Image stitching using gradual image warping in autonomous driving

To improve object recognition and tracking in autonomous driving, we first create a seamless panorama. Object recognition can benefit from image stitching, especially at the borders of  individual images when an object is only partially visible. This also prevents duplicate detection of the same objects in overlapping image areas that are to be filtered for tracking. In  this process, a homography is determined for the overlapping image area, whereby the entire image is transformed using classical image stitching methods. As a result, the  deformations propagate to further images that are to be added to the panorama. To avoid this problem, we integrated a step-by-step image warping approach into our existing stitching  pipeline. This ensures that after attaching one image to another, the outermost right and left borders of the panorama are no longer deformed. Furthermore, the panorama width remains  constant regardless of the calculated homography. We have evaluated our approach on the nuScenes dataset and the Waymo Open Dataset  for perception. In addition to a qualitative  assessment, we evaluate the resulting panoramas in terms of the deformation of the individual images as well as the deformation of labeled object instances.