Efficient planar features matching for robot localization using GPU

Matching image features between an image and a map of landmarks is usually a time consuming process in mobile robot localization or Simultaneous Localisation And Mapping algorithms. The main problem is being able to match features in spite of viewpoint changes. Methods based on interest point descriptors such as SIFT have been implemented on GPUs to reach real time performance. In this paper, we present another way to match features with the use of a local 3D model of the features and a motion model of the robot. This matching algorithm dedicated to robot localization would be much too slow if executed on a CPU. Thanks to a GPU implementation, we show that it is possible to achieve real-time performance while offering more robustness than descriptor based methods.