Detection and tracking of objects with direct integration of perception and expectation

One of the main challenges in video-based multi-target tracking is the consistent maintenance of object identities over time. We present a novel approach to that challenge that integrates tracking and detection in a single process. We thereby inherently solve the identity problem and gain additional stability of the object detection performance. For that purpose, we extend a state-of-the-art local-feature based object detector by integrating expectations resulting from tracking directly into the detection procedure on the level of features. By that combination of newly gathered and expected local features we are able to directly integrate new data-evidence with object knowledge collected in the past without changing the detection approach itself. Since our tracking approach is solely based on local features, without employing other features like color or shape, it works independently of underlying video-data characteristics and preserves the general applicability independent of object-class specifics.