Parameter optimisation for vision guided terrestrial locomotion: Multi-frame

During locomotion, Autonomous Ground Vehicles (AGV) need to map the surface of the ground ahead in order to plan a safe route of passage. Algorithms that achieve this are all derived from a basic requirement to track pixels between successive frames. This paper establishes the effect of tracking error on depth estimation with different image-plane geometries and presents an argument for tight control of the camera orientation. It is found that the effect of tracking error is invariant with camera angle for a concave image-plane but varies non-linearly with angle for both flat and inverse-flat image-planes. Pairs of frames are produced by each camera from a simulated scene and a selection of algorithms are used to compute the optical flow. It is shown that current state-of-the-art optical flow algorithms can successfully track pixels across concave and inverse-flat image-planes. This results in comparable depth estimation accuracy to a flat image-plane but allows for improved object tracking.