Video analysis, part 1: what a camera can measure

A poolside video camera is a precision measuring instrument: a stopwatch, a speedometer, and a protractor combined in one housing. Its measurements have been tested against more rigorous standards, and the limits of its accuracy are known. In practice, however, the camera is rarely used as a measuring device. Close to three-quarters of American coaches use video at least once a month, yet mainly for replay: they film the swimmer, play the clip back in slow motion, and review it at the screen (Mooney et al., 2015). Such qualitative observation is undoubtedly valuable, but it uses the instrument at a fraction of its capacity.

In sports science, the camera’s status as a measurement tool was established early on. Much of our foundational knowledge of swimming biomechanics was first extracted from film frames. Counsilman’s underwater footage, for instance, demonstrated that elite crawl swimmers do not simply pull their hands straight back; rather, the hand traces a curve resembling an inverted question mark, an observation that underpinned three decades of research into sculling propulsion (as cited in Callaway et al., 2009).

The aim of this series is to present the video analysis of swimming as empirical research has constructed and verified it.¹ The discussion is organised into five parts:

1. The scope of the method (Part 1). How does visual observation differ from measurement? Here we map out the territory: what the four standard camera views reveal, what a single consumer camera can achieve, what is required to reconstruct three-dimensional body kinematics, and what no camera can measure in principle.
2. Optics and calibration (Part 2). Next, we move underwater, where the refraction of light across water, glass, and air becomes the chief source of measurement error, and show how calibration accounts for it.
3. Measuring velocity (Part 3). We evaluate how accurately video captures swimming velocity, the parameter coaches ask about most frequently.
4. Starts and turns (Part 4). We examine how video analysis is used to improve the execution of starts and turns, where gains of a few hundredths of a second can decide the fate of an entire race.
5. Automation and artificial intelligence (Part 5). Finally, we examine computer-vision and artificial-intelligence algorithms that promise to sharply cut the manual labour required for analysis.

Observation versus measurement

Researchers draw a fundamental distinction between two modes of video analysis: qualitative and quantitative (Mooney et al., 2015).

Qualitative analysis is expert visual judgement augmented by playback tools: slow motion, frame-by-frame stepping, split-screen comparison of two trials, and edited compilations of a swimmer’s own best laps used as a personal benchmark (self-modelling). Its purpose is to answer the question: what is the fault?

Quantitative analysis addresses a more precise question: what is its magnitude? To answer it, anatomical landmarks on the body are tracked frame by frame (digitising), mapped into spatial coordinates in metres, and used to compute durations, velocities, and joint angles. The output is a set of numbers that can be objectively compared against a previous month’s baseline, another athlete’s data, or reference values in the literature.

Of course, qualitative observations can also be compared across time: a coach may record that a swimmer lacked a high elbow at the catch in March, and observe in May that the flaw persists. Yet a visual impression lacks a graduated scale: it registers the presence of an error, but cannot quantify how much that error has been reduced.

A scale is indispensable whenever changes fall below the threshold of naked-eye perception. Consider the glide following a wall push-off: on video replay, a coach can see that the swimmer’s streamline appears tighter, but cannot determine whether the glide velocity has actually increased, or by what percentage. Empirical evidence underscores the value of numerical precision: when coaching instructions during push-off training were supplemented with quantitative feedback, swimmers increased their mean glide velocity over five weeks from 1.74 ± 0.16 to 1.84 ± 0.09 m/s (approximately 6 per cent), a faster improvement than qualitative feedback alone produced (Thow et al., as cited in Mooney et al., 2015).

The remainder of this series is devoted to quantitative analysis: its technical prerequisites, its empirical yield, and its potential pitfalls.

Camera geometry and recording parameters: where the camera stands

The placement of the camera largely determines which variables can be measured in principle. In the research literature, recording setups differ in four methodological respects (Mooney et al., 2015):

1. Medium and viewing plane. A camera may be positioned above or below the water surface and aligned with one of the swimmer’s anatomical planes of motion. In coaching practice, this taxonomy reduces to four standard views, each capturing a distinct aspect of stroke mechanics (Figure 1). Importantly, competition race analysis is restricted to above-water views; it therefore yields split times, relative positions, and stroke rates. While these metrics are highly effective for evaluating race pacing and modelling future race strategies, a thorough biomechanical analysis of stroke technique requires an underwater perspective (Mooney et al., 2015).

