A new artificial intelligence system that can recognise yoga poses with more than 93 per cent accuracy could help transform digital coaching, home-based rehabilitation and the way physical movement is monitored. Developed and tested by researchers from the University of East London, Nirma University, Imperial College London and Doctor On Click, the model is designed to interpret human posture from images and video, potentially allowing software to provide immediate feedback without requiring a person to be physically present with an instructor or therapist.
The study, published in Scientific Reports, evaluated four newly developed AI models for their ability to identify different yoga poses. The strongest performer, known as Hierarchical CoAtNet 1, achieved an accuracy rate exceeding 93 per cent during testing, substantially improving on earlier approaches reported by the researchers. Although yoga provides the immediate testing environment, the technology belongs to a much broader field of computer vision systems that attempt to understand human movement, body position and physical performance through ordinary visual data.
At the heart of the system is a strategy intended to reflect the way people naturally organise visual information. Instead of treating every yoga pose as a completely separate category, the model first learns broader relationships between groups of poses and then distinguishes between their more specific variations. These relationships form a hierarchy: poses may share similar body configurations, transitions or alignment patterns even when they have different names. Teaching the model these connections can make it more effective at separating movements that appear visually similar while retaining information about their distinctive features.
Hierarchical CoAtNet 1 combines this classification strategy with the CoAtNet family of neural-network architectures. CoAtNet is designed to bring together two important methods used in deep learning: convolutional processing, which is highly effective at detecting local visual patterns such as edges, joints and body contours, and attention-based processing, which can examine relationships across a wider image. In a yoga photograph or video frame, that combination may help the system understand both small details, such as the angle of a limb, and the overall configuration of the body.
The researchers report that the model was also fast enough to operate in real time under the conditions used for evaluation. It processed images in approximately 16 to 17 milliseconds per batch and reached around 65 to 70 frames per second during streaming inference. In practical terms, this could allow a camera-based application to analyse movement continuously rather than waiting for a recording to finish. A digital coach might identify a pose as it is being performed, while a rehabilitation platform could monitor whether a user is maintaining a prescribed position or moving through an exercise correctly.
That speed is important because useful movement feedback depends on timing. A system that responds several seconds after a person has moved may be less helpful than one that can identify a change in posture immediately. Real-time processing could enable applications to highlight an incorrect alignment, signal that a pose has been held long enough, or guide a user through a sequence of movements. It could also help instructors and healthcare professionals review posture quality and movement patterns over time, although the technology would need to be carefully validated before being used to make clinical decisions.
Dr Laura Vanderbloemen, senior lecturer at the University of East London and a co-author of the study, said that accurate movement recognition could make coaching and rehabilitation more accessible. People who live far from specialist services, have mobility challenges or cannot afford regular in-person sessions may benefit from tools capable of offering basic guidance at home. The technology could also support remote supervision, allowing professionals to observe patterns in a person’s movement and use that information to personalise instruction. However, an AI model that recognises a pose is not the same as a clinician capable of diagnosing an injury, assessing pain or adapting treatment to complex medical needs.
The potential applications extend beyond yoga. Human-pose recognition is increasingly being investigated for exercise monitoring, sports training, workplace safety, physical therapy and assistive technologies. A system trained to understand body geometry could, with additional data and testing, help identify whether an exercise is being performed consistently, detect changes in range of motion or track progress during recovery. In robotics and interactive devices, similar techniques could allow machines to respond to human gestures and body positions. Yet performance can vary when lighting, camera angle, clothing, background, body shape or mobility limitations differ from the conditions represented in the training data.
Those limitations make further research essential. High accuracy on a controlled dataset does not automatically guarantee reliable performance in a crowded living room, a dimly lit clinic or a home where the camera captures only part of the body. Yoga poses can also be performed by people with different proportions, flexibility levels and physical abilities, meaning that a system should not interpret every deviation from an idealised posture as an error. Future versions will need broader and more diverse datasets, testing with real users and assessment against expert judgement. Privacy will be equally important, since camera-based health applications may process sensitive information about a person’s body and movements.
The researchers say their work demonstrates how artificial intelligence, computer vision and movement science can be combined to develop responsive digital health tools. The model’s hierarchical design may be particularly valuable because it gives the AI a structured way to learn relationships between movements rather than relying solely on isolated labels. If its performance can be reproduced in less controlled environments and across different populations, the approach could contribute to a new generation of coaching and rehabilitation systems that are faster, more affordable and easier to access. For now, the study offers a promising technical step, while highlighting the need for rigorous clinical testing before such systems are treated as substitutes for qualified human professionals.
Subject of Research: Artificial intelligence for yoga-pose recognition, digital coaching, computer vision and movement monitoring.
Article Title: New AI model could improve digital coaching and rehabilitation
Web References: https://doi.org/10.1038/s41598-026-54558-1
References: Scientific Reports, DOI: 10.1038/s41598-026-54558-1
Keywords
Artificial intelligence, yoga pose recognition, computer vision, machine learning, deep learning, CoAtNet, neural networks, biomechanics, digital health, physical rehabilitation, physical therapy, patient monitoring, movement analysis, real-time inference, healthcare technology, University of East London
