Multiple extended object tracking using Gaussian processes

The goal of multi-object tracking is to estimate the number of objects and their states recursively over time. In the presence of extended objects, i.e. objects with an extent that is not negligible in comparison to sensor resolution and which give rise to multiple measurements per time step, the tracking task is even more challenging: To fully utilize all available information and to achieve accurate estimation results, measurement models that are able to represent the object shape as well as algorithms that tackle the aggravated track-to-measurements association problem are necessary. In this work, the Gaussian process measurement model is integrated in the recently developed labeled multi-Bernoulli filter for extended objects. Thus, the ability of Gaussian processes to estimate and represent a wide range of free-from shapes is combined with a principled approach to the multiple extended objects tracking problem. The filter performance is demonstrated for simulated and experimental vehicle tracking using a laser scanner. For this particular application, a new, approximately axis-symmetric covariance function is additionally introduced.