Text description

The figure is one scene at real sizes: a pool 2 m deep with lanes 2.5 m wide, seen from above and behind the swimmer’s right as a cut-away block (the near side wall is removed; a dashed outline marks the piece of it where the side cameras stand). A front-crawl swimmer swims in the third lane from that wall. The four cameras are drawn as wire symbols, each with its field of view; below the overview is the picture the highlighted camera records, taken at a moment of the stroke that its view shows well:

1. Above water, side view. Position: on the deck, 0.9 m above the water and 6.75 m from the swimmer’s line. Shown at: the arm recovery. Field of view: above-water segments of the body. Metrics captured: stroke rate, stroke timing, and arm recovery; this is the standard perspective for competition race analysis (split times and pacing).
2. Above water, head-on view. Position: on the deck at the end of the pool, 1.0 m above the water and 5.5 m in front of the swimmer’s head, looking down the lane. Shown at: the breath. Field of view: head and recovering arm. Metrics captured: head posture, breathing mechanics, and frontal symmetry of the arm recovery.
3. Under water, side view. Position: on the side wall, 1.0 m below the surface and 6.1 m from the swimmer’s line. Shown at: the catch. Field of view: the full body in profile. Metrics captured: horizontal body alignment, catch and pull trajectory, and kick depth and timing.
4. Under water, head-on view. Position: on the end wall, 0.8 m below the surface and 4.85 m in front of the swimmer’s head. Shown at: the pull. Field of view: frontal projection of the underwater stroke. Metrics captured: pull width, hand trajectories relative to the body’s centreline, and kick symmetry.

The two above-water pictures show the swimmer’s underwater parts as those cameras see them: the light bends at the surface, and they appear squeezed just under it (Part 2, Figure 1).

Animation sequence: the cameras are highlighted one at a time, about three seconds each, with that camera’s picture below the overview; then all four together, their pictures two by two.

2. Dimensionality (2D vs. 3D). A single camera records motion in a single plane (2D). Movements occurring strictly parallel to that plane are registered at true scale, whereas any movement oblique to the plane undergoes foreshortening, a perspective distortion that introduces measurement error. If you hold a ruler flat against a window, it appears at full length; if you tilt one end toward yourself, its projection on the glass contracts. Exactly the same distortion affects a swimmer’s arm as it sweeps toward or away from a lateral camera. Eliminating perspective error requires reconstructing coordinates in three dimensions (3D) using two or more synchronised, spatially calibrated cameras. Although 3D analysis demands specialised hardware, synchronisation, and calibration, it remains the standard in academic research, whereas applied practice relies predominantly on 2D setups (Mooney et al., 2015).

3. Camera motion. Recording cameras operate in one of three kinematic modes: static, panning, or trolley-mounted. A static camera measures motion against a fixed frame of reference, but its field of view is limited to a few stroke cycles. A panning camera rotates from a fixed point (either manually on a tripod or via a motorised PTZ mount) to track the swimmer across the entire pool, though at the cost of a continuously changing scene geometry from frame to frame. A trolley camera moves parallel to the swimmer along the pool deck, capturing above-water, underwater, or dual-medium footage (Mooney et al., 2015).

4. Frame rate. Temporal resolution is governed by the camera’s frame rate. In a systematic review of thirty studies on swimming video analysis, the majority of quantitative measurements were conducted at 25 to 50 frames per second, with only isolated studies reaching 200 fps (Mooney et al., 2015). Technically, any contemporary smartphone surpasses the frame rates on which most of this scientific literature was built. In practice, however, the primary constraints on underwater image fidelity are shutter speed and the optical properties of water, issues we address in Part 2.

