Propagation based tracking with uncertainty measurement in automotive applications

Tracking keypoints through a sequence of images is one of the first major steps of structure-from-motion approaches. This paper presents a robust method to extend from two-frame keypoint matching to multi-frame tracking. The transition from matching to tracking is implemented through the propagation of the predicted 3D position of the keypoint. The uncertainty level of the tracking results is calculated based on the uncertainty of the motion parameters. Experiments on KITTI dataset as well as on a synthetic dataset show that dense and accurate tracking is attainable.