Cricket Fast Bowling Optimization Using Machine Learning Pose Estimation Modeling

Abstract

Fast bowling in cricket is a biomechanical process that involves multiple phases: Run-up, Jump, Ball Delivery, and Follow-through. The goal of my project was to analyze the jump and ball delivery phases using computer vision, ML-based pose estimation, and physics-based analysis to identify key differences between professional and novice bowlers, providing both visual feedback (annotated videos) and data-driven insights (statistical clustering).

After evaluating multiple models, I developed a program using the Python-based Mediapipe Pose Estimator library, which utilizes 33 pose points. Among 34 biomechanical parameters, the study identified 17 key parameters that affect bowling performance, with a focus on arm, leg, wrist, and foot positioning, as well as wrist speed. The original dataset was created from 30 professional and 5 novice bowlers’ (46 balls) MP4 videos.

The software reliably captured and annotated the video of novice bowlers, overlaying body angles on video frames and maximum wrist speed for subjective analysis. Statistical clustering with Dynamic Time Warping revealed that novice bowlers formed distinct biomechanical clusters separate from professionals, highlighting inefficiencies in their technique.

Using Dynamic Time Warping, the data were successfully aligned despite the different time frames of the videos. The novices in this study achieved skill parity with professionals, ranging from 35% to 64%. The rate of change in the right leg, foot, and wrist did not significantly impact the bowling action. Novice bowlers exhibited lower wrist speed and greater variability in joint angles, which affected ball velocity and led to inconsistent mechanics.

This study employed a data-driven approach to enhance fast bowling techniques, demonstrating the potential of AI and biomechanics to improve sports performance. Future work on this study will expand to incorporate a multi-person model, allowing for the analysis of a broader range of videos and enabling 3D pose estimation to enhance the accuracy of the angles.