Capabilities of a single consumer camera

What is the minimal hardware configuration required for quantitative video analysis? For competition race analysis, a single consumer-grade camera suffices. A study published in the Journal of Biomechanics demonstrates this clearly: the authors employed basic video recording not as an experimental method under evaluation, but as a standard measurement tool (Washino et al., 2019).

The experimental setup was deliberately minimal:

• Hardware and placement: A single consumer camera (640 × 360 pixels, 120 frames per second) was mounted in the stands (approximately 5 m above the water surface and 30 m from the lane) and panned manually.
• Calibration and digitising: Spatial scale was defined by markers attached to the lane ropes at 5, 10, 15, and 20 m; digitising was performed via manual frame-by-frame stepping in a standard media player.
• Measured variables: Four kinematic parameters were extracted from each maximal 50 m trial (Figure 2): (1) total race time, from the start signal to wall touch; (2) clean-swimming velocity across the 10–20 m and 35–45 m segments (calculated as the 10 m distance divided by the time taken for the head to pass between the corresponding rope markers); (3) stroke rate, computed over three consecutive stroke cycles; and (4) stroke length, derived as segment velocity divided by stroke rate.

Text description

The figure joins a plan of the pool to the camera’s own view, both taken from one scene at real sizes (the pool and ink of Figure 1).

• Plan: A top-down view of the 25 m pool to scale, with a distance scale (0–25 m) on the near deck and the swimmer’s lane between two lane ropes marked at 5, 10, 15, and 20 m. The camera is placed in the stands opposite the middle of the pool (12.5 m; the study gives only its height and distance), 30 m from the lane (dotted line) and 5 m above the water. Its line of sight stays on the swimmer’s head, and its field of view, 10 m wide at the lane, turns with it; the pan angle from straight across is shown as it changes (−5° at the 10 m marks, +14° at the 20 m marks).
• Trajectory and measurement zones: The athlete performs a maximal 50 m front-crawl swim, out to the 25 m wall and back along the same lane. Each 10 m measurement window is shaded as it is swum: 10–20 m on the outbound lap (along the far half of the lane), and 35–45 m of total race distance (the 15 m to 5 m marks on the pool scale) on the return lap (along the near half).
• Camera frames: Two frames from the camera, each shown once the swimmer’s head reaches its marks. A dashed line joins the two lane-rope marks at that distance, and the front of the head is just on it. Frame 1: head at 10 m, pan −5°, 4.62 s, frame 554. Frame 2: head at 20 m, pan +14°, 10.21 s, frame 1225, 671 frames (5.59 s) later.
• Output metrics: A summary table records four variables as each measurement is completed (mean values for 14 swimmers filmed without markers; Washino et al., 2019):
  1. Time (start signal to touch): 26.97 s.
  2. Segment velocity (10 m ÷ 5.59 s): 1.79 m/s.
  3. Stroke rate (over 3 cycles): 0.89 Hz.
  4. Stroke length (V ÷ rate): 2.04 m.

The swim is modelled from these means: the head reaches 10 m at 4.62 s, and from there to the touch the swimmer keeps an even 1.79 m/s, so both windows give the same velocity. The stroke length of 2.04 m is the group’s mean stroke length, not the ratio of the two means (1.79 ÷ 0.89 = 2.01).

Animation sequence: the swim plays at twice real speed; each frame stays an empty dashed outline until the head reaches its marks, and after the touch the table holds its final values.

This raises a natural methodological question: how accurate are measurements obtained from a manually panned camera? For timing variables, Parrott (2023) provides a direct empirical answer across 560 race starts filmed at four national- and international-level competitions. A single consumer camcorder panned from a tripod near the 25 m mark (50 fps) was evaluated against a criterion standard: a stationary camera (100 fps) aligned perpendicular to a calibrated 15 m line. Start time was defined as the interval from the start signal to the swimmer’s head crossing the 15 m mark.

