Towards Optimized Physical Activity: A Fusion of Graph Convolutional Networks and XGBoost for Exercise Analysis

This research explores the implementation of the GCN-XGBoost with attention mechanism within the detection and correction of exercises through computer vision, utilizing the 2-dimensional pose coordinates from the MMFit
^[
1
]
dataset. The objective is to enhance the precision and reliability of squat performance analysis, leveraging the robust, scalable machine learning capabilities of XGBoost. By processing and analyzing two-dimensional skeletal data, the study aims to identify key postural deviations and provide corrective feedback in realtime, potentially reducing the risk of injury and improving overall exercise effectiveness. Initial results demonstrate the model's capability to discern between correct and incorrect exercise forms with high accuracy, indicating the feasibility of applying advanced machine learning techniques to personal fitness and rehabilitation contexts. This paper details the methodology, from data preprocessing to model training and evaluation, highlighting the contributions to the fields of sports science and machine learning in physical activity monitoring. We have used MMFit dataset for the model that we have trained. We use the 2D coordinates of 18 of the 33 joints in the human body as the main features to categorize various exercises and forecast the accuracy of the poses identified by the Google's AI framework MediaPipe [2]. In this paper, we first classify an individual's exercises (squats, lunges, bicep curls, situps, pushups, triceps extensions, dumbbell rows, jumping jacks, dumbbell shoulder press, lateral shoulder raises) and check whether the pose associated with an exercise is correct. The final model trained has a 98% final validation accuracy.