Localization Using Feature Matching in Near-Random Textures
Precise vehicle localization, whether performed via visual scene registration or special-
ized sensors such as GPS, is currently limited in robustness and accuracy. We propose
a computer vision-based localization system that improves precision by matching fea-
tures in small patches of asphalt or carpet imaged with a downward-facing camera.
The key insight is that such images will have enough near-random detail to retrieve
their correct position in the environment.
In our system, the environment is first processed to construct a knowledge base.
Then during the retrieval phase, we apply multilevel search to precisely localize the
observed patch in the known environment. We adjust the patch registration algo-
rithms previously developed for natural scenes to the case of ground environments
with near-random textures. Our system is marker-free and robust against orientation
as well as illumination changes. By combining temporal coherence with classification
techniques, we can instantaneously provide accurate localization predictions. We also
envision several potential applications including vehicle navigation system and precise
control for 3D fabrication.