To identify round (or any other) shape which interlaced or "connected" with background, as it shown on the posted image, would require pattern recognition algorithm, cross-correlation what I'd think of, and there is no time to do it. All you have is about 1.2 millisecond during 20 lines in the beginning of new frame, and that makes task impossible, especially for varying size of the object depends on distance from the camera.
Edited:
Better way is comparison between pre-stored image w/o ball (desk & net), and image with it, than algorithm would outputs X Y coordinate where "activity - movement" has been detected. But even with this "simplified" - no shape tracking task, math would take too mach time. Giving one compare or subtract 2 cycles operation, 16 MHz cpu can do 8 op/usec, 8000 op/millisec,. image size 300x300 = 90 000.
The best chance to fulfill a mission is make a ball "only" one object in video frame, I mean create a "sharp" contrast, that only the image of the ball would get in the video buffer. Black background or colorful - yellow filter in front of cam definitely would help to increase a contrast.