3D-recognition (4 degrees of freedom), Jan Wedekind and Manuel Boissenin

Microscopes have very limited depth-of-focus. This imposes a variety of constraints when building imaging and vision solutions for microscopes. The MMVL vision software for the MiCRoN-project uses a geometric hashing technique to compare all images in an image stack with the scene.

Even with recent advances in computing performance, geometric hashing may require in excess of one second to locate a micro-object in the scene image, thus using this technique in a real-time feedback loop is unfeasible. To speed things up, the MMVL vision software uses hough tracking to keep track of objects in subsequent frames. Using recognition and tracking together, it is possible to achieve real-time object tracking for micro-objects.

Current results

The software has been tested extensively under various conditions. A sample of these are shown below. The algorithms have also been stress tested in a real working environment using a live videofeed.

If you find that any of these microscope imaging solutions is useful to you application then please contact a member of staff for more information.

Capacitor handling

image size: 384x288 (jpeg files input)
initialisation: 77.0531 seconds
average time/frame: 2.3597 seconds
maximum time/frame: 6.6389 seconds

Artificial sequence

image size: 768x576 (jpeg files input or gstreamer input)
initialisation: 172.666 seconds
average time/frame: 0.7303 seconds
maximum time/frame: 7.4322 seconds

Solder sphere

image size: 768x576 (video for linux input)
initialisation: 12.5038 seconds
average time/frame: 0.1538 seconds

maximum time/frame: 0.1693 seconds

Micro-gripper

image size: 384x288 (jpeg files input)
initialisation: 16.8425 seconds
average time/frame: 0.0617 seconds
maximum time/frame: 0.2294 seconds

Intersection of pipette with focussed plane (3-DOF only)

image size: 1280x960 (jpeg files input)
initialisation: 7.14593 seconds
average time/frame: 3.3958 seconds
maximum time/frame: 6.6800 seconds