Getting started in OpenCV, recommendation for API functions to use

I'm just getting started with OpenCV and I'm planning to build a robot with computer vision. I am looking to make this robot recognize classes of objects as well as individual instances. In a sense, a Haar-like feature capability for general classes and BIGG for specific instances. I am essentially looking to 开发者_运维问答make something like this: http://www.youtube.com/watch?v=fQ59dXOo63o in the video, kinect is used, but I'll only be using a single camera. If you watch the video, you'll see that the kinect is shown an object and learns after a few seconds to recognize the new object. This is essentially what I want to do; instead of creating thousands of templates and training the software all at once, I want to make this process a semi manual one where the robot learns a single object at a time. I have no limitation on the type of object to be learned, everything is fair game.

Because I'm dealing with a potentially large amount of objects that will be trained, I'm worried about performance issues. If I have 10,000 objects trained, I would imagine that my laptop might choke on some of the algorithms. I am currently pretty overwhelmed by all the different techniques that the documentation has and I've little idea of what do use.

How would you guys tackle this problem?

thanks

Here are the questions you've asked about (whether you realize it or not):

• Object detection
• Object classification
• Object recognition
• Segmentation
• Normalization
• Machine learning

Each is an entire subject unto itself, and there is no "right" answer for your needs. You need to experiment and find that magical combination of algorithms that works well for your problem domain.

Also, The kinect has an advantage that a normal camera doesn't, which is depth. Plain old 2D recognition is ridiculously hard.

However in the spirit of giving a useful answer, check out the V1 algorithm by Nicolas Pinto which simulates the object detection capabilities of humans.

http://pinto.scripts.mit.edu/Code/Code