The comparison established two key findings:

1. Validity (Accuracy). The panning camera overestimated 15 m start times by a mean of +0.026 s (95% limits of agreement: −0.048 to +0.100 s). This error margin falls well below 0.187 s, the threshold calculated in the same study as the smallest practically meaningful change in start time.
2. Reliability (Repeatability). Across ten repeated digitisations of the same trials, the consumer panning camera achieved an intraclass correlation coefficient (ICC) of 1.000, slightly outperforming the criterion camera itself (ICC = 0.994). The explanation lies in temporal discretisation: a lower frame rate produces larger steps between consecutive frames, reducing observer ambiguity over which specific frame captures the head crossing the line.

Consequently, the complete set of standard race-analysis metrics (total time, splits, segment velocity, stroke rate, and stroke length) requires no specialised laboratory infrastructure. This tier of video measurement, routinely deployed by national federations at major championships, can be replicated from the spectator stands with a single consumer camera and any standard video player. The test above verified one of these timings, the 15 m start, against a criterion instrument; how accurately video captures velocity is the subject of Part 3.

From point particle to body kinematics: the scope and limits of 2D underwater video

Competition race analysis treats the swimmer as a point particle: it records only when the head crosses a reference line, along with mean segment velocity and stroke frequency. To analyse the mechanics of the stroke itself, the camera must be submerged. The most accessible configurations for coaching practice are a single underwater camera positioned at the pool wall, or a synchronised pair of cameras filming the same volume from above and below the water surface.

These systems need not be stationary. An operator can walk alongside the swimmer carrying an underwater camera mounted on a pole. Commercial telescopic pole rigs are available for coaching (SwimPro, n.d.), or an equivalent setup can be constructed from an action camera and a standard photographic monopod. Adding a second camera above the waterline on the same shaft creates a mobile dual-medium rig.

Moving the camera introduces a strict requirement: in the absence of a static background, spatial scale must be established by reference markers within the lane itself. For example, Born et al. (2024) walked an action camera mounted on a 1.55 m pole along the pool deck at a depth of roughly 20 cm over a 25 m distance, using vertical marker ropes suspended at 5, 10, 15, and 25 m for spatial calibration. The systematic review by Mooney et al. (2015) documents a wide range of such 2D configurations:

• a single underwater camera to quantify dolphin-kick kinematics;
• a synchronised above- and underwater pair to track technical breakdown in butterfly under fatigue;
• two underwater cameras mounted on a trolley to evaluate inter-arm coordination in Paralympic swimmers.

From such recordings, researchers determine the duration of individual stroke phases, inter-limb coordination, kick frequency and amplitude, and planar projections of hand paths, foot paths, and joint angles (Mooney et al., 2015). Furthermore, combining two synchronised cameras (positioned 30 cm below and 30 cm above the surface) into a single split-screen display yields high diagnostic reliability: when experts evaluated swimmer technique against a fault checklist using split-screen footage, inter-rater and intra-rater agreement (re-tested one month later) was near-perfect (Silva et al., 2007).

Nevertheless, all single-plane setups remain bound by a fundamental geometric constraint: the camera records a two-dimensional projection (2D), whereas swimming locomotion is inherently three-dimensional (3D). Any motion directed toward or away from the lens undergoes perspective foreshortening, analogous to the tilted ruler against the windowpane, and is further degraded by radial lens distortion near the edges of wide-angle frames. Consequently, 2D video introduces four systematic errors:

1. Joint-angle distortion: Elbow and shoulder angles are misrepresented whenever the plane of flexion rotates out of the camera’s viewing plane.
2. Loss of lateral depth: Inward and outward sculling movements of the hand beneath the torso are measured wrongly from a side view.
3. Misrepresented axial rotation: Body roll around the longitudinal axis is misrepresented by a lateral camera.
4. Projective artefacts: Even the apparent geometry of a trajectory can be illusory. Viewed head-on, the hand’s path traces a familiar S-shaped curve; yet the swimmer is not pulling in an S-pattern. Rather, a three-dimensional spatial trajectory has merely been projected onto a flat plane (Prins, 2022).

Accurate measurement of joint angles, true spatial trajectories of the extremities, and the velocity of the whole-body centre of mass therefore requires three-dimensional (3D) reconstruction.

