Turn visible movement into something a coach can examine.
An athlete’s movement contains details that can be difficult to follow consistently at speed. A changing ankle position or repeated leg motion may matter, yet a video alone leaves the observer to find and compare those details. My thesis asked how machine learning could make that review more useful.
Cycling provided one example, but the underlying research question was broader: could a system analyse body movement and present useful attributes to coaches? The challenge sat between computer vision and interpretation. Coordinates are a technical result; understanding what deserves another look is a different problem.
The project used pose estimation to extract joint positions frame by frame from standard video, then examine body, leg, ankle, and movement attributes. MediaPipe supported the tracking work. The product perspective was to make the resulting motion data readable for people whose expertise is sport, rather than machine learning.
Individual points become useful when they can be reviewed across a sequence. Move through this illustrative signal to see the relationship between a frame and a measured attribute.
Synthetic signal for explanation. This is not an athlete assessment or a medical interpretation.
Use video as the input to the analysis. The subject, view, and visible movement provide the context for everything that follows. A useful review starts with footage that lets the relevant body points be seen rather than assuming every frame contains dependable information.
Extract body keypoints and examine how their positions change over time. This connects individual detections to movement patterns and attributes, making it possible to look beyond an isolated pose toward the motion occurring across the sequence.
Present movement information for a coach or analyst to inspect alongside the athlete’s action. The intended question is what warrants closer attention. An output becomes useful when its meaning and uncertainty remain understandable to the person reviewing it.
The research starts with standard video input and a pose-estimation pipeline. That keeps the central investigation focused on extracting useful movement information from imagery rather than assuming a specialised motion-capture environment.
Movement is a sequence, so a single set of coordinates cannot carry the whole explanation. The analysis connects detected positions to temporal changes, with cycling providing a concrete setting for examining repeated motion.
Confidence, missing detections, and noise sit between a video and its interpretation. The design principle is to communicate those limits, avoiding a polished chart or overlay that implies more certainty than the underlying detection can support.
This is a thesis and research project exploring movement analysis, including real-time attributes. Its outputs support investigation and coaching review; they should not be read as a validated diagnosis or prediction of injury. No numerical accuracy, latency, or cross-sport validation result is claimed here.
How does a change in camera angle affect the usefulness of an attribute? When should a weak detection be withheld rather than visualised? Which explanations help a coach distinguish a movement pattern from tracking noise? These remain valuable questions for further evaluation.
A little less logging. A little more living.