Improving Robustness of Tracking. For more robust tracking of maneuvering persons, we extended this tracking method with an interacting multiple models (IMM) approach that dynamically mixes four different motion models depending on the evolution of a tracked per- son’s trajectory and the prediction error: 1) A constant velocity (CV) model with low process noise, 2) a CV model with high process noise, 3) a Brownian mo- tion model, 4) a coordinated-turn model. In lab experiments, this combination helped to reduce the number of track losses caused by abruptly stopping or turn- ing persons, without us having to manually tune process noise levels, especially in smaller or very crowded environments where motions are less linear. Our implementation also optionally subscribes to a second DetectedPersons ROS topic, on which detections of a high-recall, low-confidence detector (e. g. a primitive laser blob detector) can be published. After regular data association, we perform another round of data association on these detections only with tracks that have not yet been associated with a detection (from the high-precision xx- xxxxxx). This simple modification can improve tracking performance significantly if the main detector has a low sensitivity. Finally, an implementation of a track initialization logic as described in [6] helps to prevent false track initializations in case of spurious misdetections.
