A Quick Feature Detecting Method Applied in Robot Vision

Detecting scale-invariant feature is very important in robot vision fields, such as object recognition and vision-based localization. However, the methods detecting features always have a lot of computation and can not meet the real-time demand. To solve the problem, a quick method for detecting interest points is presented. It is based on a nonholonomic pyramid frame, whose influence on the repeatability is analyzed in theory in this paper. The method computes the Harris corners in each level in image nonholonomic pyramid scale space and uses difference of Gaussian to select the interest points. The points are robust to image translation, image rotation, image noise, scale changes, illumination changes and so on. Most important of all, the method can ensure the performance and evidently decrease the computation time at the same time. The experimental results have certified its validity.