Multiple Target Tracking With Uncertain Sensor State Applied To Autonomous Vehicle Data

In a conventional multitarget tracking (MTT) scenario, the sensor position is assumed known. When the MTT sensor, e.g., an automotive radar, is mounted to a moving vehicle with uncertain state, it becomes necessary to relax this assumption and model the unknown sensor position explicitly. In this paper, we compare a recently proposed filter that models the unknown sensor state [1], to two versions of the track-oriented marginal MeMBer/Poisson (TOMB/P) filter: the first does not model the sensor state uncertainty; the second models it approximately by artificially increasing the measurement variance. The results, using real measurement data, show that in terms of tracking performance, the proposed filter can outperform TOMB/P without sensor state uncertainty, and is comparable to TOMB/P with increased variance.