The computer knows where the camera is in relation to the ground and searches for instances of vehicles (the class of the vehicle and its position on the ground) which match the edge data in the image.
Once a vehicle is recognised, it can be tracked through long sequences of video, independently of rotations and occlusions by known structures.
Image of Typical Road Scene with Extracted Edges.
Poses of cars recovered automatically from edge image, using the
algorithm described in
Tan, Sullivan and Baker, 1994 (see Papers)
Camera View, overlayed with the scene interpretation.
Top View from Virtual Camera, showing accurate pose recovery from a single camera 200 metres away.