Feature-based probabilistic data association and tracking - a novel approach capable of tracking objects under splits, merges and occlusions
Uncertainties in the sensor data such as measurement noise, false detections caused by clutter, as well as merged, split, incomplete or missed detections due to a sensor malfunction or occlusions (both due to the limited sensor field of view and objects in the scene) make multi-target tracking a very complicated task. Thus one of the big challenges is track management and correct data association between detections and tracks. In this contribution we present an algorithm for visual detection and tracking of multiple extended targets under occlusions and split and merge effects. Unlike most of the state-of-the-art approaches we utilize low-level information integrating it in a unified approach based on a threshold-free probabilistic conception. The introduced scheme makes it possible to utilize information about composition of the measurements gained through tracking of dedicated feature points in the image and resolves data association ambiguities in a soft decision using a globally optimal probabilistic data association approach. Beside existence evolution consideration we also exploit the spatial and temporal relationship between stably tracked points and tracked objects, which along with observability analysis, allows us for reconstruction of compatible measurements and thus correct track update even in cases of splits, merges and partial occlusions of the tracked targets.