Rolling-Shutter Correction for Monocular Vision-Based Obstacle Mapping
Open Access
- Author:
- Dunbabin, Oliver
- Graduate Program:
- Aerospace Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- April 23, 2020
- Committee Members:
- Eric Norman Johnson, Thesis Advisor/Co-Advisor
Jacob Willem Langelaan, Program Head/Chair
Jacob Willem Langelaan, Committee Member
Puneet Singla, Committee Member - Keywords:
- Rolling shutter
progressive scan
Monocular vision
Visual mapping
SLAM
feature-point estimation
obstacle detection
obstacle avoidance
EKF
UKF
time-delay
time delay
time delay compensation
mapping algorithm
UAV
UAV mapping
localization
Simultaneous localization and mapping
Cameras
Image sequences
Layout
Trajectory
camera
image sequences
trajectory
image sensors
Kalman filters
motion estimation
nonlinear filters
SLAM (robots)
kalman filter
real-time vision
inverse depth parametrization
monocular SLAM
monocular simultaneous localization and mapping
extended Kalman filter
feature initialization
camera motion estimates
inverse depth
inverse depth parameterization
monocular mapping
rolling shutter time delay
GPS denied
navigation
feature-extraction
robot vision
computer vision
rolling-shutter model - Abstract:
- The ability of an autonomous vehicle to detect its immediate surroundings is of paramount importance for its successful operation in cluttered environments. This research describes an obstacle detection algorithm for an Unmanned Aerial Vehicle (UAV) suitable for applications where non-visual mapping techniques may perform poorly. Feature-point detection and mapping using a monocular vision sensor is a particularly attractive hardware choice for this application due to its relative simplicity, economy of scale, and cost. Although many visual mapping techniques currently exist, limited focus has been directed towards the use of a rolling-shutter camera, due to its progressive scan nature of image acquisition and subsequently coupled time-delay. The focus of this thesis is to enumerate the primary source of time-delay borne from using a rolling shutter camera for visual mapping: that is, the dependence of feature-point measurement time-delay with its location in the image frame. Because a rolling shutter progressively scans pixel rows (or columns) in a sequential manner, a feature-point observed at different locations within the image will have been captured at different time-points. Furthermore, an image frame is usually only stamped with one time-stamp, hence we get a disparity between recorded time of measurement and actual. Presented here is a means of compensating for this time-delay with the aim of improving mapping performance. Unlike the limited library of past work on this matter, where the vehicle’s trajectory is approximated by a piecewise polynomial function with a sparse number of control points, this work parameterises a camera frame’s line scan time, then draws on a buffer of the vehicle’s past states to interpolate to our approximated time-point. This approximated state is then used to sequentially update our database of landmark position estimates. The contribution of this current work is to present a recursive map estimation algo- rithm which compensates for this time delay in both an Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF) framework, using an inverse depth parameterisation of landmark location. The estimator performances of these two filters are then directly compared, with and without time-delay compensation, in both Monte Carlo simulation and in hardware.