Skip to main content

Human activity recognition and estimation of joint kinematics in Ballet

Team: Danica Hendry, Dr Amity Campbell, Prof Peter O’Sullivan, Dr Luke Hopper, Prof Leon Straker, Prof Tele Tan

CIC specialists: Dr Kathryn Napier, Dr Richard Hosking, Dr Kevin Chai

Professional and pre-professional ballet dancers have an intense physical training regime that involves repeated jumping and leg lifting tasks that are associated with the development of lower limb and lower back pain. Currently training load is estimated from dancer’s written diary entries (that can be imprecise and biased), while the investigation of joint kinematics requires sophisticated and expensive laboratory based optical motion capture systems. Using affordable wearable sensors, we were able to develop a machine learning based approach that can accurately measure dancer training load and joint kinematics in female pre-professional ballet dancers.

The first part of this study aimed to better quantify training volume using wearable sensors that incorporated an accelerometer, a gyroscope and a magnetometer. Kevin Chai developed a human activity recognition model that identified six different jumping and leg lift ballet movements with a classification accuracy of over 80%. The ability to recognise specific jumping and leg lifting tasks not only enabled the quantification of training volume in dancers, it also allowed us to then investigate the quality of movement during specific leg lifting tasks. The second part of this study therefore developed a second machine learning model that was capable of using data from wearable sensors to predict hip and trunk angles during commonly practised ballet leg lifting tasks.

Kathryn Napier and Richard Hosking developed a joint kinematics prediction model that predicted hip angle as a measurement of leg height, trunk sagittal plane angle, and also identified which leg was performing the task. The average root mean squared error (RMSE) and mean absolute error of the peak angles (MAE) was between 5.5-7.5° for hip angles. As dancers can lift their legs up to 140° (hip angle), this degree of error is considered very small.

The models developed were robust enough to identify jumping and leg lifting movements and to identify how often and how high dancers lifted their legs during leg lifts in a real-world ballet training class. These models are now being used in a longitudinal study to provide further insight into the factors influencing a dancer’s pain and injury risk.