3.2. Positioning and Tracking 41ture model. In both cases, the transition probabilities p( s| s) are circular bended onedimensional Gaussians with mean 2. 5 and variance 0. 3 representing average move-ment speed of 2. 5 m per sampling step. The problem of sample impoverishment onthe particles is not discussed.An evaluation is conducted on synthetic data from the WAF model. The scene is anoffice style floor with an area of around 1200 m2. The WAF model is configured withattenuation coefficients for three different materials. 6 APs are evenly distributed.The emission probabilities are used to generate measurements for one reference pathwith length 25 m that visits three rooms in the scene. The reported offline LE is lessthan 0. 8 m.On the empirical radio propagation model an evaluation with real measurements wasconducted. 14 APs were evenly distributed in the same scene. On the same path asabove, the reported averaged offline LE is less than 1 m. In both experiments 500particles were used.The last investigated implementation of a PF which uses RSSI readings can befound in[5]. The propagation model is an empirical one and is described in[6]. Theweighting is based on emission probabilities that are Gaussians with a variance of5 dBm. The indoor environment is described in a 2D-map that is used to constrainthe movements during the particle sampling process. The transition probabilitiesare further modified in crossway zones to reflect the higher probability to turn indirection. Sample impoverishment is detected by a drop of the combined weights ofthe particles and is compensated by feeding new uniformly distributed particles intothe system.The evaluation is conducted on a rectangular indoor office scene with an area of1600 m2. An averaged LE of 1. 9 m is reached but the number of APs and the lengthof the path were unfortunately not further defined.3.2.3 Nearest Neighboor based ApproachesFor completeness, an overview of localization systems that ignore the sequentialnature of the tracking problem or that are simply only designed for the positioningproblem, will be given. These systems use the Mean Squared Error(MSE) as thedistance measure between the received RSSI vector and the stored RSSI values fora location. The location with the minimum MSE is selected as the most probablelocation. Therefore, the employed technique is labeled Least Mean Squared Error(LMSE) and was also derived in the background chapter 2.7.A problem in the approach of LMSE is given by the result, that largely differentlocations can have similar MSEs. As a reaction to this problem, enhanced LMSE-based methods were developed. One of them is described in[18] as the so calledcloseness elimination scheme. Another in[17] where the locations at the top smallestMSEs are combined.The system described in[8] evaluates a basic LMSE driven localizer that is based ona raytracer generated progation model. The raytracer is configured with differentmaterial coefficients that are taken directly from the literature. By using materialparameters from[30] the performance of the system is given by a RMS-LE of 2. 28 m