Overview
My martial arts journey began in my first year at university when I joined the Japanese Jiu Jitsu club. This quickly developed into a strong interest in grappling, leading me to also join the Brazilian Jiu Jitsu club and eventually help refound the Judo club at Newcastle University. At the same time, computer vision has been increasingly used in professional sports coverage to provide performance analysis and tactical insights. I wanted to explore what similar analysis might look like if applied to Judo competitions.
Build Process
I began by applying OpenPose to Judo competition footage to evaluate how well pose estimation performs in a fast-paced grappling environment. The goal was to understand what meaningful, sport-specific insights could be extracted from the detected body positions. One of the fundamental concepts in Judo is kuzushi (balance breaking). As an initial experiment, I wrote a Python script that monitored torso angle and flagged moments when it exceeded a defined threshold, with the aim of identifying when a competitor's balance was being disrupted. The results were mixed. Pose detection was not perfectly reliable during fast exchanges or occlusion between athletes, and torso angle alone did not consistently correlate with genuine balance breaking situations.
This led me to experiment with estimating each athlete's centre of mass (CoM) using torso position data. This approach highlighted an aspect of Judo that is often less explicitly discussed: successful throws frequently occur when an opponent's centre of mass is moved outside their base of support. In techniques such as Seoi Nage, this can also be achieved by positioning your own centre of mass beneath your opponent's, creating a mechanical advantage that enables the throw.
I later added a visualisation comparing the estimated centre of mass with the athlete's foot positions. This made it possible to clearly observe how shifts in balance develop over time and how changes in body positioning influence throwing opportunities.
What I Learned
This project served as an introduction to applying computer vision techniques to real-world sporting analysis. It demonstrated how automated analysis could reveal patterns and metrics that are difficult to notice in real time. The project also highlighted future possibilities, such as training models to automatically recognise specific techniques or attempted attacks, potentially providing coaches and athletes with new forms of performance feedback.