Sensor Management for Identity Fusion on a Mobile Robot

This paper addresses the problem of identity fusion, i.e. the problem of selecting one of several identity hypotheses concerning an observed object. Two problems are considered. Firstly the problem of representing and fusing measurements relating to identity hypotheses is treated. Here it is shown that if the hypotheses are chosen to be concerned with object properties rather than object identity, Bayes' combination rule outperforms Dempster-Shafers. Further some points are made concerning the numerical properties of the normalization in Dempster-Shafers combination rule, and of the use of a priori knowledge in identity fusion. Secondly the problem of selecting the most appropriate sensing action is addressed. A number of methods from the literature is described and compared, and a simple and computati-onally cheap metric is proposed and shown to be very eecient. 1 The Identity Fusion Problem Consider a mobile robot navigating in an uncertain environment that needs to be able to identify certain objects in its path. Speciically it needs to be able to determine if an object in front of it is a door. The frame of dis-cernment for this problem can be chosen as: where the mutually exclusive and exhaustive set of hypotheses are: H 1 : The object is a closed door H 2 : The object is an open door H 3 : The object is a window H 4 : The object is a person H 5 : The object is something else For this, the robot is equipped with a suite of sensors. Using these, the robot can then perform a sequence of measurements each stating the identity of the object and fuse these using for instance Bayes' or Dempster-Shafers rules. To obtain identity measurements like z s = H 1 , however, the sensor data will need to be preprocessed. In this processing some sensory information can be lost. The sensor data might be more ambiguous or uncertain but then be forced to choose a favorite hypothesis to provide a nice input to the identity fusion. Instead of letting measurements declare identity, more information can be preserved in these by relating them to object properties. A number of properties that separates a door from its surroundings could for instance be: z 1 : Squareness (a door is a fairly square object) z 2 : Color separation with wall z 3 : Depth (especially when open) z 4 : Dimensions (within certain min-max limits) z …