4.4. Devices 57The case is detected by a drop of the combined emission probabilities p( x| s) ofall samples by several orders of magnitude. This sum is always available, as it isused for the weight renormalization described in 2.6.2. A possible strategy to resolvethis situation is given by feeding new uniformly distributed particles into the statespace. Another strategy that works well on the presented system is given by simplydeactivating temporarily the geometry constraints. After recovery of p( x| s), theconstraints will then be reactivated.4.3.2.4 Result SequenceThe Particle Filter returns three different result sequences. As for the HMM ViterbiDecoder, the first sequence sT1evaluates all x1Tand represents an offline run. Thissequence is reconstructed after processing all T time frames that uses informationstored in a backtracking table. This table stores for each time frame t and eachstate s the state s that was responsible for spawing s. The best sTis used as thestarting state to reconstruct the sequence by iteratively following the state linkingbackwards.The second sequence s1Tthat is returned, is given by the particles that have thehighest emission probability p( xt| st) for each time frame t. As in the HMM ap-proach, this stands for an online result as it does not depend on the remainingfuture measurements xt+{ 1..T- t}.The third sequence sT1is also an online result but it contains the locations thatare the geometric means of the top N particles ordered by p( xt| st). The averagingis more expensive since a sorted representation of the particles is needed. But byutilizing the skiplist data-structure, that is used also in the HMM pruning technique,described in 4.3.1.4, keeping track of the best particles is performed efficiently. Thisresult sequence leads to lower localization errors than the non-averaged variant.4.4 DevicesNo special attention is given to design of the device software. After starting thelocalization service, it enters a simple loop that pushes the last RSSI readings overHTTP to the Server. For reading RSSI values from the available Wi-Fi-APIs, theservice chooses a device specific implementation. Currently two such implementa-tions exists, one for linux/libpcap based devices and another one for android baseddevices. Localization results can then be analysed with a browser to fetch variousreporting and visualization pages from the server.4.5 Fat ClientThe Fat Client is used for debugging and evaluating the different aspects of the local-ization framework. It is build on a GUI toolkit that embeds the VTK- VisualizationToolKit for analysing 3D-data from different sources. Furthermore, the Fat Clientembeds an interactive scripting environment in the form of a Python interpreter for