Automatic arrow scoring: Parag 2017 study

A master's thesis in computer science at Leiden University, carried out at the request of the Dutch archery federation (NHB). The federation wanted a cheap, portable system that would record the score and position of every arrow without archers or helpers having to write them down after each end, and without losing the order of the shots. The programme, written in Java with the OpenCV library, works on video from a camera placed close to the target and at an angle: it recognises the face from its blue, red and yellow rings, corrects the perspective, picks out each new arrow by comparing frames and assigns it a score and a position.

On the six videos available the target was recognised every time, and 29 of 32 arrows were given the correct score. The three failures have specific causes. One arrow arrived during a sharp change in light, added to the video on purpose, and was rejected by the safety filter; another looked too short in the image to be recognised as a shaft; a third moved the paper as it went in, and the programme extended the arrow's line until it scored 9.85 instead of a 10. The decimal part of the score was also sometimes slightly off, because the programme measured from the outer side of the shaft rather than the inner side, as the federation requires.

The material comes from NHB training sessions: six videos and 14 images, used to tune all of the programme's parameters. Four videos were unedited, and in two the author added noise or a drop in light with editing software; two were filmed with a Logitech C920 PRO HD webcam and two with an unidentified, sharper camera. The background of the videos, with trees and moving branches, shows that at least some of the shooting was outdoors.

The face is found with colour ranges in HSV space on the blue, red and yellow, then straightened with a homography and rotated until the rings look as circular as possible. New arrows are looked for in the difference between one frame and an earlier one, using the Hough line transform; when more than 20 lines are found the frame is discarded as noise. The score depends on which ring contains the arrow's entry point, and the position is converted into coordinates that archers and spectators can read. White and black are not handled, because their colours get confused with the background and with each other: the programme only assigns scores from 10 to 5.

The author says so openly: with no new videos available in time, he tested the programme on the same six videos used to tune it, and asks readers to treat the 29 out of 32 with caution. A system tuned on certain images tends to work well on exactly those; how it would perform on another range, in different light, with another camera and another face, remains to be seen. There are few arrows, the shooting distance and bow type are not given, and the thesis does not say what reference was used to decide the correct score.

The programme also cannot see arrows in the white and black rings, and cannot work out the orientation of the face, because the numbers printed on the paper are not recognised. It is a thesis, not peer reviewed, and the author presents it as a prototype to build on.

The thesis says nothing about shooting itself, but it describes clearly what automatic scoring with a camera aimed at the target depends on. The difficulties are ones anyone who shoots outdoors will recognise: clouds suddenly changing the light, arrow shadows on the paper, branches moving in the background, paper shifting on impact. One of the remedies suggested has nothing to do with software: stretching the face tightly on the boss, because paper that moves lengthens the image of the arrow and pushes it into the next ring. The Dutch federation's aim also matters, which was to keep the order and position of every arrow as well as the total: entering scores by hand after each end, as the federation did, loses that information, and recovering it was the reason the federation wanted such a system.