157When the motion model depends on terrain information, the use of terrain-based densities (and conditional densities) becomes potentially much more effective than in the static case. In the static case, the prior density gradually becomes less and less influential with time as the estimates become more peaked. When the targets' movements depend on the terrain features, though, the terrain information can remain highly influential over time.
VII. CONCLUSIONSIt is clear mathematically that the use of prior information can improve both association and estimation algorithms. In practice, however, using even moderately complicated prior densities is computationally expensive. In many cases, the use of prior information does not result in significant improvements, especially when used for data association. With limited computational resources it may be preferable to devote the available resources to other techniques, such as multiple hypothesis algorithms, batch algorithms, or the use of higher level reasoning. The terrain information might also be used in a gating algorithm, rather than to construct a prior density.While using the prior density can be expensive, there are cases where the extra expense is justified and where the prior provides especially valuable information. This suggests a hybrid approach. Dynamically allocating the available resources to tradeoff between the various approaches available should result in an algorithm with high performance across a broader range of input statistics. The problem then becomes to determine heuristics for effectively making the tradeoff decisions. Analyses such as those we have presented for the static problem can serve as a first step for developing these heuristics.Another area that needs further research is the representation scheme for the prior densities. The appropriate representation is a function of the class of densities to be represented, the expected class of sensor report and target state distributions, and of the particular association-estimation algorithms being used. While we have simply discretized over a rectangular grid, many other schemes are available, such as Gaussian mixtures, vectorized representations, neural nets, etc. In particular, for the cases we have considered, Fourier-domain density representations (characteristic functions) might be appropriate, especially if they can be efficiently bandlimited to the most useful frequency ranges.