Image Detection Question

[docmur]

Hey Guys

I want to build camera based image detection into a project, just for fun.  I've used image sensors in the past, but they were intensive resource wise and sucked power like it was free. 

I've also used web camera's in the past and used raw image processing, but again it was very resource intensive. 

What are some ways people have done this?  What I want to do in effect, is take pictures of my environment every X seconds and process them for edges and surfaces. 

I know I could grab a number of different image sensors or a web camera and just throw this together, but does anyone know of a more elegant solution?

Again this is just for fun, not a serious level project.

Thanks
Docmur

[helius]

http://opencv.org

[Kilrah]

https://www.adafruit.com/products/1906

[helius]

The Kinect uses photon time-of-flight sensors to build a 3d point cloud. Having depth data makes object reconstruction easier, but the device itself does not do feature recognition.

[CatalinaWOW]

I am not sure I know what you mean by elegant.  There are some pretty fundamental trades in object detection, and no real way around them that I know.  Edge detection, for example, requires an absolute minimum of comparing two pixels (and performs poorly at that minimum level).  So the only way to further reduce processing is to reduce pixel count.  If you know a priori the size of the objects you wish to detect you can match the resolution to that size and reduce pixel count.  Or you can reduce processing by only looking for changes around previously detected objects.  That leads to a whole family of solutions that does intense processing occasionally with delta processing on the frames between.  As you pointed out, reducing frame rate is an option.

The on chip processing approaches mentioned above don't fundamentally change any of this, they merely apply custom hardware to part of the solution and reduce the size of the data channel heading downstream. 

None of these solutions works in a general case, but can work quite well in a well defined problem space.  As an example the processing required to detect a red ball in a green field can be quite simple.  Detecting a single, moving, known size and contrasting object in a grey scale image is also relatively easy and computationally achievable.  Can you narrow down your objective to allow some form of simplification?

[docmur]

Thanks for all the input