The phases of the shot can be recognised and timed automatically from video

The summary today

It can be done, at least in part, and the figures also show how coarse the measurement still is. The strongest result dates from 2024: on fifty-eight competition clips of world-class archers, filmed at thirty frames per second, a system based on pose estimation times the interval between anchor and release with a mean error of 0.34 seconds against an expert's manual measurement. It works because at anchor the angle of the drawing arm almost stops changing, and it only works if the arm is tracked: relying on the string near the face pushes the error to 1.4 seconds. In 2025 a classifier working on joint positions extracted from the videos of twenty archers recognises three phases with 90 per cent accuracy in recurve and 76 in barebow, and it fails exactly where posture changes least, between drawing and aiming and the expansion into the release. Before video, automatic segmentation came from sensors on the bow, and in 2025 it also comes from a wristband: related work, but it does not start from video. An error of a third of a second is enough to count the timings of a competition, not to judge an arrow, and nobody has yet compared these systems with several timekeepers or measured the error at the boundary of each phase.

Also phrased as: "automatic video analysis"; "artificial intelligence in archery"; "pose estimation"; "timing the shot from video"; "shot phase recognition"

Evidence

• 2020
  Context Archers Published later

  5 intermediate and advanced recurve archers with a gyroscope on the bow: the release phase is cut out automatically as the 1.2-second window in which angular velocity changes most, and differences in shake, posture and speed between shots by the same archer show up in windows of about 0.5 seconds. It is a sensor on the bow, not video.

  Archery shots visualization by clustering and comparing from angular velocities of bows · Midori Kawaguchi, Hironori Mitake, Shoichi Hasegawa

• 2021
  Context Archers Published later

  11 male university recurve archers, about 880 minutes of recording with an accelerometer on the bow: a decision tree recognises the shots with a precision of 0.97 and a recall of 0.94 when the moment of release is known, and 0.72 and 0.80 without that information, then isolates the aim to measure its tremor. Again it is a sensor, not video.

  Archery Skill Assessment Using an Acceleration Sensor · Takayuki Ogasawara, Hanako Fukamachi, Kenryu Aoyagi, Shiro Kumano, Hiroyoshi Togo, Koichiro Oka

• 2024
  For Archers Published later

  58 frontal video clips at 30 frames per second of world-class recurve archers in competition, from the Paris 2024 Olympics and Korean national team events: against an expert's manual measurement with Dartfish, the time from anchor to release is measured automatically with a mean absolute error of 0.34 seconds and R squared 0.95, and the start of the anchor with an error of 0.29 seconds and R squared 0.91.

  AI-Based Analysis of Archery Shooting Time from Anchoring to Release Using Pose Estimation and Computer Vision · Seungkeon Lee, Ji-Yeon Moon, Jinman Kim, Eui Chul Lee

• 2024
  For Archers Published later

  In the same study, the arm angle alone gives an error of 0.512 seconds (R squared 0.93) across the 58 clips and 0.54 on the 11 in which the string could not be detected, while the method based on the string near the face alone is off by 1.4 seconds (R squared 0.89), because it starts counting as soon as the bow enters the face box: the difference from the proposed method is significant (p below 0.05).

  AI-Based Analysis of Archery Shooting Time from Anchoring to Release Using Pose Estimation and Computer Vision · Seungkeon Lee, Ji-Yeon Moon, Jinman Kim, Eui Chul Lee

• 2024
  Context Archers Published later

  The mechanism that makes the time measurable: the anchor is detected when the mean change in the drawing-arm angle stays below one degree for ten consecutive frames, and the release when the moving average over ten frames exceeds five degrees. These are empirical thresholds, tuned on footage at 30 frames per second.

  AI-Based Analysis of Archery Shooting Time from Anchoring to Release Using Pose Estimation and Computer Vision · Seungkeon Lee, Ji-Yeon Moon, Jinman Kim, Eui Chul Lee

• 2025
  For Archers Published later

  Videos of 20 recurve and barebow archers, joint positions extracted with YOLOv8 and three classes, set-up, drawing and aiming, expansion and release: on the test set accuracy is 90 per cent with SVM and 87 with Random Forest in recurve, 76 and 75 in barebow.

  Pose Classification in Archery Sports Based on YoloV8 Using SVM and Random Forest Methods · Yuridikta Adha Muslim, Bedy Purnama, Bayu Erfianto

• 2025
  Context Archers Published later

  In the same study the confusion lies entirely between drawing and aiming and the expansion into the release, where posture changes very little: in barebow, with SVM, the drawing and aiming class has a recall of 39 per cent, while the set-up is the most recognisable phase in both styles, with an F1 of 98 per cent in recurve and 86 in barebow.

  Pose Classification in Archery Sports Based on YoloV8 Using SVM and Random Forest Methods · Yuridikta Adha Muslim, Bedy Purnama, Bayu Erfianto

• 2025
  Context Archers Published later

  6 adolescent professional archers, 4 girls and 2 boys with an average age of 15.5, in official competition, with an accelerometer on the bow-hand wrist: an LSTM network using total and differential acceleration recognises drawing, aiming and release with 95.99 per cent accuracy, against 84.53 with the raw axes alone. It is a wearable sensor, not video.

