Inspiration
The Anterior Cruciate Ligament (ACL) is one of the four major ligaments in the knee that provides stability and helps with side-to-side movement. A torn ACL is a fairly common injury, especially in football (soccer), and can range from a partial tear to a complete rupture. In fact, ACL tears are one of the most devestating injuries in soccer. Football is a sport that inherently stresses out the knee joint to a significant level. In detail, high-speed and rotational movements, which are typical in the sport, inevitably increase the risk of an ACL rupture. Such movements include:
1) Stop and start movements 2) Sudden changes in direction 3) Jumping and landing awkwardly 4) Player-to-player contact
Studies have shown that women footballers are subject to an ACL injury nearly 2.2 times higher than male footballers (Montalvo et al., 2019). The consequences and general impact can be enormous to the team, the club and most importantly, the player themselves.
A four-season prospective UEFA study involving 596 players from 15 elite women’s teams recorded 1,527 injuries. ACL injuries represented only 2% of injuries, but shown the greatest burden, with a median absence of 292 days—approximately nine and a half months (Hallén et al., 2024).
Concluding, it is more than evident that women face a markedly greater ACL burden than men — a disparity that science cannot afford to ignore anymore.
bold What makes women footballers so susceptible to an ACL injury?bold
Unfortunately, this is not a just-one complete explanation for why this phenomenon occurs. Throughout, the years, multiple possible contributors have been proposed and discussed, but the expert stress that the explanation is not yet concrete. Many support that an ACL injury is multi-factorial. Mechanics, anatomy, preparation, load and other contextual factors can interact.
Footballers are repeatedly exposed to sprinting, cutting, pivoting, jumping and sudden changes of direction. These reactive movements are highlighted as vital movements in the sport. The more explosive the movement is, the greater the forces transmitted through the knee. From a movement point of view, female and male athletes are biomechanically different. They essential pose significant differences in hip structure, knee position and other anatomical factors.
Furthermore, women’s football has become professional in a short amount of time recently however, physical preparation and infrastructure have failed to improve in the required pace.
Therefore, it is of utmost importance to protect female footballers' health, well-being, careers and longevity on the sport itself. Movement analysis can turn cutting, deceleration, landing and joint behaviour into measurable biomechanical evidence. Through it, specific movement factors can be identified — and physical qualities such as strength and agility can then be targeted.
Researchers already use match footage to understand how these injuries occur. A study conducted on systematic manual video analysis examined 37 ACL injuries in professional women’s football and found that 80% of injuries among field players occurred during horizontal movements such as sprinting, changing direction, stopping, and lunging, (Four distinct ACL injury patterns, 2024).
Therefore, we strongly believe that through the utilisation of technology, data analytics, computer vision and AI, we may not be able to eliminate ACL risk, but we can understand it better and continue to reduce it.
boldCan we make real-world njury footage (through a single broadcast camera) easier to examine, compare, and audit an ACL injury? bold
What it does
ACL Movement Analytics Lab accepts documented women’s-football ACL injury footage and produces interpretable, evidence-aware 2D Movement Stories.
The basic algorithmic flow of the software is described as:
1) Create an injury-event case and attach one or more video views, subject to availability.
2) Isolate the relevant video clip featuring the injury.
3)Identify the target athlete and define the visible movement window.
4) Correct the athlete’s tracking and mark unavailable intervals.
5) Generate quality-controlled 2D pose and movement measurements.
6) Explore movement phases through synchronised video, trajectories, and source frames, if they can be reliably constructed.
In addition, the user can:
1) Compare cases using only measurements supported in both cases.
2) Investigate case-level patterns, evidence coverage, and data provenance.
Focusing on producing an application that consists of a highly responsible AI-element we ensured that:
1) Every result remains connected to its supporting video frames. Unsupported measurements appear as gaps instead of being converted into zero or falsely estimated.
