I want to optimize my program, in which I am using color object tracking algorithm described here. The only difference is that I am using cvBlob library, instead of cv::moments (cvBlob was faster and more accurate). Using profiler (valgrind + kcachegrind) I have found that ~29% of time is taken by colorspace conversion method (cv::cvtColor; I am tracking objects in three colors). I am converting from BGR to HSV.
I've read in some papers that using YCbCr colorspace is even better in color tracking. Is it good idea to convert from BGR to YCbCR? It should be slightly faster, as it requires less multiplications (I am not sure about that -- I do not know how OpenCv does it internally). Does this algorithm need some changes, or can I just convert lower and upper boundaries for tracked color from HSV to YCbCr, and then use inRangeS method, as I did with HSV?
Is there any way to get the frame from driver in YcbCr (or YUV)? I am not asking about HSV, because this is not supported by v4l2, AFAIR.
Do you have any other ideas? I don't want to use IPP or GPU.