Three-dimensional underwater analysis: the dilemma of manual digitising versus motion capture

Reconstructing underwater motion in three dimensions (3D) rests on two methodological pillars: the spatial calibration of several synchronised cameras, and anthropometric modelling.

1. Calibration and 3D reconstruction (DLT). Before recording, a calibration frame bearing control points of known spatial coordinates is positioned within the overlapping fields of view of several synchronised cameras (Mooney et al., 2015). From these points, software computes each camera’s position and orientation. Once an anatomical landmark is identified in at least two camera views, the direct linear transformation (DLT) algorithm reconstructs its 3D coordinates frame by frame (Callaway et al., 2009). Underwater, light refraction across the water–glass interface introduces unavoidable optical deviations that must be absorbed by the calibration model (Callaway et al., 2009).
2. Segmental body modelling. Raw 3D coordinates alone do not yield joint angles or the whole-body centre of mass. The reconstructed points must be mapped onto a 3D biomechanical model in which each body segment (upper arm, forearm, thigh, etc.) is assigned a known length, mass, and segmental centre of mass, parameters that can be individualised for a given swimmer using as few as two calibrated photographs, front and side (Sanders et al., 2015).

The central practical challenge lies in how anatomical landmarks are tracked across thousands of video frames. Researchers face a trade-off between two approaches.

Approach 1: manual digitising and the labour bottleneck. In manual digitising, a human operator clicks every landmark on every frame of every camera view. Data from Mooney et al. (2015) illustrate the severity of this bottleneck: digitising 30 dive starts (10 swimmers × 3 trials) requires roughly 7.5 hours, while digitising just four stroke cycles of a single swimmer (6 cameras, 19 landmarks, 1,620 frames) takes 27 hours. A movement lasting a few seconds in the pool demands more than three working days at a computer. Early efforts to automate this process without markers offered only partial relief: when Ceseracciu et al. (2011) adapted a clinical silhouette-based algorithm to track the right shoulder, elbow, and wrist across six underwater cameras (five front-crawl sprinters), automated 3D reconstruction took 2 to 3 hours of computer time per trial with no operator involved, compared with 6 to 7 hours of an operator’s manual marking in Simi Motion.

Approach 2: optical motion capture and the observer effect. To automate tracking fully, optical motion capture uses spherical markers (typically 19 mm in diameter) attached to the skin or swimsuit over bony landmarks (Olstad et al., 2014). Coated in a retroreflective material that returns light directly toward its source, the spheres reflect powerful LED strobes mounted around the camera lenses, appearing as bright, easily segmented dots against a dark background. Because water heavily attenuates light and strips standard land markers of their reflectivity, underwater motion capture requires either specialised cameras equipped with high-powered cyan strobes and custom markers (e.g., Qualisys; Olstad et al., 2014; Washino et al., 2020) or land-based systems (e.g., Vicon) filming through submerged viewing windows (Monnet et al., 2014).

However, marker-based automation introduces a classic observer effect: the markers alter the very performance they are meant to measure (Washino et al., 2019). When 14 well-trained swimmers performed maximal 50 m front-crawl sprints with and without the 25 retroreflective spheres required to reconstruct a full 3D segmental model and centre of mass, the marker condition significantly degraded performance:

• Total 50 m time increased by 3.9 ± 1.6 per cent (28.05 s vs. 26.97 s, P < 0.001).
• Clean-swimming velocity fell by 3.3 per cent, stroke rate by 1.2 per cent, and stroke length by 2.1 per cent.
• Blood lactate concentrations did not differ significantly between conditions, indicating comparable physiological effort.

A hydrodynamic calculation confirmed that the parasitic drag of the 25 spheres (approximately 5.3 N in total) accounted almost exactly for the observed velocity loss, consistent with earlier reports of a 7 to 10 per cent increase in passive drag from a 24-marker set (Mooney et al., 2015).

