|
Object Recognition and Tracking with AVM Navigator
Kamlesh Thakur from India [32 posts] |
11 year
|
Hello,
Am using AVM Navigator v0.7.4.1 licensed version with RR v2.46.37.
With object recogntion module in AVM navigator, I could train an object especially a human eye (with front facial position) and AVM tracking algorithm can track the trained eye object very well. However, I noticed that the object recognition and tracking works only for the person whose eye object image was trained and not for any human eye in same front facial position with same distance from camera as that of trained object.
In short, I am looking for an algorithm/program which can recognize and track a human eye object (when a person is sitting in front of camera with front facial position) for an eye object and any person (who sits in front of camera with front facial position at a distance of 1 feet to 2 feet) and not just only for a particular eye object image used during Learn Object process in AVN Navigator- object recognition
|
|
|
|
EDV [328 posts] |
11 year
|
|
What if you could try to train AVM on eye of several different humans? I think it would help to solve this problem. You should just click on “Learn object” button and choose the same “Eye” object in “Enter object name” dialog window for additional training on eye of other people. If you will have collection of humans eye images inside AVM search tree then it properly increase possibility of recognition anyone other eye that was not learned.
|
|
|
|
Kamlesh Thakur from India [32 posts] |
11 year
|
Hello,
Thanks for the update.
At present am working on my academic project for opto-mechatronic eye tracking system. The system has a camera and a desktop turret mechanism equiped with a pan axis motor and a tilt axis motor (see the attached picture). The camera is fitted on top of the pantilt unit of the turret mechanism. The turret mechanism is interfaced with RR and AVM navigator software. The objective of the project is, when a person sits infront of the camera (in front facial position at a distance of 1 feet to 2 feet), the camera should recognize and track one of the eye object (preferably the left eye), start tracking the recognized eye object and the turret mechanism should position the camera so that the recognized eye object is in centre of the image acquired by camera.
At present I noticed that the system works fine with AVM navigator but only for a person whose eye object is learned to the system. I want to make this system generic for every human eye. The system can be trained with eye objects for few people but it's not practical to train it for every human eye.
Secondly, I have a database of eye object images (.jpg files) on my computer. How can I put these jpg image files in AVM search tree?
|
|
|
|
EDV [328 posts] |
11 year
|
Now additional training from JPG images is possible only by hand with helping “Load image” and “Scale image” modules. You should first select JPG image in “Load image” module from your database and further press “Learn object” button in “AVM Navigator” dialog window.
 program.robo
|
|
|
|
Kamlesh Thakur from India [32 posts] |
11 year
|
Hello,
I noticed that, AVM Navigator v0.7.4.1 can recognize and track the human eye object only when the object is placed at a same distance as that of the object distance during learning process. Please see the attached JPEG image file for more details. In Image1, the eye object is recognized as it is at the same distance as that during learing process. However, when the eye object is moved few centimeters away from the camera (image2), the same eye object is not recognized for tracking.
Is there any remedy to overcome this limitation such that the learned eye object can be recognized when it's moved close or away from the camera within the vicinity of 1 feet to 2 feet?
|
|
|
|
EDV [328 posts] |
11 year
|
|
You could try to use higher resolution of video camera (for example 960x720 pixels) for more scale possibility of eye object.
|
|
|
|
Kamlesh Thakur from India [32 posts] |
11 year
|
Hello,
I tried with a camera resolution of 960x720 pixels. It helped a little bit but not much. With the increased camera resolution, now the system can recognize and track an eye object when the face is moved closer to the camera few centimeters from it's centre position. However, the system is still unable to recognize and track the same eye object when the face is moved few centemeter away from the central position.
During the practical operation of the system it's most likely that the person will sit in front of the camere with an offset (few centimeters) from the central position (position when the eye object was learned in AVM Navigator)
-Kamlesh
|
|
|
|
EDV [328 posts] |
11 year
|
What key image size do you use (80x80 or less)?
Try to use default value that will set if you click “Set key image size (New)” button and further press “Yes“.
|
|
|
|
Kamlesh Thakur from India [32 posts] |
11 year
|
I did not use the default key image size of 80x80.
-Kamlesh
|
|
|
|
EDV [328 posts] |
11 year
|
|
But key image size 80x80 is optimal for object recognition with AVM. Save your current recognition data and try to set 80x80 key size and also you can try to increase camera resolution. I hope it could help.
|
|
This Thread
