Analysis of the Squat Exercise from Visual Data

Squats are one of the most frequent at-home fitness activities. If the squat is performed improperly for a long time, it might result in serious injuries. This study presents a multiclass, multi-label dataset for squat workout evaluation. The dataset collects the most typical faults that novices make when practicing squats without supervision. As a first step toward universal virtual coaching for indoor exercises, the main objective is to contribute to the creation of a virtual coach for the squat exercise. A 3d position estimation is used to extract critical points from a squatting subject, then placed them in a distance matrix as the input to a multi-layer convolution neural network with residual blocks. The proposed approach uses the exact match ratio performance metric and is able to achieve 94% accuracy. The performance of transfer learning as a known machine learning technique is evaluated for the squat activity classification task. Transfer learning is essential when changing the setup and configuration of the data collection process to reduce the computational efforts and resources.