Vehicle Navigation
There are many applications for sensor fusion in the vehicle navigation space. Applications and possible sensors include:
Applications:
- Open sky (Ag fields, Sod farms,…)
- Occasional tree blockage (parks, commercial mowing, …)
- Orchard (transitions from open to blocked)
- Stadium (little sky visibility)
- Golf (tree blockage and terrain)
- Residential mowing (low-cost potentially complex environments)
- Industrial machine control: outdoors, indoors, seamless transition
- Construction (grading)
- Road vehicle nav (Snowplow,…)
- Undercanopy (terrain; surveying, UXO,…)…
- Military: R-gator, M-gator, Packbot, …
Sensors:
- GPS
- RF-LPS
- Acoustic ranging
- Laser systems: RTS, SICK, ArcSecond
- Compass
- Odometry
- Inertial (gyros and accelerometers)
- Perception of environment: Optical odometry, vision, stereovision
- Intentional landmarks: Magnets, buried wires, lights, visual markers
- Laser elevation (grading)
Military vehicle navigation -- RGator
A flexible sensor fusion system was implemented in the RGator product, a military scout vehicle, shown below.
Sensors included GPS (with RTK as an option), 3axis accelerometers, gyros, and magnetometer, and wheel odometry. Design process focused on the systems engineering characterization and trades of sensor suite and grade, including simulation and field trial postprocessing to minimize systems cost at the required performance.
For more information, please see RGator nav paper
Dead-reckoning correction sensors
For some types of absolute position, constant access to an absolute measure of position is not required and deadreckoning between occasional absolute position measures is sufficient. Deadreckoning can often be good to 10cm over 10-30-60s, depending on the motion of the vehicle and accuracy of wheel odometry. Below are 2 examples of sensors for correcting deadreckoning.
Buried magnets
While infrastructure is often not desirable for positioning systems, magnets buried in yards or roadways require no power supply and minimal maintenance.
The video below shows an android smartphone used as a sensor passing about 6 inches from a magnet. A particle filter is used to associate the observed 3axis magnetic field with a dipole model of the magnet. This shows that within a 1-foot range, the magnet is sensed reliably, and the position is accurately estimated within 1 inch.
Visual boundary classification/detection
Another form of deadreckoning correction is the visual detection of an area boundary. The results below are from a downward pointing camera about 5 inches above a yard. The type of surface is classified based on observed texture and hue/saturation. The texture is based on the spatial correlation scale. The edge is also detected during transitions from one surface classification to another.
This type of deadreckoning correction gives 1 axis of absolute positioning correction for every edge transition. To keep position accurate in 2D, edge transitions must occur often in naturally different orientations, or a path constraint must be placed on the vehicle to ensure favorable geometries.
Vehicle navigation -- GPS denied environments
Our primary technique for navigation in GPS-denied environments is to use RF positioning sensors for absolute location input. For more on these techniques, please see the section on RF Location.
An example of this system is shown below operating in a domed stadium.
Stadium mower
Many sports stadiums have very high elevation angle horizons from the field or in the case of domed stadiums, have the sky completely blocked. Even the best case sky visibility in stadiums only allows 1-2 GPS satellites, so other measurements must be used.
In our case, we used RF location via micro-impulse ranging to active beacons surrounding the field. We used a total of 6 beacons, and with multipath ground bounce fading especially at longest range operation (>500ft), the minimum number of beacons visible was 4. This still yields plenty of overdetermination, since on a plane surface we'd need only 2 that are widely angularly separated for operation. Also, as in GPS-based vehicle navigation, odometry and inertial sensors allowed deadreckoning even if short outages of many beacons occurred.
The platform was a robotic mower, designed to mow a ballpark outfield with precise stripes. This system is show below.
The entire fusion engine with RF ranging electronics was placed in a prototype enclosure (show below), and it operated for an entire baseball season, over 1000 hrs of mowing, without issue.
Prototype beacon
Short video showing navigation:
Car fusion system example
As an example of our proprietary fusion software, the example below shows the filtered navigation state on a road vehicle using
- single frequency GPS (ublox)
- 9-axis IMU: accelerometer, gyro, magnetometer triads
with a state comprised of
- vehicle position, velocity, acceleration; orientation and rates
- vehicle velocity/acceleration in body coordinates, to constrain sideslip and pitch vs groundplane
- rotation rates and accelerations modeled at first-order gauss-markov processes, with time constants and steady-state deviations selected to match vehicle behavior
- measurement biases and scales also modeled as GM with long time constants and prior spec uncertainties
Software allows selection of filters (unscented kalman filter, both standard and square-root information, and particle filter), recording and playback of serial frames, and visualization. For instance, the following plot shows estimated position and heading, with vehicle speed superimposed color-coded and coincident with 1s GPS measurement locations.
Dead-reckoning and vehicle suspension effects
Deadreckoning on a vehicle is usually achieved via a combination of
- Vehicle odometry, from encoders on steering and drive train
- Inertial sensors (accelerometers and rate gyros) attached to vehicle body
Both of these types of inputs are susceptible to vehicle suspension assumptions:
- Tire give between the ground and wheel/axle
- Independently-suspended wheels (wheel to vehicle body/frame has spring/shock independent of other wheels)
See a section from a report on vehicle nav, while exemplifies the types of systems analyses we do
Vehicle Control: Intended-Path Tracking
see technical note for more information