Automatic Classification of Squat Execution Based on Inertial Sensors and Machine Learning

The squat is one the most popular and important sports activities for developing strength and power to lower limbs and is commonly included in conditioning and rehabilitation programs. A suitable squat technique execution minimizes the risk of knee injury and ensures maximal activation of the leg muscle. However, there are a variety of squat techniques that are of personal preference for the development of specific muscles, but many of these are recommended to carry out in a functional range of 0–50 grades of knee flexion. Therefore, in real sports scenarios, the classification of squat execution patterns requires quantitative measures to define its optimal performance, which is still considered a challenge. This work presents a novel approach for automatically classifying the squat execution and determining its suitable execution. For doing that, a measurement system for capturing the joint angles from the lower limbs is proposed, which is based on a set of seven inertial sensors placed at each corporal segment of the lower limbs. The inertial sensors are aligned and calibrated by applying a quaternion strategy. Once, the joint angles are acquired from lower limbs, these are used as an input vector to an automatic classifier, which determines whether squats were performed correctly or incorrectly. This classification problem is faced as a binary classification problem, for which two classifiers were independently applied, both SVM and MLP classifiers, respectively. The performance of the proposed strategy was evaluated with 6 volunteers, 3 men, and 3 women, respectively. The obtained results report the prediction rates of 92% and 72%, respectively. This research highlights the relevance of overcoming limitations and challenges in the use of inertial sensors and artificial intelligence techniques to improve the accuracy of capturing and analyzing squat execution.