Strategies to mitigate marker drag. Researchers have explored three avenues to reduce or eliminate added drag:

1. Flat retroreflective tape. Flat tape markers barely protrude from the body and so add practically no form drag; in one study, two tape markers on short stand-offs were placed on the spine to stay visible near the air–water interface (Washino et al., 2020). In controlled tank tests, markers covered with marine-grade (SOLAS) retroreflective tape and filmed through glass were localised as accurately underwater as in air (Raghu et al., 2019).
2. Flat high-contrast skin or suit marks. Non-reflective circles painted on the skin or applied to the suit can be tracked on standard video, provided contrast is sufficient. While white circles on a black suit were tracked with no operator correction at all, front-crawl swimming generated enough bubble interference to require manual operator correction in 14.6 to 33.6 per cent of frames, depending on the software (Magalhaes et al., 2013). Moreover, all flat markers suffer from foreshortening error when viewed at oblique angles (OptiTrack, n.d.), a constant issue given the continuous axial rotation of a swimmer’s torso. Notably, none of the literature reviewed here directly compares flat and spherical markers on swimmers for either hydrodynamic drag or tracking accuracy.
3. Marker-set reduction. The group that measured the spheres’ drag demonstrated that reducing the marker set from 36 spheres to just 4 trunk markers still predicted vertical centre-of-mass position with an intraclass correlation coefficient of 0.776 relative to the full marker set (Washino et al., 2023).

The cost of a complete 3D picture. Consider the instrumentation required to capture the ideal biomechanical dataset across a 200 m swim: continuous centre-of-mass velocity, body depth, and frontal surface area (Washino et al., 2020). Achieving this took 15 underwater and 10 above-water motion-capture cameras operating at 100 Hz, 36 markers, a full-body 3D scan of each athlete, and a fitted digital human model. Yet this extensive array covered only an 8 m measurement zone within a 25 m pool (one stroke cycle per lap), tested only three swimmers, and relied on markers that, as the same laboratory had just shown, slow swimmers down. (Over the 200 m, the recorded velocity fell from 1.63 to 1.36 m/s alongside a sinking centre of mass and rising frontal area.)

The sheer scale of that equipment list highlights the methodological impasse. As the authors themselves concluded in 2019, a markerless motion-capture system would likely be the optimal solution, but “has yet to be developed for underwater activities” (Washino et al., 2019). Whether that limitation still holds is the subject of Part 5.

The fundamental limit of the method: what no camera measures

Suppose technological progress eventually resolves the trade-offs described above, giving us affordable underwater systems free from both the labour of manual digitising and the hydrodynamic drag of surface markers. Even under those ideal conditions, video analysis will remain bound by a fundamental physical limit.

That limit is the boundary between kinematics and kinetics (dynamics). A camera records spatial coordinates across successive points in time; consequently, its domain is strictly kinematic: it captures the geometry of motion and its time derivatives (velocities and accelerations). The forces that generate or oppose that motion never appear in the optical image.

In terrestrial biomechanics, this limitation is overcome by pairing cameras with force plates embedded in the ground. In free swimming, however, the athlete interacts solely with an uninstrumented fluid medium. Video footage cannot directly register the pressure differential across the palm, the form and wave drag resisting the body, or the propulsive force generated by the kick. While these kinetic quantities can be inferred by feeding kinematic data into fluid-dynamic models, every such estimate compounds two distinct sources of uncertainty: the theoretical simplifying assumptions of the model layered atop the measurement errors of the camera. Video reveals the kinematic effect (the trajectory of the body) while leaving the kinetic cause unseen.

Summary of Part 1

1. Metrological status. A poolside video camera is a quantitative measuring instrument whose margins of error have been empirically tested and published in the scientific literature.
2. Competition race analysis (2D from the stands). A single hand-panned consumer camera times the 15 m start to within a tenth of a second and also yields total time, segment velocity, stroke rate, and stroke length.
3. Three-dimensional technique analysis (3D underwater). Reconstructing the 3D kinematics of body segments requires either dozens of hours of manual digitising or retroreflective markers that measurably slow the swimmer down. Complete 3D analysis remains a laboratory procedure rather than a poolside tool.
4. The kinematic boundary. All video analysis is restricted to kinematics: a camera combines a stopwatch, a speedometer, and a protractor, but cannot function as a dynamometer. It measures motion, never force.

