Viewpoint normalized images for feature based landmark detection

Image feature based landmark detection is widely used in machine vision for mobile robotics applications. Advanced image feature extraction algorithms offer robust keypoint detection as being invariant to numerous transformations such as translation, rotation, scale change. Even slight changes in brightness, contrast or viewpoint do not affect feature point matching abilities. The most widely used advanced image feature extraction algorithms such as SIFT (Scale Invariant Feature Transform) and SURF (Speeded Up Robust Feature) both have these features. Both algorithms implement feature point detection and descriptor generation. Descriptor vectors are calculated around feature points to distinguish and match these points based on the image content. In the application of mobile robotics viewpoint invariance is essential. As the robot moves in its environment it must detect landmarks from widely different viewpoints to improve efficiency of SLAM algorithms. This is essential in loop closing situations. In this paper we evaluate a method to improve viewpoint invariance based on the additional data provided by range image sensors to supplement traditional feature extraction algorithms. Planes are extracted from range images based on local surface normal histograms and feature point matching is evaluated in viewpoint normalized image planes. We present a simulation framework and results for the selected algorithm. We compare feature point matching with and without the improvement.