This technology is a method for registration of terrain models created using stereovision on a planetary rover. Most 3D model registration approaches use some variant of iterated closest point (ICP), which minimizes a norm based on the distances between corresponding points on an arbitrary 3D surface where closest points are taken to be corresponding points. The approach taken here instead projects the two surface models into a common viewpoint, rendering the models as they would be seen from a single range sensor. Correspondence is established by determining which points on the two surfaces project to the same location on the virtual range sensor image plane. The norm of the deviations in observed depth at all pixels is used as the objective function, and the algorithm finds the rigid transformation, which minimizes the norm. This recovered transformation can be used for visual odometry, rover pose estimation, and feature handoff.
Single cycle instrument placement (SCIP) is the single greatest autonomy need for the next generation of Mars rovers. The goal of SCIP is to enable a planetary rover to approach and place an instrument on a scientifically interesting point on the terrain from a distance of ten meters. This must happen within one command cycle, so that after an operator selects a science target and uploads a command, the next response from the rover is the requested science measurement from the target.
The first step in SCIP is the navigation of the rover to a location that places the point of interest within the workspace of the arm that carries an instrument. Uncertainty about the exact target position and accumulated rover localization errors requires that the rover actively keep track of where the target is in relation to itself as it navigates towards it. Once positioned, the rover evaluates the target to ensure the instrument can be safely placed, and then moves it into place with the arm.
Terrain model registration can solve both the target tracking and target handoff problems. Tracking is done by registering successively acquired terrain models of the target area to the initially acquired model of the target. Tracking also provides information about rover motion between views. Handoff is done by registering the target models from two sensors. Registration of 3D surface models is an attractive approach for rover localization. As long as the lighting conditions permit the acquisition of images for stereo, the resulting 3D surface models are independent of the lighting conditions. This is attractive compared to 2D approaches that might have difficulty with tracking features or recognizing places when lighting conditions change.
During testing, surface models were not “cleaned” in any way, and the results are still promising. Other reported approaches require mesh regularization and cleaning in order to ensure that there are no outliers before minimizing a norm that is sensitive to large deviations. These steps may improve the results achievable using robust estimation, but empirically are not required for it to work.