Recent advances in computer vision for sports are aimed at tackling problems such as skeleton detection to analyze body joints and postures, ball, shuttle, or player localization and tracking, and improving the quality of decision-making in sports such as LBW in cricket or line-call in tennis. Multiple problem dimensions of different sports continue to pose various technical challenges.

In the problem domain of object tracking and analysis, while existing literature has focused on tracking balls or humans, only a few works target ball motion analysis. The current technique to estimate the translational and rotational velocities of a cricket ball involves multiple high-speed specialized cameras and a tracking system. In our work, we target a low-cost computer vision pipeline for this task. We began with a framework requiring two cameras to extract the 3D spatiotemporal information of the ball and estimate its rotational magnitude and direction at short time intervals. We record our data with GoPro Hero Black 10 cameras with 120 fps. To this end, we propose a non Deep-Learning (DL)-based 2D ball detection and localization framework that exploits photometric, geometric, and temporal consistency across video frames.This is followed by 3D trajectory estimation and plane fitting to find the spin axis. Next, we attempted ball analysis with a single camera. We analytically derive a closed-form linear relationship between optical flow and angular velocity by exploiting physical constraints that stem from motion and projective geometry. We synthetically create a 3D point cloud representing the ball and verify the proposed approach. We study the trade-off between camera frame rate and rotational speed and show when this relationship can break down. We could obtain satisfactory results for real rotational ball motion.

We are also working on the problem of dynamic novel view synthesis of complex deformable dynamic events using Neural Radiance Field (NeRF). Most of the existing works focus on static scenes; a few address dynamic situations. We observe that prior dynamic scenes in NeRF literature assume restricted motion or a simple motion targeting primarily humans. We aim to address the novel view synthesis problem in a single-view setup for more complex scenes. We have obtained preliminary results on some videos captured from our mobile phones. We plan to collect real data from different sports centers in India for carrying out relevant sports-related research activities.