Ball Spin Estimation Using Deep Learning and Ellipse Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining the spin axis and spin rate of a moving ball, such as using fixed markers and radar/camera technology, are rigid and costly, requiring specially marked balls and complex feature extraction processes.
Innovation Solution
A deep learning network is trained on series of images of spinning balls to estimate spin rate and axis using feature points, forming 3D circles and 2D ellipses, with data augmentation and convolutional neural networks for accurate prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fixed markers are used on the ball to determine spin axis and spin rate, then measurement precision is improved, but device complexity and cost increase due to requiring specially marked balls and additional hardware
Solution Approach 1:
The patent uses image copies from video frames to create synthetic training data. Instead of requiring physical markers on balls, the system captures images of balls with natural features, creates synthetic views through image processing and transformation, and uses these copied images to train the deep learning network. This eliminates the need for special markers while maintaining measurement capability.
Solution Approach 2:
The patent replaces the mechanical/optical marker-based measurement system with a computational deep learning system. Instead of using physical markers that reflect light to radar/camera systems, the invention uses a trained neural network that processes natural ball images to extract spin information, substituting mechanical measurement with computational analysis.
2Measurement precision
If fixed markers and feature extraction techniques are used, then spin measurements can be obtained, but the process becomes time consuming and computationally intensive
Solution Approach 1:
The patent performs preliminary action by training the deep learning network offline before actual spin measurements are needed. The network is trained on large datasets of ball images with various spin conditions, lighting, and angles. This pre-training allows the system to perform rapid spin measurements during actual use without time-consuming feature extraction and matching operations.
Solution Approach 2:
The patent replaces time-consuming mechanical feature extraction and matching processes with a pre-trained deep learning network that can directly estimate spin parameters from images. The computational heavy lifting is done during training, allowing real-time or near-real-time inference during actual measurements.
3Measurement precision
If specially marked balls are required for spin measurement, then measurement accuracy is improved, but adaptability decreases since users must acquire specific equipment
Solution Approach 1:
The patent achieves universality by designing a system that works with any ball type that has visible surface features. The deep learning network is trained on diverse ball images including different types (golf balls, baseballs, tennis balls) with various natural features like dimples, seams, and patterns. This allows the same system to measure spin across multiple sports and ball types without requiring special markers or equipment.
Solution Approach 2:
The patent uses image copying and transformation to create synthetic training data that represents various ball types and conditions. By capturing images of real balls and creating transformed copies with different spins, angles, and lighting, the system learns to recognize spin patterns across diverse ball types without needing physical markers on each ball type.
Data Source
AI summary
Embodiments are disclosed for determining a spin rate and axis of a ball using deep learning. In some embodiments, a method comprises: training a deep learning network on training images of spinning balls, each spinning ball having at least one feature point in a time series that forms a two-dimensional (2D) ellipse image in a 2D plane; capturing, with at least one camera, a series of images of a ball; predicting, with the trained deep learning network, spin measurements associated with the ball based on the series of images; determining a spin rate of the ball based on the spin measurements; determining coefficients of a 2D ellipse model based on the spin measurements and the spin rate; and determining, with the at least one processor, a spin axis of the ball in 3D space based on the 2D ellipse model and the spin rate.