2) Uncertainty also changes the outcome. The system can produce a supported Movement Story, narrow the result to a defensible interval, report that no sustained transition was detected, or withhold phase analysis entirely.
bold ACL Movement Analytics Lab does not diagnose an ACL injury, predict injury risk, estimate ligament loading or forces, reconstruct true 3D biomechanics, or determine what caused an injury.bold
How we built it
We built ACL Movement Analytics Lab as an end-to-end local research application in Python 3.12. We proceed with listing the most relevant technologies used, as well as their contribution to the software.
1) OpenCV: handles video processing, cutting, frame access, overlays, and human-correctable target masks.
2) Ultralytics YOLOv8 Pose : Provide COCO-17 body landmarks that create markerless pose estimation.
3) Pandas and NumPy: Contribution to the pipeline that follows the following order.
a) Preserves the original pose coordinates. b) Rejects low-confidence or unreliable landmarks. c) Interpolates only short, supported gaps. d) Smooths valid continuous segments. e) Calculates projected 2D angles, orientations, normalised distances, paths, and temporal changes. f) Keeps unsupported measurements null throughout the analysis.
4) HTML, CSS, and JavaScript: Used for the UI of the software, as a web application.
5) JSON, CSV, and Parquet: used for storing annotations, processing decisions, quality states, and source-frame provenance.
bold Key Algorithmic Decisions: *bold *
Inevitably, some players' whole movement order is unlikely to be reliably and accurately captured for various reasons such as: occlusion, video quality, angle view and so on.
Therefore, a missing-aware phase engine combines supported movement descriptors to identify sustained multivariate changes. Minimum-duration and evidence rules prevent a single noisy frame from becoming an artificial movement phase.
Cross-case comparison uses reliability-weighted, robustly scaled case-level measurements. Only mutually supported measurements contribute, and each result reveals its shared-evidence support. Multiple replay views of the same injury remain part of one injury event instead of being counted as independent cases.
bold Prototype Scale and Testing:*bold *
For this hackathon prototype, we processed 52 human-reviewed video views. The reliability of the codebase is supported by an automated test suite containing 338 passing tests.
Challenges we ran into
bold Working with real broadcast footagebold
Football broadcast footage is not designed for biomechanical measurement due to various reasons. Examples include, camera motion and perspective, zoom, replay speed, partial or even severe occlusion, players overlapping with our target, and changing viewpoints. All these can all alter a projected 2D measurement to a significant extend.
boldThis limitation is supported by broader markerless-motion research. bold
Specifically, a 2025 systematic review of 53 studies found that differences between markerless and marker-based measurements ranged from 0.2° to 28.6°, depending on the task, joint, movement plane, and technology (Yoma et al., 2025).
Hence, our degree-valued extracted results are explicitly described as projected 2D angles, and bold not true anatomical 3D joint angles bold.
bold Maintaining the correct athlete's pose*bold *
Pose models can switch to another player or misplace a joint while still creating a plausible-looking skeleton. Therefore, I have addressed this through human-drawn athlete regions, correction keyframes, unavailable-interval annotations, temporal quality checks, and visual pose overlays.
bold Guarding against false certaintybold
A major challenge was deciding what the system should do when evidence was incomplete. I decided that unsupported samples are to remain missing, restricted movement phases to continuous evidence, and blocked comparisons without enough shared measurements.
bold Preserving statistical independencebold
A single injury may appear in several replays and thousands of frames. Treating those views or frames as separate injuries would artificially inflate the dataset. We therefore model the injury event as the independent case and keep its video views nested beneath it.
bold Adapting the pose technologybold
MediaPipe was our initial pose-estimation choice, but its native Metal service crashed in our desktop environment. Because the failure occurred outside Python, it could not be safely recovered. Instead, I created a replaceable pose-backend contract and implementing YOLO Pose as a reliable fallback.
Accomplishments that we're proud of
I am proud that using Data Science/Machine Learning/Computer Vision and Agentic AI, I've built a pose analysis and similarity engine software instead of a single pose-estimation prototype.
ACL Movement Analytics Lab integrates video preparation, human annotation, pose extraction, evidence evaluation, movement measurements, temporal phase analysis, interactive visualisation tools and statistics for dataset exploration, and movement similarity comparison across cases.
