Object Detection and Land-marking for Human Action Recognition in Single to Multiplayer Sports by Hybrid Approach
Aryaman Tiwary, Jayant Dhidhi, Vijay Kumar Gupta, Bipin Chandra Mandi · 2024
Sporting events require strategic planning and proper analysis of the performance of players. Decision-making before or after a game helps the coaches to provide the players with appropriate guidance and training. Human Action Recognition models are the best tool to implement in this scenario. It not only helps improve the gameplay but also focuses on other factors. The broadcasting team can use it to show the highlights and Physicists can have a look into how a player is getting injured and prevent any future mishaps. Our paper proposes the use of two different algorithms to detect and estimate players, positions during a match in any sport. Firstly, object detection is done using YOLOv7 model then the pose is estimated using OpenPose. Additionally, we delve into a comparative analysis of different pose-estimation algorithms. Furthermore, we assess the performance of YOLOv7 in both single-player and multiplayer sports contexts, providing valuable insights into its efficacy and applicability.The average detection time per frame is $\mathbf{1. 4}$ seconds, with a peak accuracy of $\mathbf{0. 9 9}$ for single sports and 0.871 for multiple sports involving player and object detection, resulting in an overall accuracy of 0.8649 for the proposed framework. Through this comprehensive approach, we aim to revolutionize sports analytics and elevate the standard of performance analysis in the realm of athletics.