Archers' heart rate and tremor from video: An 2025

The idea is a good one and worth grasping before the faults: measuring the heart rate and tremor of elite archers without touching them, taking the data from television coverage of competitions. Heart rate is estimated from the colour of the skin of the face, because blood passing through absorbs light differently, and arm movement is measured by tracking feature points in the image after recognising the figure and its skeleton with a neural network.

The material is nine world championship videos taken from YouTube, from which, after exclusions, fifty ten-second segments at twenty-five frames a second remain. Each segment carries a score, and the scores are grouped into three categories: ten, nine, eight or less, with twenty, twenty and ten segments.

The results all run the same way. The estimated heart rate is linked to the score with a rank correlation of minus 0.618, which is a lot, and in the distributions the arrows worth ten sit around seventy and eighty beats a minute while the others spread between ninety and a hundred and ten. Right-arm movement, that is the string arm for a right-handed archer, is linked to the score at minus 0.393; the bow arm is linked to nothing, minus 0.136 with p equal to 0.345. In the ordinal regression heart rate and right arm stay significant, including together, and the two measures are almost independent of each other, with linear correlations between 0.11 and 0.21.

So far the picture is coherent and interesting. The problem is that almost none of it rests on a verified measurement: the heart rate is never compared with a monitor, not even on one person; the movement is in pixels, a unit that changes with the camera's zoom; the unit of analysis is the segment and not the athlete, and how many athletes there are is not written.

And there is one sentence, in the discussion and in the conclusions, that on its own says how far the data were handled: reducing tremor in the dominant arm by one unit, one reads, would gain more than six points. The score in this study has three categories between eight and ten: six points do not fit inside.

The route is declared with a flowchart following the STROBE guidelines, which deserves acknowledgement: nine YouTube videos to start, then background-only scenes, blurred or low-quality footage, clips without face or body visible and clips with camera movement are excluded, and fifty valid segments remain.

For heart rate: the face is located and tracked with a detection model, skin pixels are isolated by filtering in the YCbCr colour space, with blue chrominance between 133 and 177 and red between 77 and 127, and the pulsatile signal is derived from the periodic variations in colour. Only values with a confidence score of 0.7 or above are kept. For movement: object recognition with YOLO version 8, skeleton extraction, definition of right and left arm regions, selection of the feature points nearest those regions, and tracking of the coordinates with displacement computed between consecutive frames. Micro-movements within a five-pixel threshold are considered. For each segment the mean is taken.

The statistics: normality with Shapiro-Wilk, homogeneity of variance with Levene, and then an independent-samples t-test or a Mann-Whitney test according to the assumptions, to compare the ten-point group with all the others. Variance inflation factors for collinearity, all close to one. Pearson correlations, which come out low between heart rate and movement, and Spearman's, used for the link with the score. Ordinal logistic regression with the score as a three-category outcome and heart rate and movement as continuous predictors. All in Python with SciPy and statsmodels.

The sample size is justified in two ways. The first is precedent: others have worked with thirty-two points and with eleven. The second is a power analysis conducted after the fact using the coefficient of the movement variable from the full model, which gives a power of 0.9486.

Arm movement, one reads, is measured from the start of the aiming phase to the release of the string, while the means are computed across the whole ten-second segment: whether the moment of release falls inside or outside the averaging window is not clarified.

The fault to start from is the easiest to check. In the discussion one reads that the heart rate coefficient, minus 0.1166, indicates that one beat less corresponds to about 0.12 points more, and that the right-arm coefficient, minus 6.1747, suggests reducing tremor by one unit would gain more than six points. In an ordinal logistic regression the coefficient is a change in the logarithm of the relative odds, not in points: minus 0.1166 per beat means each beat less multiplies by about 1.12 the odds of landing in a higher category. And the outcome has three categories between eight and ten, a range of two points: no predictor can move it by six. The sentence is not a slip of language, it is the way the article translates its own numbers into training advice, and it travels from the discussion into the conclusions.

The second is the unit of analysis. Fifty segments from nine videos, and how many different archers are inside them is written nowhere. If the segments come from few athletes, the observations are not independent and a regression treating them as fifty separate cases overstates precision; and above all the link between heart rate and score becomes largely a comparison between people rather than within one person. Archers who shoot tens more often are also the better trained and the calmer, so a lower heart rate and a higher score may be two consequences of the same level. No mixed-effects model, no clustering by athlete.

