OVERWATCH adds self-calibrating cross-camera tracking

Homography is the right cheap hack here: it will never solve arbitrary camera geometry, but for ground-plane scenes it buys a lot of usefulness without blowing the Jetson budget.

• –The self-calibration loop is the practical part: shared foot points, RANSAC, periodic relearning, and reprojection-error flushing are enough to survive minor camera drift in real deployments.
• –The three-path fallback design is stronger than the homography itself, because it degrades gracefully instead of turning a bad calibration into a hard failure.
• –HSV-based re-ID is the obvious weak link; it is lightweight, but similar clothing, occlusion, and lighting changes will cause identity swaps well before the tracking stack runs out of compute.
• –The singleton pipeline and broadcast snapshot model are smart for edge hardware: more viewers do not mean more GPU inference, which is the right scaling choice here.
• –The biggest limitation is structural, not tactical: once cameras are steeply elevated or the scene stops being planar, a learned 3D/world model will outperform ground-plane projection.