Natural Feature Tracking for Autonomous Navigation

Unmanned Platforms like mobile rovers operating in natural environment rely in many cases on autonomous vision systems. One important component in the processing chain within vision guided navigation is the tracking of landmarks which leads to a subsequent update of the sensor (vehicle) position including pointing parameters. This paper describes how an approach coming from stereo matching can be adapted to the image flow approach. The multiresolution structure of Hierarchical Feature Vector Matching (HFVM) is utilized for an algorithm to become independent of motion prediction, Kalman filtering and smoothness of the pointing parameters: A low resolution dense disparity map between two subsequent frames is gained by HFVM, followed by Feature Vector Matching on interest points in high resolution. This reduces drastically the search space for landmark tracking and fully exploits the robustness of HFVM. We report on experiments simulating the autonomous descent of an unmanned spacecraft over unvegetated terrain which can be directly extrapolated to the vehicle case. First, the surface morphology is
 reconstructed using stereoscopy from a lateral motion component. Relying on this digital elevation model (DEM), a set of interest points (landmarks) is known both in object space and image space which, after tracking these points in subsequent image frames, allows for calibration of each camera position. The paper concludes with information about accuracy and robustness of the calculated positions in dependence of the various influencing parameters (update rate, landmark properties, topographic environment, viewing angles, DEM resolution).