Bow sensor to compare shots: Kawaguchi 2020 study

A group of computer scientists at the Tokyo Institute of Technology fitted a sensor to the bows of five recurve archers and built a programme that, without anyone watching the videos by hand, groups similar shots together and shows the instant at which two shots part company. The sensor measures the angular velocity of the bow, that is, how the bow rotates in three directions, presented in the programme as roll, pitch and yaw; shots are compared with dynamic time warping, which measures how alike two curves are even when they last different lengths of time, and then gathered into groups with the k-means method.

Between 104 and 177 analysable shots came from each archer, and the programme split them into eleven to sixteen groups for each of the three phases into which the shot is cut. The differences the system brought out last about half a second and are of three kinds: a continuous movement against the same movement made in two stages, a movement lasting more or less time, a movement made more or less quickly. In one archer the arm extended in two steps in one group of shots and in a single one in the other; in one woman the bow was raised higher than usual and the draw broke into two stages; in a third archer the fingers stayed hooked on the string at release, and those shots had more shake during aiming; in the same archer, in one group, the bow rotated at release at roughly twice the speed. In yet another archer, aiming lasted about 4 seconds on average in one group and about 2.5 seconds in the other, and the shorter group shook more.

The coach interviewed judged those differences relevant for working out where to improve, and the archers said they had not noticed them by themselves.

The five participants were three men and two women, all shooting recurve, ranging in level from intermediate to entry at the all-Japan championships; the authors define advanced as a personal best above 650 points in the double score at 70 and 50 metres and intermediate as around 600. They shot at 50 metres on a square target 82 centimetres a side, 110 to 180 arrows each, losing a few to sensor faults.

The sensor is a nine-axis inertial unit of 30 grams fixed to the bow, with a range of 300 degrees per second, sampled at 1,000 hertz; of all its data only the angular velocities are used. In parallel two cameras filmed at 300 frames per second, one from the side and one from behind, synchronised with the sensor. After each shot the score, the position of the arrow on the target, a spoken comment and the archer's own rating on seven levels, from very bad to very good, were recorded.

The signal is first cut into three phases: drawing, aiming (where the angular velocity about one axis stays between minus 2 and plus 4 degrees per second) and release (the 1.2 second window around the maximum). A filter then removes frequencies above 5 hertz, to take out the bow's own vibration. The number of groups for each phase was chosen with the elbow method. The programme, written as a Windows application, shows the comparison between two groups, the comparison between two single shots, the cost graph that marks where the two curves move apart, and a video in which the two shots are superimposed, one in blue and one in red, cut to the stretch where the difference appears. The differences then described in the paper were spotted by the first author watching those superimposed videos.

The sensor sits on the bow, and that decides everything else: the system sees the behaviour of the bow, not that of the body. An arm extending in two stages or a bow raised higher than usual reach the sensor as different shapes of the curve, and the authors themselves acknowledge that reading those shapes goes through the video. Movement is measured as rotation only, with no displacements, and the archers interviewed ask for exactly what is missing: the position of the shoulder, the drawing hand, the relation between the movement of the bow and that of the body.

Score and the archer's own rating were recorded shot by shot, but never related to the groups: it is not known whether the "good" groups score more. One participant says as much in the interview, that score, feel and shot shape sometimes do not go together, and that he would like to see that relation. The system, then, finds differences without ranking them by quality.

Five archers, no quantitative evaluation of the programme, and as a user study one interview with a coach and five comments: the judgement that the differences matter is one person's opinion. The person who identified the differences in the videos was the first author, alone, knowing which shots the programme had separated. One contradiction in the figures also remains: the sensor is stated as sampled at 1,000 hertz, but the paragraph on the filter gives a sampling frequency of 200 hertz, and the paper does not explain which applies to the analysis.

Anyone who films their own shooting to compare shots knows the work is long, and that with a hundred or two hundred arrows the pairwise comparisons are too many to make by hand. This study shows a way out: a sensor on the bow and an algorithm that holds similar shots together and marks the exact point where two shots move apart, which is then watched on video.

What comes out says something even to those without a sensor. In archers of a good level, over a hundred or so arrows, the differences that remain are not different positions but different timings: a movement in two steps instead of one continuous one, an aim of 4 seconds instead of 2.5, a quicker release. They last about half a second, and the archers had not noticed them; according to the coach interviewed, advanced archers and beginners do not notice them equally, and beginners need someone to read them.

There is no data on what those differences cost in points: score and arrow position were recorded, but never compared against the groups. So the study does not license saying that the group of shots with the longer aim is the better group, not even in the archer in whom that difference was found. The contribution is to the method of observation, not to a technique to prefer.