Evaluation of a canonical model approach to probabilistic data association in tracking with particle filters

This paper evaluates a probabilistic approach to data association in a class of tracking problems characterised through intermittent, sparse observations. Examples are tracking of a specific target, such as a suspicious person or a car in urban environments. The used data stems from disparate, often simple detectors, each capable of detecting one type of a feature, such as a license plate, car type, its color, etc. It is also assumed that the detectors are dispersed throughout the environment and have a limited range. Such partial observability combined with the inherent ambiguity of the detected features makes data association challenging. Traditional gating methods are in such settings often not suitable. This paper investigates a probabilistic approach to data association in a combination with Particle Filtering methods. We introduce a composite probabilistic sensor model which incorporates the knowledge about the ambiguity of the origin of an observation of a specific feature (e.g. a vehicle type, fragment of a license plate, etc.). The presented approach is based on canonical probabilistic models and provides a clear guidance for the determination of adequate modeling parameters. The impact of the proposed model is evaluated theoretically and empirically. Moreover, causal probabilistic models of dynamic processes are used for the identification of potentially critical modeling deficiencies and properties of the resulting systems while a testbed is used for empirical evaluation of the overall filtering performance under specific conditions. The experiments confirm the theoretically predicted properties and show that the presented modeling approach allows robust tracking also with ambiguous features.