A Squat Classification Model that can classify the quality of your squat based on squat depth, gaze direction and how neutral your spine is

As part of the third year of my Mechanical Engineering degree at the University of Bristol, I completed an Individual Research Project on a topic of my choice. I chose to focus on sensors, specifically exploring how sensor technology can be used to analyse the quality of a person’s squat. One of my primary objectives was to develop a solution that is affordable, accessible, and easy to use for the general public. To achieve this, I opted to use a standard mobile phone in combination with a Human Pose Estimation model. This approach allowed the solution to run directly on a mobile device, provided the user was able to record themselves performing the movement.

Skills

  • Machine Learning (TensorFlow)
  • Data Processing
  • Real-world testing
  • Report Writing

Narrated Video

Abstract

The squat is a fundamental exercise performed daily that requires precise technique to avoid injuries.
Traditional methods for learning correct squat form are often costly or lack real-time feedback, limiting their
effectiveness and accessibility. This study addresses these limitations by developing an innovative real-time
analysis system using pose estimation and machine learning technologies. The system employs the MediaPipe
BlazePose model to capture and analyse squat motions frame-by-frame, extracting landmark data that is then
evaluated by a Squat Classification Model (SCM). Based on a 1D convolutional neural network, the SCM
assesses squat quality using three critical metrics: the optimal knee angle between 55° and 65°, correct gaze
direction (forwards or upwards), and spine neutrality. Trained on the synthetic InfiniteForm dataset, the SCM
achieved an overall accuracy of approximately 90% after 50 epochs, with the depth measurement showing
particularly strong performance (average metric of 89.9%). However, the model was less effective at
recognising incorrect gaze and spine alignments, with average metrics of 63.1% and 55.9%, respectively. Realworld testing confirmed these findings, highlighting the model’s robustness in depth classification but
indicating areas for improvement in gaze and spine analysis. Future research will refine the SCM by utilising
a more suitable dataset and extending the model’s application to other exercises and sports, potentially
enhancing training techniques and injury prevention strategies across various physical activities.