This week I did a lot of work on the object recognition part, to really explore the capabilities of OpenCV and finalize our design for this part of the project. As a group we also spent a lot of time doing the design report.

I decided that the best way to do fast object recognition for this project is based on the geometrical differences between the 3 different types of objects. In experimenting with online images, I found that edge detection works pretty well because the differences between a person and their environment, or a car and their environment, etc. tend to be much more different color wise than the difference between a typical background. So with a bit of blurring it barely picks up the background when finding the edges. I also explored SIFT key points, which allow you to capture the important (scale invariant) features of an image, and once the noise is removed via blurring and edge detection, it really allows you to capture the geometrical differences between the images. All of these operations are blazing fast in OpenCV, and I figure that by extracting some features (as many as we have time for without breaking our requirement), and using them for machine learning, this algorithm will work quite well for our case of object recognition, and be pretty speedy. More details and sample images are in the design doc.