Pose Pentathlon

Disclaimer: extremely vibe coded. This is for a one-off event. The foundation (video stream capture + rtmlib pose detection) was taken from a previous project - done cleanly, everything on top is quick AI code.

Interactive open-day exhibit: a single webcam runs real-time 2D pose detection (RTMO / COCO-17). Students see their own skeleton mirrored live, then an operator runs them through a 5-event athletic mini-pentathlon for a composite score and athlete title.

Design doc: docs/POSE_PENTATHLON_PLAN.md.

Quickstart

conda create -n pentathlon python=3.10 -y
conda run -n pentathlon pip install -r requirements.txt
conda run -n pentathlon python main.py

Requires a CUDA-capable GPU (onnxruntime-gpu, pinned to 1.20.1).

Controls

Key	Action
SPACE	Start circuit (from ATTRACT) / advance / next student
N	Skip current activity
R	Abort, reset to ATTRACT
Q	Quit

Status

• M1 — Mirror + ATTRACT (done) — live mirrored skeleton + pulsing "STEP IN" prompt
• M2 — Display layer + state-machine spine (done) — UI primitives + screen renderers (instructions / countdown / HUD / transition / results); Circuit state machine; full ATTRACT -> ... -> RESULTS flow walkable
• M3 — Formal Activity ABC (skipped) — duck typing covers it; see CLAUDE.md
• M4 — Real events: High Knees + Vertical Jump (MVP) (done) — hysteresis-based rep counting, hip-centre jump tracking, per-event skeleton highlight, vertical power bars with persistent peak markers. Score endpoints in config.py likely need lab-tuning.
• M5 — Reaction Wall (done) — two simultaneous target circles in a reachable zone, wrist-keypoint hit detection with tunable HIT_RADIUS_MULT, hit count scored via cfg.SCORE_MAP["reaction_wall"]. Circuit composition is now driven by cfg.CIRCUIT_ACTIVITIES, so individual events can be commented out for testing.
• M6 — Punch Power (done) — per-frame |horizontal wrist velocity| / arm_length (arm_length is a bend-invariant running mean of upper-arm + forearm). Velocity is the 5-frame MA from PoseDetector; a deadband zeros sub-jitter motion; live power bar with persistent peak marker. Originally paired with Javelin via a shared displacement-burst detector; Javelin deferred to M8.
• M7 — Stick the Landing (done) — replaces the original Javelin event as #5. 3-phase composite: (1) hold a single-leg stand for 3 s → static balance sub-score from multi-keypoint sway across elbows + wrists + hips + knees (flailing arms tank the score), (2) hop sideways to a glowing target ~0.75 leg-lengths away on the standing-leg side → landing-accuracy sub-score from hip-x distance to the target, (3) hold steady for 3 s after landing → landing-stability sub-score. Composite = 0.4 balance + 0.4 land-stability + 0.2 accuracy → 0..1000. Graceful degrade: phase timeouts zero that sub-score and continue. First multi-phase activity — introduces a _StickPhase enum + per-phase dispatch as a template for future composite events.
• M8 — Polish: instruction images, day leaderboard, optional sound; Javelin if time permits.

Project files

• main.py — capture/pose loop + circuit dispatch
• circuit.py — State, StubActivity, Circuit, build_demo_circuit
• activities.py — real event classes: HighKneesActivity, VerticalJumpActivity, ReactionWallActivity, PunchPowerActivity, StickTheLandingActivity
• ui.py — primitives (panel, progress_bar, power_bar, big_digit, text helpers) + screen renderers
• config.py — single source for tracking / UI sizing / state-machine timings / SCORE_MAP

Reused from rtmlib_vfx

• capture.py — CaptureThread (threaded non-blocking webcam capture)
• pose.py — PoseDetector, OneEuroFilter

Copied verbatim except for two lines stripped from pose.py (a Warp CUDA sync that's redundant when ONNX is the only GPU consumer).