Automated Machine Learning Evaluation of Squat Movement Patterns Using Smartphone and CoreML

Marta Sokolowska, Michal Podpora

Abstract: There is a vast diversity of Machine Learning use cases, and both the hardware and the methodology itself are constantly getting better. This article presents a very specific use case – the recognising, classifying, and evaluation of human physical activity by using just a video stream from a smartphone and potentially instructing people how to exercise to get better results. The paper presents two key aspects of the challenge: (1) defining the problem in the physical activity domain and choosing a proper ML approach and tool, and (2) conducting the experiment using consumer-grade hardware (a smartphone) to evaluate if the proposed methodology is applicable using an easy-to-use tool and widely accessible hardware for users with limited (or without) Machine Learning experience. Chapter II presents the process of gathering and preparing the dataset, while the next chapter presents the creation of a Machine Learning model, model training using the acquired dataset, and the model performance. It was possible to obtain the result of 92.4% validation accuracy for the squat movement pattern classification model. The trained model can be used in an ML smartphone application for evaluation of user-provided examples of squats.