Ming Chen

Player-tracking data drives modern team analysis, but optical tracking systems are installed only in professional arenas and their data is rarely public. Broadcast video exists for almost every game. I wanted to find out how much reliable spatial data can be recovered from the broadcast alone.

The pipeline detects players and the ball, estimates the broadcast camera's pan, tilt and zoom frame by frame from the painted court lines, tracks each player, and projects every position onto a top-down court in meters with an attached uncertainty.

The biggest risk is a confident but wrong camera estimate, which silently moves every player on the court. I built a quality gate that grades each frame and withholds court coordinates when the fit is doubtful. It is tested on real failure cases, not only synthetic ones.

The system runs end to end on real broadcasts, including a 113-second multi-camera sequence. Keeping one ID per player through crossings is the main open problem. Human-reviewed annotations are in progress so that tracking accuracy can be measured rather than estimated.

Can two ECG recordings identify the same unseen person—even if they were captured at different times or with different devices?

Our team used a K2 score-level ensemble of three independently trained CNNs and one Raw GRU to compare 128-dimensional ECG embeddings.

I developed beat-CNN experiments, supported controlled architecture comparisons and cross-team validation, and investigated longitudinal morphology changes behind difficult matches.

Identified longitudinal changes in ECG waveform shape as a major challenge in matching recordings from the same patient over time.

Per-possession metrics such as on/off and RAPM depend on judgment calls that public numbers make silently: when a possession ends, how technical free throws count, who gets credit when a substitution happens between free throws.

I built a play-by-play database for 1996-97 → 2025-26 with two levels of validation, and possession logic driven by a versioned rules file. Possession counts are compared with pbpstats, a public reference, for every team and game.

From 2016-17 the share of identical counts fell from about 98% to about 78%. The cause was clock precision: reading time in whole seconds would restore the match, but would misclassify real possessions. I kept the exact time and documented the difference.

All 60 season files are loaded, with stints and possessions built and a read-only Streamlit dashboard on top. On/off, lineups, team efficiency, a matchup table and a RAPM prototype (with standard errors) are built; next: my own rule profile.