TRN v2

What is Terrain Relative Navigation?

In Terrain Relative Navigation (TRN), a vehicle's position is estimated by comparing terrain measurements (e.g. sonar or altimeter) with a terrain map.  The vehicle is most likely to be over areas of terrain whose profiles (in the map) most closely correlate with the profiles measured by the vehicle.  A simple 1-D example is shown in the figure below.

Engineering implementations of the TRN concept take different forms, but are generally involve probabilistic representations of vehicle state estimate over varoius degrees of freed.  Depending on the application, these representations can be highly non-linear, and various TRN formulations have been developed for use in the various applications.

Prior work in TRN

One of the earlist uses of TRN was in cruise missile guidance systems in the late 70's.  The algorithm, called TERCOM (terrain contour matching), matched a series of altitude measurements against a map to correct a missile's position estiamte.

Since the advent of GPS, TRN has seen little use in aerial and surface vehicles.  However, as underwater vehicles have become more capable and prevalent, TRN has become the subject of a good deal of research.  Noteable underwater TRN work has been done at WHOI, CMU, Stone Aerospace, KTH Stockholm, the University of Sydney, and many others.

TRN research in the ARL

In conjunction with the Monterey Bay Aquarium Research Institute (MBARI), the ARL is interested in characterizing and extending the capabilities of TRN algorithms.

Two TRN formulations have been used in the ARL: a generalized discrete Bayes filter, and a particle filter.  Both filters have been used with logged sensor data to demonstrate correctness and convergence rates of TRN estimates as well as the ability of TRN to work using various terrain sensors including multibeam sonars, 4-beam doppler velocity loggers, and single-beam altimeters.  The convergence of a TRN particle filter using multibeam sonar measurements is shown in the figure below.  The particle distribution is shown after 0, 2, 4, and 12 measurement updates.  After 12 measurement updates (over ~75m of travel), the weighted mean of the distriubtion (cyan +) has converged to agree with the Kearfott DVL/INS dead-reckoned estimate (yellow +).

<<Particle Filter Convergence Image>>

Since April 2008, a version of the generalized discrete Bayes filter TRN aglorithm has been fielded on an MBARI Dorado-class autonomous underwater vehicle (AUV).  While the algorithm had previously been shown to work using logged data, this field demonstration shows that the algorithm can run in real time, using the modest CPU power available onboard the vehicle.  TRN and INS solution trajectories from an April 24 dive appear in the figure below.

<<Field Operation Image>>

Under-Ice Navigation

For safety and science reasons, a vehicle operating beneath an iceberg most know its position with respect to the ice.  In the case of a free-floating iceberg, this is a particularly difficult problem.  Not only do icebergs translate, but they rotate freely, rendering inertial heading references useless over all but very short timescales.  Simulations in the ARL have demonstrated the ability of TRN algorithms to perform successfully in the presence of terrain rotation.  However, these algorithms rely on a-priori knowledge of an iceberg terrain map.  Ongoing work in the ARL is focused on the application of TRN concepts to the problem of mapping terrain which is both translating and rotating.  One candidate approach involves relative pose estimation between overlapping sonar swaths using TRN techniques.  A second approach leverages the idea that large-area, low-resolution maps can be created quickly compared to the timescale of an iceberg’s rotation, and then used in TRN algorithms to navigate the vehicle over slower, lower-altitude trajectories for the creation of higher resolution maps.  Preliminary data for this study come from a shipboard system deployed in Antarctica in June 2008, while the long-term goal of the research is to enable autonomous operation beneath icebergs (below).

<<Antarctica Ship Image>>

<<Under-Ice AUV Image>>

Detecting Terrain Changes

Additional TRN work in the ARL is focused on detecting changed terrain in TRN algorithms.   The success of terrain navigation relies on the ability to match current terrain measurements with previously stored terrain maps.  Thus if the true terrain differs from that predicted by the map (depicted in red on the left image),  the measurement registration will be hindered and the terrain-navigation filter will be prone to divergence.

<<Detecting Terrain Changes Image>>

In the image below, the TRN filter diverges as a result of the vehicle unknowingly crossing a region of where the terrain differs by ~5 meters from expected (e.g. from a landslide). To avoid this divergence, the vehicle must be able to detect when terrain has changed in real-time such that it can ignore the associated measurements.  Current research in the ARL is aimed at developing on-line, real-time terrain change detection, thereby increasing the robustness of TRN applications in dynamic environments.