Every movement claim can be traced to the following stages:
1) The original injury-event case. 2) The selected video view. 3) The human annotation session. 4) The contributing pose landmarks. 5) The measurements evidence status. 6) The exact supporting source frames.
I am especially proud that at every step of the analysis stage onwards, I have utilised responsible AI in producing results and assisting the user into sensible and responsible decision-making and results interpretation.
In detail, real analysed cases demonstrate four possible outcomes:
1) Supported analysis 2) Limited-interval analysis 3) Supported finding of no transition 4) Withheld results
Furthermore, I am also proud that movement phases are constructed only when the evidence sufficiently supports them. In the context of our responsible AI realisation, the engine does not force a fixed number of phases. Instead, it identifies sustained changes across multiple supported measurements, prevents isolated noisy frames from creating false transitions, and limits or withholds phase construction when the evidence is insufficient.
The similarity engine follows the same responsible approach. It handles multiple views of the same injury as one distinct case and compares only measurements supported in both cases.
Four distance measures are available:
1) Weighted Manhattan (L1) distance for overall movement difference. 2) Weighted Euclidean (L2) distance for larger deviations. 3) Weight Cosine similarity for movement-pattern direction; 4)Weighted Soft-cosine similarity for relationships between measurements.
Each similar case demonstrates its shared evidence. We should note that the movement similarity observed does not imply the same injury mechanism, biological cause, or diagnosis.
It was of utmost importance for me that the AI-component of the system did not force itself to provide an answer. Instead, it could recognize when the available evidence does not justify a concrete result.
What we learned
Valuable lessons have been learnt throughout this process.
Firstly, I learnt that human inspection is not just a temporary workaround for imperfect AI. It supplies essential context that pose estimation cannot determine reliably, including which athlete is relevant, whether the intended movement is visible, and whether the case metadata is trustworthy.
Furthermore, I also learnt that a confidence warning next to an unreliable number is not enough for correct human interpretation of results. Uncertainty becomes meaningful only when it changes what the system calculates and displays. Therefore, it is of vital important to include human operators in the loop, at every step the AI takes into generating a result or lack of therefore.
In addition, data design matters equally as model selection does. Therefore, by separating injury events, video views, frames, annotations, raw coordinates, processed coordinates, and derived measurements, I ensure that the analysis stays accurate, reliable and overall protected from false sample sizes and hidden transformations.
After all, published research firmly supports the value of combining computation with human review. A 2024 professional women’s-football video study used three expert raters and a standardised observation process. This is strong evidence that contextual interpretation remains essential even in formal video-analysis research (Systematic video analysis study).
What's next for ACL Movemement Analytics Lab
Future work denotes the expansion of the independently sourced and human-annotated case library. I also plan to work with biomechanics, sports-medicine, Responsible AI, privacy, and footage-rights experts to achieve the following:
1) Measure inter-rater agreement. 2) Review of the movement-comparison rankings. 3) Evaluate pose performance across a number of factors such as camera angles, resolution, motion blur, occlusion, athlete scale and so on. 4) Improve target tracking and human-correctable athlete segmentation.
boldSelected scientific references*bold*
Montalvo AM, et al. What’s my risk of sustaining an ACL injury while playing football (soccer)? A systematic review with meta-analysis. British Journal of Sports Medicine. 2019;53(21):1333–1340. Full text
Hallén A, et al. UEFA Women’s Elite Club Injury Study: a prospective study on 1527 injuries over four consecutive seasons. British Journal of Sports Medicine. 2024;58(3):128–135. Full text
Four distinct patterns of anterior cruciate ligament injury in women’s professional football: a systematic video analysis of 37 match injuries. British Journal of Sports Medicine. 2024. Full text
Yoma M, et al. Reliability and validity of lower-extremity and trunk kinematics measured with markerless motion capture during sports-related and functional tasks: a systematic review. Journal of Sports Sciences. 2025;43(17):1703–1730. PubMed
Log in or sign up for Devpost to join the conversation.