Calibration of tracking systems using detections from non-cooperative targets

Tracking algorithms are based on models: target dynamic and sensor measurement model. In most practical situations the two models are not known exactly and are typically parametrised by an unknown random vector θ. The paper proposes a Bayesian algorithm based on importance sampling for the estimation of θ. The input are detections/measurements collected by the tracking system from non-cooperative targets. The algorithm relies on the particle filter implementation of the probability density hypothesis (PHD) filter to evaluate the likelihood of the measurement set history conditioned on θ. As a byproduct, the proposed algorithm can also output a multi-target state estimate over time. An application to sensor bias estimation is presented in detail as an illustration. The resulting sensor-bias estimation method is applicable to asynchronous sensors and does not require prior knowledge of measurement-to-track associations.