Posts
The Camera's Pulse
// reading a rider from the broadcast feed instead of a data leak
Strip the idea of "reading a rider from the broadcast feed" down to its most boring version, and it stops being speculative at all. Tracking cadence, out-of-saddle position, or a subtle change in how smoothly someone is pedaling is a kinematic problem — motion and position over time — and that's exactly what pose estimation and optical flow already do well in other sports. Golf swing analysis and gait labs solved versions of this problem years ago. Nothing about a cyclist's leg moving in a repeating pattern is harder than a golfer's swing plane. If any team or broadcaster wanted to build a cadence-reading tool off existing race footage today, the computer vision to do it already exists, off the shelf, in adjacent industries.
Reading heart rate from a face on video is a different, much harder claim — and a real one, not invented for this series. Remote photoplethysmography works by detecting subtle, otherwise invisible shifts in skin color caused by blood flow with each heartbeat, extracted from ordinary video. It's an active academic field with a real body of published research behind it, not a speculative leap.
The problem is exactly the condition a mountain stage guarantees. rPPG's accuracy depends heavily on the quality of the video and how much the subject's head is moving, and demonstrated results for a moving subject top out around 15 km/h running, with roughly 1.8 bpm of error — in a controlled lab, not a helicopter shot panning across a peloton grinding uphill at 25 to 40 km/h, in a helmet, in sunglasses, sweating, surrounded by other riders. Every one of those conditions works against the signal rPPG depends on. Recent research explicitly frames "motion-robust" rPPG as the open problem it's trying to solve, using techniques like optical flow correction — which tells you plainly that it isn't solved yet.
Put the two technologies back side by side and they land in genuinely different places. Kinematic reading — cadence, position, the visual "easy gear at 190bpm" bluff this series opened with — is achievable now, with existing tools, by anyone motivated to build it. Physiological reading — heart rate from broadcast video — is a real research trajectory, moving toward motion robustness, but not race-usable today by any evidence we can find. If a WorldTour team or a broadcaster is doing visual deception analytics at all right now, the kinematic version is the far more plausible bet. The physiological version is a story about where the field is headed, not where it currently stands.
The maturity of pose estimation and optical flow in adjacent sports, and rPPG's documented motion-artifact limitations, are both independently sourced below. That kinematic cadence-reading is cycling's "most plausible" current vector is our inference from that maturity — no source confirms any team or broadcaster is actually doing it. The claim that rPPG is not yet race-usable is a direct reading of the cited research's own stated limitations, not our extrapolation.
rPPG's sensitivity to video quality and head motion, and current research explicitly pursuing motion-robust methods via optical flow correction, are drawn from IEEE and arXiv publications on motion-robust remote photoplethysmography, treated as Tier 1. Demonstrated accuracy for a moving subject at approximately 15 km/h running with roughly 1.8 bpm error in controlled conditions is drawn from peer-reviewed rPPG performance research, treated as Tier 1. A documented limitation regarding rPPG's reduced reliability under intense exercise specifically is drawn from a peer-reviewed study on PPG motion artifacts, treated as Tier 1. A broader review of camera-based vital sign measurement is drawn from a Springer-published survey, treated as Tier 2 corroborating context. No source describing pose estimation or optical flow applied specifically to professional cycling broadcasts was found; that application is presented here as inference from adjacent-field maturity, not as a documented cycling-specific tool.
This series opened with a viral claim about digital twins and glycogen-burn prediction that didn't survive contact with how a power meter's data actually travels — and it closes, four posts later, with a genuine argument about where the next real edge is likely to come from. In between: a regulatory system that bans by name instead of by category, a prediction model that's been quietly running in Tour broadcasts for six years, a documented legal history of exactly the kind of deception the original claim imagined, and a hard technical line between what a camera can plausibly read today and what it can't yet.
The honest version of the story was less cinematic than the viral one, and more durable. Sensor regulation will keep lagging sensor shipping. Prediction, once public and unremarkable, diffuses into whoever wants it next. And the real competitive frontier, if this series' forecast holds, won't be a better model — it'll be whoever first understands their rival's model well enough to make it wrong on purpose.
No comments:
Post a Comment