Electronic archery scoring: Danielescu 2021 study

The paper describes a prototype electronic scoring system with two cameras mounted on a frame around the boss, one above and one at the side, about 18 cm from the edge of the face and 10 cm in front of it. The software, written in C++ with the OpenCV library, finds the face and its rings in the images, locates the arrow by comparing the pictures taken before and after the shot, and awards the value of the innermost ring touched, as the rules require when an arrow touches a line.

Out of 16 arrows, the top camera scored 14 correctly (87.5%) and the side camera 11 (68.75%); taking the higher of the two values for each arrow, the overall result stays at 14 out of 16. The author puts the errors down to poor calibration of the side camera and estimates that fixing it would give at least 93.75%. By way of comparison she reports, from an email by the chair of USA Archery's Officials and Rules Committee, that judges are called to about 30 arrows out of 18,000 in a tournament, and from this derives a maximum human error of 0.17%.

The procedure has several steps. First the cameras are calibrated, the empty face is photographed and the user clicks on the centre of each face. The software finds the edges of the rings with the Canny algorithm, fits ellipses to them and builds a mask that excludes everything outside the face; a bilateral filter and erosion and dilation remove part of the noise. After each arrow the user presses a button, the software subtracts the previous image from the new one, isolates the shaft with a vertical structuring element and looks for the point of the arrow closest to the centre. It then applies the ring masks from largest to smallest: the last ring that still contains the point gives the score.

The system assumes a face with a white background, rings distinguishable by colour or by an edge, and an X at the centre. The work was done in the summer of 2007, during a research internship at Iowa State University, and posted on arXiv in 2021.

The test is minimal: 16 arrows, one face of one size, one session, with nothing on distance, lighting or how many arrows landed near the lines, which are the difficult cases. The comparison with judges does not hold up: the 0.17% is not a measured error but the share of arrows on which a judge is called, and in competition the archers read their own scores at the target, with the judge stepping in only on doubtful arrows.

The author herself lists the problems: the cameras focused poorly, changing light confused the rings, and the edge detection left precisely the inner rings incomplete, the ones worth the most. The cameras see the face at an angle, as an ellipse, and the transformation that would turn it back into a circle is only sketched out. The text has not been peer reviewed and should be read as a feasibility study from almost twenty years ago.

For archers today the study is mainly of historical interest. It shows why automatic scoring with cameras is hard: arrows on the lines, light that changes through the day, faces seen at an angle and small inner rings are the same problems any system must solve to be accepted in competition. At 87.5% accuracy the prototype misread two arrows in sixteen, and the author herself concludes that changes are needed to come close to the accuracy of a judge.

For anyone thinking of using a camera or an app to record their own shots in training, the paper suggests that the quality of the camera and its calibration matters at least as much as the software: in the prototype the errors came from there, and the inner rings, the smallest, were the hardest to detect.