Data association for visual multi-target tracking under splits, merges and occlusions

Abstract

In this contribution we present an algorithm for visual detection and tracking of multiple extended targets which is capable of coping with merged, split, incomplete and missed detections. We utilize information about the measurements' composition gained through tracking dedicated feature points in the image and in 3D space, which allows us to reconstruct the desired object characteristics from the data even in the case of detection errors due to limited field of view, occlusions and sensor malfunction. The proposed feature-based probabilistic data association approach resolves data association ambiguities in a soft threshold-free decision based not only on target state prediction but also on the existence and observability estimation modeled as two additional Markov chains. This process is assisted by a grid based object representation which offers a higher abstraction level of targets extents and is used for detailed occlusion analysis.