A region-growing based clustering approach for extended object tracking

In the case a scenery consisting of multiple moving objects has to be observed and analyzed by using radar, it may occur that extended objects cause more than one observation. As a consequence, a conventional tracking algorithm, that bases on the assumption of point objects, has to process lots of observations, generates several tracks per object and thus is slowed down distinctly. Moreover, it is necessary to sort out and merge the tracks before they can be used. In order to avoid these problems, a clustering algorithm for radar-based object tracking is presented in this paper. The algorithm combines, assigns and discards observations before they are passed on to the tracking. Thereby not only the observations are utilized, but also the existing tracks. Furthermore, the method proposed and its benefits are tested in the example of an automotive object tracking system.

I. Introduction

Supposing that a multi-beam radar sensor is to be used to detect and track objects, it may occur that large objects, for example vehicles, intersect various beams simultaneously. As a consequence, one obtains more than one observation per object. And if the scenery consists of more than one object, it is impossible to associate the observations immediately. As a result, a lot of superfluous, transient tracks have to be created and maintained, which causes a lot of processing load and slows down the whole object detection process considerably. Moreover, objects induce a variable number of tracks that have to be sorted out and merged by another processing step.