Are you doing your vision tracking on color images or depth map images from a device such as the Kinect or similar? I got a reliable boulder tracker working from a depth image using OpenCV on a Jetson. Finding the center of the boulder requires finding the moments of the contour in the image using this software, or a function like minEnclosingCircle. My code from 2014 could prove useful:
https://github.com/rr1706/vision2014...r/src/ball.cpp