On Lunar On-Orbit Vision-Based Navigation: Terrain Mapping, Feature Tracking Driven EKF

Graph-based terrain relative navigation with optimal landmark database selection

“…We assume accurate feature detection and data association for all tests, as these issues have already been well addressed in the literature, and set the landmark recognition probability of (2) to be uniform across all landmarks. Existing feature tracking methods, such as those presented in [14], [20], [21], could integrate with our presented algorithms without modification. An example simulated camera image with landmark observations is shown in Figure 7.…”

“…Our approach extends upon that of [21] by incorporating landmarks into the smoother using a location prior as both an initialization point and a measurement, which anchors the estimator in the global coordinate frame, while also allowing for opportunistically tracked features to be added into the landmark database and estimated online. Unlike the MSCKF, the smoothing approach allows opportunistic feature measurements to be immediately incorporated and utilized by the filter after only two observations, rather than waiting for them to exit the sensor field of view.…”

“…Singh and Lim showed that TRN accuracy could be improved by adding the positions of tracked landmarks into the extended Kalman filter (EKF) state vector [21]. The landmark database was used to provide the initialization for the landmark location rather than an absolute truth position, and the landmarks were incorporated into the state vector upon first observation, allowing their positions to be successively refined, effectively achieving sub-pixel tracking resolution within the prior map.…”

“…Several landmark-based TRN systems for planetary landing have been proposed in the literature, including approaches using image patches [14], craters [21], SURF features [28], and Harris corners [11], [20] as landmarks. The majority of these do not mention the means they used to select which features to store in their onboard databases, presumably storing all possible landmarks in the region for the best possible nav- igation accuracy.…”

Terrain-Relative and Beacon-Relative Navigation for Lunar Powered Descent and Landing

“…Depending on the accuracy of the camera, maps, and databases, when these techniques are used in conjunction with an extended Kalman filter, the position of the spacecraft can be estimated with an accuracy on the order of 40 m or less [16]. This method, however, lacks robustness because extreme lighting conditions can significantly change the appearance of the craters and landmarks become sparse near the lunar poles.…”