Sports Camera Motion-Tracking System Development -- 3

We're building a system that works similarly to sports auto-tracking cameras (e.g. Veo), but with a simpler, classical computer-vision approach — no deep learning or trained models needed. What we need: Camera stitching: Combine footage from two fixed cameras (mounted on one rig, overlapping field of view) into a single panoramic image (~180°), using calibration/homography. The cameras don't move relative to each other, so this should be a one-time calibration applied per frame. Motion-density tracking: From the panoramic feed, detect where players are concentrated (background subtraction / foreground blob density, not per-object classification) and use that to drive an automatic pan/crop — i.e. a virtual camera that follows the action without a human operator. Output: a single, smooth, watchable video stream (no visible seam/ghosting at the stitch line, no jittery panning). What this is NOT: we are not looking for a deep-learning/ball-detection/player-identification system. Classical CV techniques (homography, background subtraction, centroid tracking, Kalman/low-pass smoothing) are the right toolset here — please don't propose a large AI/ML infrastructure buildout. Ideal candidate: Strong hands-on experience with OpenCV (stitching, homography, background subtraction) Has shipped at least one real-time video processing project (not just offline/batch) Comfortable working with two-camera / multi-camera rigs and frame synchronization Bonus: experience with sports or surveillance camera systems