counting objects with Jevois

Question

I have a Jevois connected to an Arduino Uno. I can connect just fine but i am trying to teach the camera to detect a small thin object, count them and sort by color. i am able to see the target area and adjust it to the aspect ratio i nedd but when i send the teach command, it says it was saved but I am not seeing the green box nor do i see the object being tracked in the Amcap. I am looking at colored disks from the side. the disks are about 1.25 inches tall and about a quarter inch thick. i have them stacked and i am attempting to use the camera to track how many of each color are in the stack. i have mytarget set for about 10 percent to 35-40 percent for the target box. I changed the vision sensor to a JeVois 1.3MP Sensor with 90deg no-distortion Lens. I am looking at the stack from about 2 inches away. could any of these factors be giving me an issue? Thanks

Answer 1

It looks like you are using the ObjectDetect module, right? This module is a bit tricky with simple objects, because it identifies objects by keypoints and associated descriptors. This works very well with complex, flat objects with non-repeated textures, for example a cereal box or a book cover. In those, the algorithm looks for points that have a very unique local appearance, and it matches those between the image you saved and the live image captured by JeVois.

If the object is too simple, then it may lack distinctive keypoints. You can try to adjust parameters hessian, goodpts, and distthresh (see http://jevois.org/moddoc/ObjectDetect/modinfo.html) but that may not be enough.

For colored disks, you may want to check out

http://jevois.org/moddoc/FirstPython/modinfo.html

and modify the python code to look for disks instead of a u-shape.

Or this color-based algo:

http://jevois.org/moddoc/ObjectTracker/modinfo.html

or this for several colors:

http://jevois.org/moddoc/ArUcoBlob/modinfo.html

Comments

Thank you for that insight. I work with robot vision in an industrial setting so most of my experience is with industrial smart cameras. they have many of these features pre programmed. This is a bit new to me and i'll have to play a bit more. I'm a bit new to writing my own algorithms but i was trying to work within the pre existing vision modules.