Yoga-pose AI model tops 93% accuracy in testing

An image of a yoga pose can move through the University of East London team's system in approximately 16 to 17 milliseconds per batch under testing conditions. Its best-performing model, Hierarchical CoAtNet 1, identified poses with more than 93% accuracy, according to the study published in Scientific Reports.

The model does not learn every pose as a completely separate label. It first recognizes broader families of poses, then learns the specific variations within them—a structure that mirrors how people organize related movements. During streaming inference, the system reached around 65 to 70 frames per second, a speed the researchers associate with real-time feedback.

The work was co-authored by researchers from UEL, Nirma University, Imperial College London and Doctor On Click. Dr. Laura Vanderbloemen, a senior lecturer at UEL and co-author, said the approach could support people who have limited access to in-person instruction because of location, mobility challenges or cost.

So what changes in practice? A future digital rehabilitation or coaching platform could monitor posture and movement and provide immediate guidance. The researchers also suggest that professionals could use such monitoring to understand posture quality and movement patterns and personalize coaching. For now, those are potential applications: the reported result is a study under testing conditions, not evidence that a clinical system is already in service.