Tracking corners under a perspective shift, with orientation noise

In lidar tracking, a common trick is to track a corner of an object's box rather than its centroid, to be robust to perspective changes. How does this work, and if you use the object's orientation to do it, how do you incorporate the noise in orientation?