Particle filter-based information acquisition for robust plan recognition
2005 7th International Conference on Information Fusion
Plan recognition generates high-level information of opponents' plans, typically a probability distribution over a set of plausible plans. Estimations of plans, are in our work, made at different decision-levels, both company-level and the subsumed platoon-level. Naturally, successful plan recognition is heavily dependent on the data that is supplied, and, hence, sensor management is a necessity. A key feature of the sensor management discussed here is that it is driven by the information need
... f the plan recognition process. In our research, we have presented a general framework for connecting information need to sensor management. In our framework implementation, an essential part is the prioritization of sensing tasks, which is necessary to efficiently utilize limited sensing resources. In our first implementation, the priorities were calculated from, for instance, the estimated threats of opponents (as a function of plan estimates), the distance to the opponent, and the uncertainty in its position. In this article, we add a particle filter method to better represent the uncertainty in the opponent state estimate to make prioritization more well-founded and, ultimately, to achieve robust plan recognition. By using the particle filter we can obtain more reliable state estimates (through the particle filter's ability to represent complex probability distributions) and also a statistically based threat variation (through Monte-Carlo simulation). The state transition model of the particle filter can also be used to predict future states to direct sensors with a time delay (a common property of large-scale sensing systems), such as sensors mounted on UAVs which have to travel some distance to make a measurement.