If you know camera fov you could just set a fixed radius around the bright spot to look for hough circles . If not you could iterate an expanding radius and apply a mask to identify when the object ends, and then run a hough transform on that iterated area.
But because you are not running this real time, you should look into image segmentation and YOLO, which will give the centerpoint of each detected ball (but typically not its circumference). You will have to train your own image database of balls or pull a generic one from a public source like imagenet.