Fully titled: A Bayesian approach to modelling lost person behaviors based on terrain features in wilderness search and rescue
While not directly about UAV co-ordination, this paper holds some relevance to automated behaviour in the sense that it presents a way of modelling probability of a person's path over certain terrain: potentially providing a template map for autonomous searching in this area. More relevantly for what I'm interested in, the exact method it uses could have applications in the forest-fire tracking project which has recently been kicked into gear. A quick look at the method:
Bayesian statistics are a particular interpretation of what "probability" actually is; in this case, a 'belief' about a system. Taking that into the realm of computer science, a Bayesian model seems to be one where some prior probability assumption is assumed to start with, and then this is continually updated as new information becomes available. The available distribution affects directly the areas to be searched. The initial probability distribution is generated from experts in search and rescue operations, and analysis of search and rescue data which was used to generate a Markov transition matrix of probabilities of transitioning from one terrain type to another.
While not directly about UAV co-ordination, this paper holds some relevance to automated behaviour in the sense that it presents a way of modelling probability of a person's path over certain terrain: potentially providing a template map for autonomous searching in this area. More relevantly for what I'm interested in, the exact method it uses could have applications in the forest-fire tracking project which has recently been kicked into gear. A quick look at the method:
Bayesian statistics are a particular interpretation of what "probability" actually is; in this case, a 'belief' about a system. Taking that into the realm of computer science, a Bayesian model seems to be one where some prior probability assumption is assumed to start with, and then this is continually updated as new information becomes available. The available distribution affects directly the areas to be searched. The initial probability distribution is generated from experts in search and rescue operations, and analysis of search and rescue data which was used to generate a Markov transition matrix of probabilities of transitioning from one terrain type to another.
Why this could be useful:
The nature of forest fires is clearly not entirely predictable, else this project would be redundant and there would be a lot less charred woodland in the world. A Bayesian model such as this, adjusted to take into account existing data of past fires, could be used to model both high-risk areas worthy of preventative monitoring, and also (in the event a fire is detected) the most likely path of propagation for the fire. That is, with terrain and vegetation data one could build a map showing the most likely areas of fire spreading, which is then continuously updated by the incoming sensory data from UAVs and UGVs in the area, allowing adaptive behaviour and monitoring. This could, in theory, also be tied into early warning systems for population areas at risk.
Discretising the area into hexagonal 'cells' proved a useful method of simplification in the paper and could help manage computing resources in a forest fire scenario too.
Shortcomings:
The paper had a number of shortcomings worthy of discussion, stemming mainly from an abundance of assumptions. Whilst it is clear the paper was very much the beginning of possible future research some of these appeared to be so broad as to render the results questionable. Examples of this included the definition of a 'slope' (affecting a person's travel time and direction) as being any elevation difference over adjacent cells resulting in at least a 20 degree angle in terrain. Through some experience in hiking and off-road biking, a 20 degree slope is pretty harsh; amounting to a rise of greater than 1 in 3. Practically, it must be expected that much smaller inclinations than this would affect the missing person's path.
Another rather glaring omission is the lack of consideration of a possible destination. The paper quotes another study which showed that in missing-person incidents a person's desired destination 'strongly correlates' with their behaviour. This strikes me as something of a truism. Indeed, an example of geocachers claims that the location of the 'treasure' they are intending to find "play an important role in the person’s behavior in the wilderness": surely this statement amounts to the admission that people will tend to move towards a destination? This sort of consideration; while not directly applicable to forest fires (fires generally don't desire to move anywhere); would be vital in the context of trajectories and (as stated) paths tending towards populated areas.
Final word:
With some modification it seems this paper lays a potentially useful groundwork for forest fire tracking. Bayesian statistics are well established in autonomous agents and I believe it would be a method worth pursuing. It would also be instructive to see if such a model could perhaps be married to a max-sum co-ordination algorithm (possibly through influence of the utility function).
-C
No comments:
Post a Comment