Position and Attitude Determination using Digital Image Processing A UROP sponsored research project Sean VandenAvond Mentors: Brian Taylor, Dr. DemozGebre-Egziabher
Overview • Motivation • Introduction • Feature Matching • Position Determination • Google Earth Navigation Simulation • Future Work
Inertial Measurement Units • Measure linear accelerations and rotational velocities • May contain bias error • Error grows over time • $1,000 – $100,000
GPS Update Erroror Noise IMU Readings GPS Update Attitude and Position Integrate • IMU frequency: 50-200 hz • GPS frequency: 1 hz
Problems with GPS • GPS failure can occur in small UAVs - Interference - Jamming - Spoofing • University of Texas successfully spoofed an UAV GPS (2012) and an 80 million dollar yacht (2013). • RQ-170 Sentinel speculated to be spoofed and captured by Iranians (2011) • Severely limits autonomy and control without backup system.
Motivation • Small UAVs are limited by: • Price • Weight • A backup system is required to increase the reliability. This system should be: • Inexpensive • Lightweight • Self-enclosed • Robust and practical
Position Update GPS Images Processing
Proof of Concept MATLAB • OpenSURF • Algorithms Google Earth • KML files
Feature Matching • Feature: unique grouping of pixels in an image. • Given two images the program will output best estimates on matching features. • Limitations: - shadows/lighting - lack of unique image data - minimal relative image rotation These limitations can lead to mismatched feature points between images error in position estimations.
Feature Matching OpenSURF
Limitations Shadows and lighting can degrade image matching.
Limitations No good feature points
Limitations Pitch = 40º Heading = 40º
Position and Attitude Determination • PnP problem • Uses known landmark positions • At least 3 landmarks • Various PnP algorithms available. • Direct Linear Transform • EPnP • Constrained Least Square positions.
Depth Ambiguity • Information lost when converting from 3D scene to 2D image. • Need to know landmark coordinates in all three dimensions. • Non unique solution if you don’t have 3rd dimension • Algorithms breaks down when viewing 2D scene
Simpler Solution • Let Heading = Pitch = Roll = 0º
Position Determination Use image matching and FOV calculations to determine new position.
Attitude Error • Pitch of 5 degrees: estimated position error of
Simulation • Full guidance and navigation simulation in Google Earth using MATLAB • Assumptions: - Heading = Pitch = Roll = 0 - Neglecting aircraft dynamics - Fixed velocity - Fixed altitude
Destination “Aircraft Simulation” Initial image at known lat/lon New heading New image Old image Estimate new lat/lon Feature matching
Percent Overlap Image 1 Image 2 Overlap More Overlap: - Better matches - Slower Tests done with 3000 foot length step overlap of about 25 percent
Summary • Over 100 miles in simulation. • Show promise for using digital image processing for a backup navigation system. • Benefits - Lightweight - Inexpensive - Self-enclosed • Limitations - Needs unique feature points - Computationally expensive
Future Work • Couple simulation with UAV lab MATLAB script to include errors in sensors. • Post-processing from UAV lab flight tests • Adapt to allow onboard flight testing