DRIFT CORRECTION FOR CAMERA TRACKING

摘要

Techniques are disclosed for drift correction for camera tracking. In some cases, the techniques include iterative optimization of the camera poses at selected keyframes captured along the camera trajectory to reduce pose errors. Such iterative optimization may include performing an alignment process using the point cloud of a given keyframe and a point cloud made from one or more overlapping keyframes, such as via an iterative closest point (ICP)-based expectation maximization. The keyframes may then be fused to reconstruct a more accurate model of the scene, discarding the existing model. The new model can then be used for tracking and meshing. In some instances, keyframes may be selected such that the overlap with other keyframes includes enough shape features to allow for alignment with the keyframe depth point cloud. In some cases, the techniques can be performed in a particular order giving precedence to keyframes having smaller pose errors.

主权项

1. A method for camera tracking drift correction, the method comprising:
obtaining keyframes from a plurality of image frames, wherein each keyframe has a point cloud and a camera pose; mapping one or more overlapping keyframes into the field of view of a given keyframe to create a combined point cloud from the overlapping keyframes; performing an alignment process on the point cloud of the given keyframe and the combined point cloud to determine whether the camera pose of the given keyframe is to be readjusted; and in response to determining the camera pose of the given keyframe is to be readjusted, readjusting the camera pose of the given keyframe to a new camera pose.