  A Machine Learning-Based Multimodal Framework for Wearable Sensor-Based Archery Action Recognition and Stress Estimation · Xianghe Liu, Jiajia Liu, Chuxian Xu, Minghan Wang, Hongbo Peng, Tao Sun, Jiaqi Xu

• 2025
  Context Archers Published later

  4 Korean national-level archers, pose estimation from video and sequence models: it shows how consistently each one repeats the paths of shoulders, elbows and wrists from arrow to arrow, with no numerical index of consistency and no link to the score. Not every model works: in recognising the archer a simple LSTM network stops at 55.9 per cent, against 98.7 for a Bi-GRU.

  Deep Learning Sequence Network for Identifying and Analyzing Archery Shooting Patterns · Jihoon Park, Hyongjun Choi

• 2026
  Context Archers Published later

  Eight annotated images of a single archer, one per phase of the shot, with pre-trained MediaPipe and no retraining: the normalised joint localisation error ranges from 0.55 at nocking to 0.61 at release, 0.63 at aim, 0.65 at hooking and drawing, 0.68 at set-up, 0.69 at stance and 0.71 at anchor. It does not get steadily worse as the phase becomes more dynamic.

  Accuracy Evaluation of 2D MediaPipe-Based Pose Estimation for Archery Posture Detection Using N-MPJPE · Muhammad Andhika Bayu Prasetya, Harits Ar Rosyid, M. Zainal Arifin

Status history

1. 2020 Not studied Automatic recognition of the phases starts with sensors, not video. With a gyroscope on the bow, in five intermediate and advanced recurve archers, the release phase is cut out automatically as the 1.2-second window in which angular velocity changes most, and the differences between one shot and the next by the same archer sit in windows of about half a second. On video there is nothing yet: not studied.
2. 2021 Not studied A single accelerometer on the bow, eleven university archers and about 880 minutes of recording: a decision tree recognises the shots with a precision of 0.97 and a recall of 0.94 when the moment of release is known, and 0.72 and 0.80 when it has to be found from scratch, and from there isolates the aim to measure its tremor. It confirms that the shot can be segmented automatically, but still from a sensor and not from video, so the status does not change.
3. 2024 Reported Video arrives. Fifty-eight competition clips of world-class archers, from the Paris Olympics and Korean national team events, filmed at thirty frames per second: the time from anchor to release is timed automatically with a mean error of 0.34 seconds against an expert's manual measurement, and the start of the anchor with an error of 0.29. The method works because at anchor the angle of the drawing arm almost stops changing and at release it moves again; relying on the string near the face alone pushes the error to 1.4 seconds, while the arm alone holds up even when the string cannot be seen. It is evidence on archers, but from a single study and on a single interval: reported.
4. 2025 Reported Three independent studies widen the picture, and also show where it breaks. A classifier working on joint positions extracted from the videos of twenty archers recognises three phases with 90 per cent accuracy in recurve and 76 in barebow, picks out the set-up better than anything else and confuses drawing and aiming with the expansion into the release, because posture changes too little between the two. A wristband with an accelerometer on the bow wrist tells drawing, aiming and release apart in six adolescent archers in competition, but it is a sensor again. With video and pose estimation it is also possible to measure how consistently an archer repeats the paths of the arms, though not every model works. The study that classifies the phases from video is in a regional venue, works on single frames and times nothing, and the one on arm paths does not recognise phases: the status stays reported.
5. 2026 Reported A technical validation adds a detail that matters to anyone using these systems: the error with which the model locates the joints changes little across the eight phases of the shot, from 0.55 to 0.71 in normalised error, and does not get steadily worse with movement, so much so that aiming and release fare better than stance and set-up. It is, however, a single archer and eight annotated images, with a pre-trained model and no retraining: the status stays reported.

Related

Related articles

• The future of archery: technology, research, evolutionWhat has already been measured about technology in archery: virtual reality, sensors, video, the quiet eye, camera scoring and shooting machines.
• Video analysis and self-coaching in archery: the practical guideHow to film your own shot and what to look for, what the camera measures badly, when sensors help, and what research says about video, feel and the training journal.

Research lines

• 12 studiesMeasuring the shot outside the lab: from the 2005 chronometer to pose estimation
• 48 studiesBow stability and aiming time: research since 1971

Other claims in this domain

• CorroboratedBetween one and a half and two tenths of a second pass between the clicker sounding and the arrow leaving7 studies · 94 archers measuredStatus since 1990
• ReportedA camera diagonally behind the archer measures the vertical sway of the aim well, but not the horizontal1 study · 3 archers measuredStatus since 2010
• ReportedA continuous final pull through the clicker leads to better shots9 studies · 56 archers measuredStatus since 1993
• ReportedA sound or light cue about the angle of the bow arm tightens beginners' groups2 studies · 1 archers measuredStatus since 2012
• ReportedA thumb grip on the string is less tiring than a finger grip3 studies · 9 archers measuredStatus since 1990

See all claims on technique and biomechanics