56 4. DesignFigure 4.8 Scene geometry restricted sampling of the transition probability p( s| s)during the driving process of a Particle Filter. Here, 1000 valid destination statess were generated after 2184 trial samplings. The mean of the Gaussian, the sourcestate s, is located at the red voxel.As in the HMM approach, the information of blocked zones can be derived from the3D-scene. During the sampling phase of the Particle Filter, the candidate sampless, that are either in blocked zones or have blocked voxels on the straight path tothe source state s, are rejected. A result of this sampling process on a crossing isvisualized at Figure 4.8.105samples are generated for each iteration of the driving process representinga single time frame. Further increasing this number has not shown to have anypositive influence on the error rate of the PF. This relates to around 1 samples foreach m3during the initial uniform displacement on the 8900 m3volume of the UMICscene. The samples are drawn in a round-robin scheme from a pre-calculated pool ofGaussian values. Since the constraints of the 3D-scene lead to rejections of samples,and therefore a resampling from the pool, a hard limit on the number of resamplingshas been imposed. Thus, the algorithm is able to proceed timely in pathologicalcases where most samples are rejected.4.3.2.3 Sample ImpoverishmentA problem that arises in the PF algorithm is the degeneration of the current set ofstates into a region of space that has no sufficient probable connection to the reallocation. The process is stuck in that mispredicted region and cannot recover to thetrue region by moving through the sampling process. This does especially occur inthe presented PF, if the 3D-scene restrictions are applied. A problematic zone isgiven by the stairways where the possible path has the form of a corkscrew. If thesampling process"misses" the path trough the stairs, when the device moves fromone floor to another, the degeneration case is imminent.