40 3. Related WorkA PF for pedestrians, targeted to support rescue operations, that uses only footstepsensor reading is presented in[29]. The sensor reports a step length, fused from anaccelerometer and a step heading that is derived from a compass. The environment is2D based and constrained by information from 2D indoor/outdoor floor plans. Theweighting of the particles is handled by a simple zero weight if the particle has crosseda wall and constant otherwise. The transition probabilities are derived from the stepheading and step length information and are combined with white Gaussian noise.Therefore, a directional sampling of new particles, comparable to dead reckoningconstraints, is employed. The PF was enhanced by an interesting approach to sampleimpoverishment problem. The basic idea is to track the history of the particles bystoring links between the new and the source particles. Therefore, each generationof particles is stored as well. Now, if the degradation of the current generation isdetected by a drop in the combined weight of all particles, the particles are tracedback into the older sample generations and the source particles of the currently"bad" particles are replaced by new samples. From these"corrected" generations thesampling process is restarted.The evaluation was conducted on an 2700 m2indoor/otdoor scene. A tablet PC,processing the information of XSens MTi motion sensor, was used as the mobiledevice. Map constraints from a detailed 2D-map were available. 2000 Particles areused. The averaged LE is given by 1. 5 m for the basic PF implementation and an evenbetter 1. 3 m for the variant that escapes sample impoverishment with backtrackingAnother PF, that uses the same footstep sensor as the major information source,is presented in[31]. The system does also employ RSSI measurements but, theseare only used for the initialization of the particles. The particles are not uniformlydisplaced over the environment, they are constrained by a rough approximation ofthe probable localization area. The used radio propagation model for this step isa simple path-loss model without wall attenuation. The system has a notion ofthe 3D-space for multi-level building by connecting the individual 2D floor maps atjunction points and enriching the 2D map polygons with height information. This2.5D-map is used for rejecting trajectories that lead through impassable regions.The sample spawning is driven by transition probabilities that are modelled similarto[29]. The weighting of particles is also assumed to be zero if a wall is crossed. If itis a valid particle, the weight will be determined by height information for the 2.5Dmap and the height change that is additionally extracted from the footstep sensor.The transitions are modelled equally to[31] and derive their parameters from thestep length/heading information.The evaluation was conducted on a three floor office style building with a total areaof 8725 m2that is comparable to the UMIC scene. A hip mounted PC processes thefootstep data of a XSens Mtx IMU sensor. Over 6 walks with a duration of 16 mineach, a very accurate RMS-LE of less than 0. 7 m is reported.In[28], three different nonlinear filters: Fourier density approximation, a gridbasedfilter and a PF are compared on an analytical and an empirical radio propagationmodel for 2D scenes. The former analytical model is a multi material WAF modeland the latter a manual collected RSSI database. In the analytical case, the weight-ing of a particle is derived from the output of the WAF model with noise from atwo-components Gaussian mixture with the two means of- 7. 5 dBm and+7. 5 dBm.In the empirical case, the stored RSSI are interpolated and used as means for the mix-