54 4. DesignFigure 4.6 Top 300 unpruned states at the beginning of a location tracked mea-surement run visualized by the white cross symbols. The path starts on the firstfloor and ends on the ground floor by passing the stairways.concurrent unlocked insertions this has not been done in the presented decoder dueto time constraints.Figure 4.7 Distribution of 500 unpruned from 66000 total States raveled into onedimension over all timeframes of a decoding run. Darker areas indicate more un-pruned states that have to be processed at the corresponding 3D location.The addition of pruning to the Viterbi Decoder makes it feasible to decode a se-quence of measurements in real-time on a resolution of 0. 20 cm leading to 2 · 106states on commodity hardware. Although only a third of this resolution(0. 60 cmleading to only127· 2 · 106states due to the cubic nature of the problem) was em-ployed on the designed models, as further detail has not yet brought any betterevaluation results, having such a computational reserve is promising. Especially, ifpossible enhancements like further complex models such as HMMs of higher orderare assumed. Introducing for example second order HMMs leads to an expansionof the search space by the number of possible transitions, therefore also to a speedreduction of factor 75 with the chosen(5, 5, 3) model. It is expected, that this is stillcomputational tractable on current hardware.