Robust multi-Bernoulli filtering for visual tracking

To achieve reliable multi-object filtering in vision application, it is of great importance to determine appropriate model parameters. Parameters such as motion and measurement noise covariance can be chosen based on the image frame rate and the property of the designed detector. However, it is not trivial to obtain the average number of false positive measurements or detection probability due to the arbitrary visual scene characteristics from illumination condition or different fields of view. In this paper, we introduce the recently proposed robust multi-Bernoulli filter to deal with unknown clutter rate and detection profile in visual tracking applications. The robust multi-Bernoulli filter treats false positive responses as a special type of target so that the unknown clutter rate is estimated based on the estimated number of clutter targets. Performance evaluation with real videos demonstrates the effectiveness of the robust multi-Bernoulli filter and comparison results with the standard multi-object tracking algorithm show its reliability.