Once we have probabilistic sensory data from the sensor models, we use dynamic Bayesian networks to fuse all the data together. The sensor manager looks at the sensory data and picks out objects of interest (ex. using an object recognition algorithm on an image). Corresponding Bayesian networks are selected from a pre-designed library and the sensory information is propagated through these networks to obtain the new beliefs in the objects. An example of a dynamic Bayesian network for car and road detection is shown below.
In the example, the object recognition algorithm suggested the presence of a car and so a Bayesian network for car detection was activated. The inputs to this Bayesian network were the sensory data obtained from the probabilistic visual and audio sensors. After propagation, the combined belief in the presence of a car is obtained at the CAR node in the network.
A proposed modification of the Bayesian network allows us to evolve the structure of the Bayesian network through time using an auxiliary knowledge rule base. In the above example, there was no sensory information about cars at time T-1. Hence, our belief in the presence of a car was completely determined by our prior. At time T we receive sensory data that indicates the presence of a car in the environment. Our knowledge rule base may then tell us that cars drive on roads. Thus, we then modify our Bayesian network to also reason about the presence or absence of roads along with cars.
We also propose to modify Bayesian networks to incorporate a temporal constraint. This is represented as a causal link to the Bayesian network in the previous time interval. A properly chosen conditional probability matrix on this causal link can be used to weight past beliefs and use them to influence the current belief. Temporal propagation works as described in the figure below.
Whenever a particular Bayesian network becomes active, the temporal belief manager stores the belief generated by the Bayesian network in a short-term memory structure. When new sensory evidence about that object is obtained, and once again propagated through the Bayesian network, the temporal belief manager modifies the Bayesian network to include a causal link to a dummy node and copies the most recent belief about that object into the dummy node. Thus, the belief at the previous time provides direct diagnostic support to the new belief. The conditional probability matrix on this new link can be engineered to provide the requisite amount and type of support. For example, we can make the system accept new beliefs quickly but reject them slowly, increasing robustness in the event of sensor failure. The new belief is then stored in the short-term memory and can be fed back through the temporal loop at the next time step.
This page is maintained by Amit Singhal.
Last update: 11/17/97.