The third is the sampling, and it comes out of the exclusion criteria. Blurred footage, footage without face or body visible and footage with a moving camera were excluded: what remains is what the television director chose to frame closely and hold still. The close-up on an archer's face, in a live broadcast, arrives in moments of tension and after the shots that matter. The dataset is not a sample of shots, it is a selection made by television on criteria that have to do with both the score and the tension.

The fourth is the heart rate measurement, which is the main variable and is never verified. The pulsatile signal from facial colour needs stable lighting, a still face and little compression, and broadcast video is compressed with chroma subsampling, that is precisely on the colour variations the method reads. The only check is a confidence threshold of 0.7 from an undescribed model. There is no comparison with an electrocardiogram, a chest strap or a pulse oximeter, not even on one subject under controlled conditions. And heart rate variability appears in the keywords, while no variability index is computed anywhere in the article.

The fifth is the movement measurement, and here the problem is the unit. The displacements are in pixels, with a five-pixel threshold, and a pixel is not a millimetre: how large an arm appears depends on the camera's zoom and distance, which change from shot to shot and from broadcast to broadcast. Without a calibration, or at least a normalisation for apparent body size, the variable mixes tremor with optics. For a comparison between different segments it is a serious difficulty, and the article does not name it.

The sixth concerns the interpretation of the asymmetry between the arms, which is the most quotable result and has an alternative reading the article does not consider. In a television close-up the string arm and the hand are near the face and move visibly at release, while the bow arm is extended, often at the edge of the frame and partly hidden by the bow: the asymmetry of the result may be an asymmetry of visibility. And if the moment of release falls inside the window over which the mean is taken, right-arm movement is not measuring tremor during aiming but the act of release, which is another thing and which might link to the score for different reasons.

The seventh is the power analysis. It is computed after the fact on the observed effect, and power computed that way is a function of the p value: when the result is significant, it always comes out high. It says nothing about whether fifty observations suffice, and the same holds for the other argument used, that other studies have used thirty-two or eleven data points.

Two smaller things remain and one of substance. The same data are analysed twice, with three categories in the regression and with the comparison between ten and not-ten in the group tests, and the two results are presented together in the abstract with no correction. Of the shooting conditions nothing is known: distance, phase of the competition, wind, indoor or outdoor, men or women, nothing is reported, although a ten at seventy metres in wind and a ten indoors are not the same event. And the conclusions, on fifty observations with no intervention at all, recommend breathing, meditation and biofeedback as things that may directly improve competitive outcomes.

What this work really shows is a method, and the method is promising: an archer's heart rate and movement can be estimated from television footage, with nothing attached to them. For a coach it is the prospect of watching an athlete in competition and getting numbers out of it, instead of waiting for a laboratory. The road is worth following, and it is worth knowing it has not arrived yet.

The figure to take away, carefully, is the difference between the groups: in the arrows worth ten the estimated heart rate sits around seventy and eighty a minute, in the others it spreads between ninety and a hundred and ten. It is a large difference and in the direction everyone expects. But mind what it compares: different segments of probably different archers, so it can say that the calmer archer is also the one shooting tens, without saying that lowering the heart rate makes one shoot tens.

The second interesting thing is that heart rate and arm movement, in these data, are almost independent, with linear correlations between 0.11 and 0.21, and both stay linked to the score when they enter the model together. If the datum holds, it means that being agitated and trembling are not the same thing and are not corrected by the same work.

What should not be taken from here is any quantity at all. The movement is in pixels, so it depends on the zoom; the heart rate is estimated and never checked against an instrument; and the step from coefficient to points the article proposes, six points gained by reducing tremor by one unit, makes no sense on a scale running from eight to ten. None of these numbers can be used as a reference.

Nor should the practical advice be taken as written. That breathing, meditation and biofeedback improve the score is a conclusion requiring an intervention, that is a group doing them and a group not doing them, and here nobody did anything: these are fifty video segments observed after the fact.

For an archer the honest way to use this article is as a reminder of something already known, that one shoots better with a low heart rate, and as notice of a technology to keep an eye on. For anyone wanting to try it seriously, the three things missing here are the list: check the estimated heart rate against a strap, convert the pixels into millimetres with a reference in the scene, and count athletes rather than segments.