Human Boxing Motion Prediction Using Neural Networks

Abstract

Numerous thoughts that were previously deemed inconceivable have become a reality as a result of decades of technical progress and improvement. While flying automobiles are still in the far future, artificial intelligence that can predict your next move is rapidly approaching. Human motion prediction is a relatively new area of active research that is interesting for it’s potential of improving robot’s and other machinery’s ability to work with human, such as passing objects to human, and avoiding crash into human, etc. This thesis focuses on predicting human boxing moves based on RGB visual input as an artificially intelligent boxing trainer with the help of recurrent neural networks (RNNs). I study and compares the performance of six distinct neural network architectures. I have method 1, which includes four model architectures taking 3D joint data as input, and method 2, which includes two architectures that take RGB image as input. Based on the results of all my research, I have discovered the most effective and efficient architecture for scenarios with sparse data based on the outcome of my study.