Broadcast Video to Court Coordinates on a Consumer GPU -- built from primitives

Broadcast video comes in; court coordinates and behavioral features come out. The pipeline runs end-to-end on one consumer GPU (RTX 4060, 8GB) and writes per-track court positions plus behavioral fields to data/tracking_data.csv. What makes it an engineering artifact rather than a library demo is that the load-bearing pieces are built from primitives:

- Multi-object tracking from scratch. A 6D constant-velocity Kalman filter for motion prediction, and the Hungarian algorithm for globally-optimal frame-to-frame ID assignment over a blended IoU-plus-appearance cost, with a greedy fallback. Not a wrapped tracker. - A custom-trained ball detector. YOLOv8n fine-tuned to a single ball class, then exported across PyTorch, ONNX, and TensorRT for deployment. - Court homography from classical CV. HSV masking, HoughLinesP, line-intersection cornering, and getPerspectiveTransform recover the camera-to-court matrix per clip, with a static-matrix fallback and unit tests on synthetic courts. - Broadcast hardening. A SIFT three-tier strategy with inlier gating, EMA smoothing, drift re-anchoring, and replay/scene-cut suspension so player trajectories are not corrupted during graphics and replays. A two-frame confirmation gate is unit-tested. - OSNet re-ID reimplemented in PyTorch -- omni-scale blocks and depthwise-separable convolutions -- behind a layered inference backend (TensorRT to torchreid to standalone to MobileNetV2 to HSV histograms), so it runs whether or not the accelerated deps are present. - A feature layer hardened against silent corruption -- pixel-vs-feet auto-rescale, physical-validity caps, phantom-slot filtering, and roughly ten documented sentinel-leak fixes, each guard tied to a specific observed broadcast artifact.

The end result at scale: anonymous tracker slots resolved to real NBA player identities across 240-plus games -- 17,254 `cv_features` rows spanning 241 games and 252 distinct real NBA player IDs (counted directly from the local database this session).

Two things, and the second is the point. First, the CV engineering: implementing Kalman filtering, the Hungarian assignment, a classical-CV homography recovery, and an omni-scale re-ID architecture from primitives -- and hardening all of it against the silent-corruption failure modes broadcast footage is prone to -- is mid-level CV-engineer work done solo. Every guard in the feature layer traces to a specific broadcast artifact that was observed and diagnosed.

Second, the honesty discipline. The easy version of this page would claim a re-ID accuracy, a positional precision in inches, a ten-player frame rate, and a CV-feature edge. None survive scrutiny, so none appear here. The pipeline outputs court coordinates and resolves real player identities at real scale; the accuracy of those coordinates is not yet benchmarked against ground truth, and the CV features do not yet move the prediction model. Stating exactly that -- what works, and what is built-but-unproven -- is the same discipline that runs through the rest of this evidence layer.