With Scorpion Vision we continuously make new tools or weapons in the quality war to help our customers improve their services. Our latest invention is tool # 153 – CircleSegmentor. The tool will be available in Scorpion 8.0 due january 2010. The basic specification is:
- Locate multiple overlapping Circles
- Sub-pixel accuracy
- Intensity independent
- Fast – multicore ready
- Extensive Circle validation
The concept is basically:
- Sobel-filter within a region of interest
- Clever edge-decimation
- RANSAC to locate circle candidates
- PolygonMatch to verify and locate circles with sub-pixel accuracy
A test case for the tool is object counting and tracking in fish farming. The animation below shows an underwater image from a fish cage where Scorpion counts the pellets not eaten. To provide a proper counting an advanced tracking algorithm is used to follow each pellets and to collapse the multiple observations of each pellet to a single count. The CircleSegmentor provide the count and the center position of each pellets in every single frame.
The animations shows 2 images pr second – equalling the true speed. A total of 32 pellets are counted in the 45 second animation.
We will “benchmark” the CircleSegmentor’s performance towards a blob approach.
Case # 1 is 10 pellets in the image. The CS counts 10 – blob segments 3 close pellets into one object.
The corresponding blob image – blob 1 contains three pellets.
Case # 2 – CircleSegmentor counts three overlapping pellets.
The image below shows how CS are able to count three overlapping objects.
Object Counting and Tracking
is a large application area for machine vision. CircleSegmentor will find many applications within counting, tracking and robot vision and is an extremely useful extension of the power Scorpion toolbox. The sub-pixel accuracy is important for 3D stereo vision.
PolygonMatch is a trademark of Tordivel AS.