A fast implementation of the Labeled Multi-Bernoulli filter using gibbs sampling

This paper proposes a fast implementation of the Labeled Multi-Bernoulli (LMB) filter based on a joint prediction and update scheme. The joint calculation prevents the treatment of insignificant hypotheses, e.g. considering the disappearance of an object with high existence probability which additionally generated a precise measurement in the received measurement set. Further, a Gibbs sampling approach for generating association hypotheses is presented which drastically reduces the computational complexity compared to Murtys ranked-Assignment algorithm. The proposed Gibbs sampling implementation is compared to the standard implementation of the LMB filter using two scenarios: Tracking vehicles using a multi-sensor setup on a German highway and extended object tracking in an urban scenario using Velodyne data