containing Received Signal Strength (RSS) values from 6 APs is developed at a grid spacing of 1m using three different smartphones and four different directions. In our experiments, we observe up to 30dBm temporal variations in RSS values at the same grid point. Furthermore, at 1m grid spacing, it becomes difficult to differentiate between RSS variations at neighboring grid points and RSS variations at the same grid point due to hardware and direction differences. We compare the localization accuracy of our algorithms with K-Nearest Neighbors (KNN) algorithm. At 1m grid spacing, simple deterministic KNN algorithm outperforms more sophisticated probabilistic algorithms. However, when we increase the distance between neighboring grid points and make the grid spacing sparse (>1m), probabilistic algorithms perform better and labor-intensiveness of WIFI fingerprinting also decreases. We conclude that dense grid spacing is not always optimal and several performance-complexity tradeoffs can be achieved by carefully analyzing local RSS variations in the indoor environment.