Levering machine learning to distinguish between optimal and suboptimal basketball shooting forms

Basketball is a highly competitive sport and good shooting form is crucial to a player’s success. For high shooting accuracy, several body parts must be aligned, including proper leg positioning, elbow placement, hand posture, and shooting wrist curvature. With the progress in machine learning, Artificial Intelligence (AI) tools can be developed to provide feedback and personalized guidance on basketball shooting form. In this research, we compared Multilayer Perceptron (MLP), Recurrent Neural Networks (RNN), and Convolutional Neural Networks (CNN) to identify the most suitable model for integration into AI tools meant for basketball shooting form analysis. We hypothesized that CNN models will perform significantly better for basketball shooting form analysis than RNN or MLP models because CNN models are known to be better suited for Human Action Recognition (HAR) than other models. We evaluated five models to test our hypothesis - an MLP and an RNN model using Cartesian coordinates of body joints, an MLP model using angles of body joints, and two video-based CNN models using raw and cropped video data. Contrary to initial expectations, the accuracy of the MLP model with Cartesian coordinates of body joints outperformed the CNN model with cropped video (88.3% versus 83.3%). Since MLP models typically require less computational resources, they can be used to build efficient AI tools for basketball shooting form analysis in resource-constrained environments such as mobile phones.