Robust baseball pitch reconstruction using artificial neural networks with noisy data

Abstract The use of high-fidelity camera-vision and Doppler tracking systems in Major League Baseball (MLB) has created an influx of advanced analytics that have transformed game and personnel strategy. However, due to the cost and complexity of these systems, advanced statistics are severely limited in most amateur games. In this work, we develop a robust pitch reconstruction methodology using artificial neural networks (ANN) and a three-point reconstruction technique, requiring the knowledge of only three baseball spatiotemporal locations. The ANN models were trained to predict the initial baseball speed and spin rate, which are then used as initial conditions to integrate the full pitch trajectory. We use numerically simulated baseball pitches to train two ANN models, each with clean and noisy training inputs, respectively. We then performed a robustness analysis to test ANN performance with increasingly noisy data to simulate low-fidelity camera tracking systems. Probability distributions of predicted model output are calculated using the Monte Carlo method to quantify model uncertainty. We show that the ANN models accurately predict the true baseball trajectory from both high-fidelity and noise-injected testing data. The results demonstrate the effectiveness of ANN models for quick and robust pitch trajectory reconstruction using minimal input data, even in the presence of noise. The present methodology provides a first step towards enabling pitch tracking and advanced analytics in a wider variety of baseball games.