Feature-based probabilistic data association and tracking

In this contribution we present a concept for improvement of object tracking in applications that suffer from severe detection errors such as incomplete, merged, split, missing and clutter-based detections due to noisy data, sensory and algorithmic restrictions and occlusions. It is based on utilization of low-level information that is gained through tracking dedicated feature points with known relationship to the tracked objects. The proposed Feature-Based Probabilistic Data Association and Tracking Algorithm (FBPDA) can be applied not only in the field of driver assistance systems but also in surveillance applications and further video-based object tracking applications. The main requirement is the possibility to robustly track dedicated feature points in the image (and in 3D space). For this aim, both correlation-based techniques (optic flow) and correspondence-based techniques using e.g. SIFT or SURF features can be used.