38 3. Related Workis constructed by collecting indoor and outdoor measurements. Additionally to thestored RSSI information, sensor readings of the acceleration and from a compassare also available to the system. The database with the collected measurements wasinterpolated to obtain the emission probabilities p( x| s) at all states s representing2D locations on a grid. The database contains mean and variance information thatis used to model p( x| s) as Gaussian distributed. The evaluation of this probabilityfor a measurement x contains a smoothing step that is conducted by integrating overthe[- 0. 5 dBm.. 0. 5 dBm] interval around the measured RSSI value. The transitionprobabilities p( s| s) are modelled variably and depend on the sensor readings of theaccelerometer and the compass by using dead reckoning. All states are possible pre-decessor states leading to a computational complexity of T cdotS2during the ViterbiDecoding. Finally it is a first order Markov model, as only the direct predecessorsinfluence the decision process.An evaluation of the performance of the HMM localizer on synthetic data was con-ducted. The synthetic RSSI data is generated by assuming a radio propagationmodel with simple log-normal fading and combining it with white Gaussian noise of6 dBm. The scene is a 2500 m2square area with full LOS conditions. Furthermore,additional synthetic noisy accelerometer and compass information was added to thesetup. The state space contains 2500 states with edge size 1 mx 1 m. 9 APs wereevenly distributed. This environment leads to an offline LE of around 2 m for a notexactly specified sequence of 105measurements.Additionally an evaluation on two real world scenes was conducted. The propagationmodel is obtained from previously collected measurements in both cases. Both caseshave an unknown number of APs and an unknown rasterization for the state space.The measurements of the three mentioned sensors were taken with a HTC HeroAndroid smartphone. The first scene is a Christmas market with area of around2500 m2at a time of the day where nearly no visitors lead to destructive body shadow.Due to the many small tourist shops at the market, there are high NLOS conditions.For this outdoor scene an offline averaged LE was evaluated on an unclear numberof measured signals. The second scene is a combination of an indoor and outdoorarea of around 1200 m2. An offline LE of around 2 m for the indoor, and 5 m for theoutdoor part is determined. The combined LE on all 2300 measurements is givenby 4 m.Another HMM based localization system is presented in[27]. An analytical radiopropagation model was employed that is based on a multi material WAF model. Theemission probabilities p( x| s) are obtained from the WAF model that is enhanced byassuming that attenuation factors are modelled as Gaussians. The noise at thereceiver side is also modelled as a Gaussian. Both PDFs are combined to form thefinal p( x| s). As in[25], the computation of p( x| s) for an instance of the RSSI vectorx is done by integrating over the[ x- 0. 5 dBm..x+ 0. 5 dBm] interval. The transitionprobabilities p( s| s) are modelled by assuming two modes of operation. The firstmode is the movement mode which is detected by evaluating the variance from thelast 10 measurements. If the variance exceeds a preconfigured threshold, this modeis assumed to be given. The other mode can be evidently called non-movementmode. The mode configures the maximum walking speed that is used to determinethe possible transition origins s for the current state s. As p( s| s) is modelled as arasterized Gaussian, the maximum walking speed is projected onto the variance. All