How to directly process ROI of preview data and display the processed data on screen?

edit

Another approach:

// Assume "box" is a cvRect variable with the ROI dimensions

// Allocate space for a single-channel ROI (to store the grayscale ROI)
IplImage* gray_roi = cvCreateImage(cvSize(box.width, box.height), IPL_DEPTH_8U, 1);

// Allocate space for a temporary 3-channel image
IplImage* rgb_roi = cvCreateImage(cvSize(box.width, box.height), IPL_DEPTH_8U, 3);

// vid_frame is the original image. Let's set the ROI and perform a grayscale conversion.
cvSetImageROI(vid_frame, box);
cvCvtColor(vid_frame, gray_roi, CV_BGR2GRAY);

// Copy the processed image back to the original image
cvCvtColor(gray_roi, rgb_roi, CV_GRAY2BGR);
cvCopy(rgb_roi, vid_frame, NULL);

// Now reset the ROI so cvShowImage displays the full image
cvResetImageROI(vid_frame);
cvShowImage("Example", vid_frame);

This solution is inspired by the Sepia filter that you can find in an OpenCV4Android Sample "Image Manipulations".

It might be a bit slow, because it does more processing than it should for a grayscale conversion, but it does not need to make image copies like other solutions.

// Because you mentioned changing parameter from yuv to rgba
// you don't need to change it, you can convert it like this:
Imgproc.cvtColor(mYuv, mRgba, Imgproc.COLOR_YUV420sp2RGB, 3);

// define some rectangle to be the roi
Rec roi = new Rect( arguments );

// This kernel tries to apply this formula:
// gray pixel = (red + green + blue) / 3 =>  each channel has a weight of 1/3 = 0.333%
mGrayKernel = new Mat(3, 3, CvType.CV_32F);
mGrayKernel.put(0, 0, /* R */0.333f, 0.333f, 0.333f);
mGrayKernel.put(1, 0, /* G */0.333f, 0.333f, 0.333f);
mGrayKernel.put(2, 0, /* B */0.333f, 0.333f, 0.333f);

Core.transform(mRgba.submat(roi), mRgba.submat(roi), mGrayKernel);

Let us know if you have any problem with this, or if you prefer @karlphillip's solution in Java.