Detection and Classification of Weightlifting Form Anomalies using Deep Learning

Abstract

Detection and classification of weightlifting anomalies is important to prevent the risk of inflicting injury and to maximise the effect of the weightlifting exercises.  During COVID-19 pandemic, going to the gymnasium is not possible.  Thus, for those who have the necessary weightlifting equipment at home but no instructor, having automatic anomalies detection is beneficial.  Although, weightlifting form recognition or anomalies detection often require utilisation of external sensors or hardware such as motion and kinetic sensors to produce accurate feedback for the user, not many have access to these external requirements. Thus, the objective of this research is to develop a prototype that is capable of providing feedback on the correctness of weightlifting technique execution using computer vision by implementing deep learning method.  One of the popular deep learning methods for detection and recognition is You Only Look Once version (YOLO).  Since there is no publicly available dataset for training and testing purposes, videos and images on weightlifting are searched and extracted from the Internet. 387 static images were collected with 219 images for normal forms and 134 images for abnormal forms.  A confidence score in the range between 0.8 and 0.9 has been achieved during testing.  Even though the performance produced is not high which is mainly due to the size of the training data, it can still serve as a foundation for future implementation for identifying weightlifter’s technique execution and help to maximize the exercises.