Part 2 examines the optical physics of the aquatic environment and shows why, underwater, measurement accuracy is limited not by sensor resolution, but by the refraction of light across media.

The series

1. What a camera can measure (this article)
2. Water against optics: where the error budget lives
3. Velocity from video: from race splits to the stroke cycle
4. Starts and turns: hundredths of a second on camera
5. When the machine watches: markerless tracking and artificial intelligence

Parts 2 to 5 are being written; their titles are working ones.

Footnotes

1. The series is a narrative review of selected works, not a systematic search. Where it says that something has not been measured, that holds for the sources I read. ↩

References

1. Ariel Dynamics. (n.d.). Features. APAS: Ariel Performance Analysis System. https://www.arielnet.com/apas/features
2. Ariel Dynamics. (2025). APAS downloads [APAS/AI version 14.4, dated 26 August 2025]. APAS: Ariel Performance Analysis System. https://www.arielnet.com/apas/downloads
3. Born, D.-P., Polach, M., & Staunton, C. (2024). Determining validity and reliability of an in-field performance analysis system for swimming. Sensors, 24(22), Article 7186. https://doi.org/10.3390/s24227186
4. Callaway, A. J., Cobb, J. E., & Jones, I. (2009). A comparison of video and accelerometer based approaches applied to performance monitoring in swimming. International Journal of Sports Science & Coaching, 4(1), 139–153. https://doi.org/10.1260/1747-9541.4.1.139
5. Ceseracciu, E., Sawacha, Z., Fantozzi, S., Cortesi, M., Gatta, G., Corazza, S., & Cobelli, C. (2011). Markerless analysis of front crawl swimming. Journal of Biomechanics, 44, 2236–2242. https://doi.org/10.1016/j.jbiomech.2011.06.003
6. Coifman, I., Hakim, M., Weiss, G., & Riemer, R. (2026). Human pose estimation for automated biomechanical swimming analysis. Sports Biomechanics. https://doi.org/10.1080/14763141.2026.2715647
7. Dartfish. (2024, August 13). Distance, time and angle measurement tools. Dartfish Support. https://support.dartfish.tv/en/support/solutions/articles/27000048163-distance-time-and-angle-measurement-tools
8. Hong, W., Cheng, Y., Yang, Z., Wang, W., Wang, L., Gu, X., Huang, S., Dong, Y., & Tang, J. (2025). MotionBench: Benchmarking and improving fine-grained video motion understanding for vision language models. arXiv. https://arxiv.org/abs/2501.02955
9. Kinovea. (n.d.). Features. Kinovea. https://www.kinovea.org/features.html
10. Kinovea. (2026, March 21). Download. Kinovea. https://www.kinovea.org/download.html
11. Magalhaes, F. A., Sawacha, Z., Di Michele, R., Cortesi, M., Gatta, G., & Fantozzi, S. (2013). Effectiveness of an automatic tracking software in underwater motion analysis. Journal of Sports Science and Medicine, 12(4), 660–667. https://www.jssm.org/volume12/iss4/cap/jssm-12-660.pdf
12. Monnet, T., Samson, M., Bernard, A., David, L., & Lacouture, P. (2014). Measurement of three dimensional hand kinematics during swimming with a motion capture system: A feasibility study. Sports Engineering, 17(3), 171–181. https://doi.org/10.1007/s12283-014-0152-4
13. Mooney, R., Corley, G., Godfrey, A., Osborough, C., Quinlan, L. R., & ÓLaighin, G. (2015). Application of video-based methods for competitive swimming analysis: A systematic review. Sports and Exercise Medicine Open Journal, 1(5), 133–150. https://doi.org/10.17140/SEMOJ-1-121
14. Olstad, B. H., Zinner, C., Haakonsen, D., Cabri, J., & Kjendlie, P.-L. (2014). A new approach for identifying phases of the breaststroke wave kick and calculation of feet slip using 3D automatic motion tracking. In Proceedings of the XIIth International Symposium on Biomechanics and Medicine in Swimming (pp. 195–199).
