Acceleration sensors in archery: Eckelt 2020 study

This is a feasibility test, presented at a sports engineering conference, asking a simple question: is an accelerometer on each hand enough, with raw data and no processing, to build a feedback system for an archer? Six archers each shot 24 arrows at 18 metres with a three-axis sensor taped to each hand, and the hits were sorted into three categories: tens (the gold centre and the first ring), nines (the second gold ring) and everything else.

Three quantities were computed from the raw data. The first is the duration of the movement, from raising the bow to the start of the release: the shortest median belongs to the tens, but the worst shots are not the longest, since the highest median belongs to the nines. The second is the range of motion of the acceleration, that is the difference between maximum and minimum, computed over ten windows of 20 milliseconds each across the last 200 milliseconds before the release. On the bow hand the medians of the three categories resemble one another, with the only visible change on the x axis in the final 20 milliseconds; on the drawing hand the values are higher and shift more from one window to the next. The third is the coefficient of variation, the ratio of standard deviation to mean: on the bow hand it changes little from one window to the next except on the y axis, and on the x and y axes the lowest value belongs, oddly, to the worst shots.

The authors' conclusion is negative on two of the three quantities and cautious on the third: the coefficient of variation, on every axis and on the resultant acceleration, does not look like the right method, while the range of motion on single axes looks promising. A doubt and an open question remain: across the last 200 milliseconds the range was not lowest for the best shots in the window furthest from the release, and the width of the windows, set at 20 milliseconds, still needs discussing.

The measurements were taken over two days at the Vienna Indoor Archery Center. The participants were six archers of different standards: two classed as elite, three as amateur and one as recreational, four men and two women, three with a left bow arm and three with a right. Two of them were available on both days and so were measured twice. Consent was given orally and no ethical approval was sought.

On each hand a three-axis MMA7341LC accelerometer with a range of plus or minus 3 g was fixed with medical double-sided tape and then further secured with tape. The sensor's origin sat at the level of the proximal phalanx of the third finger and its y axis pointed towards that finger. The signals from all axes were digitised with an NI USB-6008 acquisition card at 1,000 samples per second and stored with a LabVIEW routine.

Each archer shot four sets of six arrows at 18 metres on a vertical FITA three-spot face. The data were neither filtered nor integrated to obtain velocity or displacement, since the point of the test was to use the raw signal. The start of the movement was identified on the y axis of the bow-hand sensor: while nocking, the arm points at the ground and the acceleration on that axis is about plus 1 g; when the bow arm rises above the horizontal the signal crosses the baseline, and that instant is the start. The end was identified on the y axis of the drawing-hand sensor, where the sudden movement at the beginning of the release shows up even in the unfiltered signal. Between the two instants comes the duration of the movement.

Range of motion and coefficient of variation were computed for each axis and for the resultant acceleration of both hands, over ten consecutive windows of 20 milliseconds covering the 200 milliseconds before the end of the movement, numbered from w1 (the last 20 milliseconds) to w10 (from 200 to 180). The choice of 200 milliseconds comes from a study giving that figure as a typical decision time. All results were grouped by hit category. No statistical analysis was carried out: the results are presented as plots of medians.

The main limit is the one the authors state in the first line of their discussion: no statistical evaluation was carried out, so the results can only point to possible trends. There are no numbers in the text, only plots of medians, and this record cannot report values because the paper does not print them.

The sample is six people, with 24 arrows each and two of them measured on two days, so about 192 shots. But size is not the problem: the way the shots were grouped is. The three categories pool the arrows of all six archers, and those six range from recreational to elite. The tens category is therefore filled mostly by the elite archers and the remainder mostly by those who shoot less well, so any difference between categories may be a difference between people rather than between shots. With that structure, one elite archer who shoots quickly with a steady hand produces the duration result on his own.

The outcome measure is coarse too: three categories, tens, nines and everything else, with no coordinates and no distance from the centre, on a three-spot face at 18 metres. The sensor's range, plus or minus 3 g, is narrow for a release, and a signal neither filtered nor integrated carries all the noise of the acquisition chain with it; the double-sided tape on the hand, in turn, interferes with one of the most sensitive parts of the movement, and nobody checked whether the sensor changed the shot. Putting the origin of the y axis on the third finger means the axes of both sensors are oriented to the hand rather than to the target, so comparing x, y and z depends on how the archer holds their hands.

Finally the picture does not close. The authors acknowledge that the range was not lowest for the best shots in the window furthest from the release, the moment their own premise had marked as decisive, and that it is therefore unclear whether one accelerometer per hand can be enough. The window width of 20 milliseconds is a choice they question themselves. It is a six-page conference paper with no stated peer review, to be read as a first attempt rather than as a result.

For an archer this work is not a result about shooting: it is a survey of a tool. The question it asks, whether an inexpensive sensor on each hand can tell an archer why a particular arrow missed the centre, got a partial and half negative answer.

The negative half is worth knowing because it touches a widespread idea, that of judging stability by how regular a signal is. The coefficient of variation of the acceleration, on every axis and on the resultant, did not separate the hit categories, and on the bow hand's x and y axes the lowest value even turned up on the worst shots. The range of motion, the span between maximum and minimum acceleration on a single axis, looks more promising, but even there the authors stop short of calling it usable.

The other pointer concerns the duration of the shot. Tens had the shortest median from raising the bow to the release, but the relation is not monotonic: the worst shots were not the slowest, since the longest median belonged to the nines. The study therefore does not allow duration to be read as an index of quality, and the authors put part of the discrepancy down to where they chose to start the clock, before the aiming phase. Above all, bear in mind that the differences between hit categories are entangled with the differences between the six archers, who ranged from recreational to elite.