Developed method of creating a high spatial resolution video from a series of panorama-based still images
Independently implemented dynamic image stitching in C (OpenCV) and integrated into existing software
1. High Quality Simulated Video from Static Images
UC San Diego - Team Internship Program
Alexander Chan | Nima Hashemi
Project Supervisor - Dr. Shay Har-Noy Technical Lead - David Schmidt
EnerView
Real-Time
Video
Stitched Mosaic
(Made from1600x1200
jpeg Images)
(1)
(2)
(3) 1. User requests region of
interest on Video Feed
2. Find corresponding area on
mosaic by matching video
and mosaic timestamp
3. High Quality Image Viewer
scrolls along mosaic in
synchronization with video
4. Suspicious object spotted
behind the tree - an airplane!
5. Grayed area of mosaic
demonstrates regions that
will be stitched dynamically
(5)
(4)
(2)
(3)
(4)
(1)
1. UAV captures video / images of ground dynamically
2. Data transmitted via airborne modem
3. Data passes through 10 mb/s data link
4. Video/Images received through ground modem
High Quality
Image Viewer
Window
Screenshots:
Extract and match image features using
SURF algorithm and NCC matching.
Use RANSAC to determine inliers
(accurate feature matches) and find a
homography relating the two images.
Apply homography and stitch images
together by position mapping and
blurring.
1. OpenCV (C++) Image Processing Library used extensively to
implement image stitching algorithm
2. IJG Library and existing EnerView System used to design High
Quality Image Viewer Window
• UAV must travel a straight path with no rotation, making the
images it is taking appear linear.
• A more robust stitching algorithm is needed to allow for
arbitrary UAV movement.
• Timestamp correlation between video and still images may
have a 3-5 second delay, causing discrepancies between
EnerView Video feed and High Quality Image Viewer.
• Images must be captured by camera supporting EXIF format.
• The EnerView System features time and CPU consuming
processes, particularly the image stitching algorithm.
• Multithreading between user interface and stitching algorithm.
• Implement fixed upper bounds on the number of SURF
features extracted and number of RANSAC iterations.
• Further image down-sampling before feature point extraction.
The current implementation demonstrates that it is
advantageous to use simulated video from still images to
identify suspect features. This is particularly true when
users are interested in zooming in on specific features.
Zooming using our approach is achievable to a very high
resolution, especially as compared to video zooming.
Direction of
video motion