Probabilistic Handling of Merged Detections in Multi Target Tracking

In this contribution we present a method for handling merged detections in video-based Multi-Target-Tracking applications. Merged detections occur when two or more objects evoke one joint detection, i.e. when measurements stemming from multiple objects cannot be resolved by the sensor. In video-based applications this is the case when the segmentation fails to separate blobs belonging to different objects. The proposed approach is based on the specification of candidate merge events and resulting data association events. We propose a concept which allows recognition of the merge events and a correct track update in case of identified merges. This is done by generating artificial (virtual) measurements (measurement reconstruction) through decomposition of respective detections. The globally optimal solution is achieved by weighting different candidate hypotheses according to their a-posteriori probabilities.