15. OptiTrack. (n.d.). Markers. OptiTrack Documentation. https://docs.optitrack.com/motive/markers
16. Parrott, H. R. (2023). The swimming start: Measurement, importance and enhancement through pre-race interventions [Doctoral thesis, Swansea University].
17. Prins, J. (2022). The biomechanics of unintended consequences — An update, Part II. American Swimming, 2022(1).
18. Qualisys. (2026, June 10). qtm-rt 3.1.0 [Python package for the QTM real-time protocol]. PyPI. https://pypi.org/project/qtm-rt/
19. Quintic Consultancy. (2026). Quintic software. Quintic Sports. https://www.quinticsports.com/software/
20. Raghu, S. L., Kang, C.-K., Whitehead, P., Takeyama, A., & Conners, R. (2019). Static accuracy analysis of Vicon T40s motion capture cameras arranged externally for motion capture in constrained aquatic environments. Journal of Biomechanics, 89, 139–142. https://doi.org/10.1016/j.jbiomech.2019.04.029
21. Sanders, R. H., Chiu, C.-Y., Gonjo, T., Thow, J., Oliveira, N., Psycharakis, S. G., Payton, C. J., & McCabe, C. B. (2015). Reliability of the elliptical zone method of estimating body segment parameters of swimmers. Journal of Sports Science and Medicine, 14, 215–224.
22. Silva, A. J., Costa, A. M., Oliveira, P. M., Reis, V. M., Saavedra, J., Perl, J., Rouboa, A., & Marinho, D. A. (2007). The use of neural network technology to model swimming performance. Journal of Sports Science and Medicine, 6, 117–125. https://pubmed.ncbi.nlm.nih.gov/24149233/
23. Simi Reality Motion Systems. (n.d.). Video motion analysis of swimming. Simi. http://www.simi.com/en/applications/sport/swimming-analysis.html
24. Simi Reality Motion Systems. (2026). Home [Notice of the closure of the Sports and Life Science Motion Capture division, April 2026]. Simi. http://www.simi.com/en/
25. SwimPro. (n.d.). Camera systems. SwimPro. https://www.swimpro.com/camera-systems/
26. TechSmith. (2019). Coach's Eye. Coach's Eye. https://web.archive.org/web/20191229133121/https://www.coachseye.com/
27. TechSmith. (2021). Coach's Eye retirement. Coach's Eye. https://web.archive.org/web/20210919234452/https://go.coachseye.com/retirement/
28. Thompson, I. M., Audot, D. A. G., Warner, M. B., Banks, J., Logan, O., & Hudson, D. (2025). Error assessment of a three-dimensional underwater motion capture methodology. Sports Biomechanics, 25(4), 485–501. https://doi.org/10.1080/14763141.2025.2514234
29. Thow, J. L. (2016). Application of 2-dimensional video analysis by competitive swimming coaches to monitor fatigue in breaststroke technique during training [Doctoral dissertation, University of Edinburgh]. http://hdl.handle.net/1842/20453
30. Washino, S., Mayfield, D. L., Lichtwark, G. A., Mankyu, H., & Yoshitake, Y. (2019). Swimming performance is reduced by reflective markers intended for the analysis of swimming kinematics. Journal of Biomechanics, 91, 109–113. https://doi.org/10.1016/j.jbiomech.2019.05.017
31. Washino, S., Murai, A., Mankyu, H., & Yoshitake, Y. (2020). Association between changes in swimming velocity, vertical center of mass position, and projected frontal area during maximal 200-m front crawl. Proceedings, 49(1), Article 60. https://doi.org/10.3390/proceedings2020049060
32. Washino, S., Murai, A., Mankyu, H., & Yoshitake, Y. (2023). Determination of subset models for predicting vertical centre of mass position during front crawl in male swimmers. Journal of Sports Sciences, 41(4), 372–380. https://doi.org/10.1080/02640414.2023.2214393