Determination of the ball's spin rate and spin axis during flight

Abstract

Embodiments for determining the spin rate and spin axis of a ball in flight are disclosed. In some embodiments, the method includes capturing a time series of images of the ball in flight and the corresponding image capture time using an image sensor; inputting the image frames and image capture time into a machine learning model using at least one processor; and predicting the ball's spin axis and spin rate based on the machine learning model using at least one processor.

Description

[Technical Field]

[0001] This disclosure generally relates to sports technology and data analysis, and in particular to ball launch monitoring. [Background technology]

[0002] Data-driven sports technology and data analysis help players and coaches better understand their performance through reliable data. One such sports technology is ball launch monitoring. For example, a ball launch monitor can use video to determine various ball parameters. Two examples of ball parameters are spin rate and spin axis. Spin rate is the speed at which the ball spins around its spin axis. A high spin rate results in a higher ball trajectory and a steeper landing angle, while a low spin rate results in a lower ball trajectory and a shallower landing angle. Therefore, it is desirable to monitor the spin rate and spin axis of the ball in flight to improve player performance in ball-related sports such as cricket, baseball, and golf. [Overview of the Initiative]

[0003] Embodiments for determining the spin rate and spin axis of a ball in flight are disclosed.

[0004] In some embodiments, the method includes capturing a time series of images of a ball in flight and the corresponding image capture time using an image sensor; inputting the image frames and image capture time into a machine learning model using at least one processor; and predicting the ball's spin axis and spin rate based on the machine learning model using at least one processor.

[0005] In some embodiments, the machine learning model includes a first neural network trained to predict the seam norm of a ball in 3D space based on a training set of ball images; a second neural network trained to predict the spin rate of a ball based on the predicted seam norm and a time difference obtained from the training set of ball images; and a third neural network trained to predict the spin axis of a ball based on the ratio of the predicted seam norm and the time difference.

[0006] In some embodiments, the training set of ball images includes composite ball images of different types of balls, where the ball images have uneven lighting, complex backgrounds, occlusion, or artifacts.

[0007] In some embodiments, the training set of ball images includes expanded ball images of different types of balls, where each ball image has at least one of random color, random contrast, random brightness, random sharpness, random shift, or random noise level.

[0008] In some embodiments, a randomly generated synthetic seam norm is used to train the first neural network.

[0009] In some embodiments, the first, second, and third neural networks are trained using batch processing, stochastic gradient descent, Nesterov momentum without dampening, weight decay, or a cosine-shaped learning rate.

[0010] In some embodiments, the first neural network, the second neural network, and the third neural network are recurrent neural networks.

[0011] In some embodiments, the first neural network comprises M blocks, each block including a transition unit for downsampling the feature size and N residual units following it, where N and M are integer values greater than 0.

[0012] In some embodiments, each of the second neural network and the third neural network includes M blocks, each block having a fully connected layer for increasing the number of inputs and N residual units following it, each residual unit including a first fully connected layer, a first activation function following it, a second fully connected layer following it, and a second activation function following it, where N and M are integer values greater than 0.

[0013] In some embodiments, the ball is a cricket ball.

[0014] Other embodiments are directed to apparatuses and computer-readable media.

[0015] Certain embodiments described herein provide one or more of the following advantages. Existing ball monitoring applications segment and extract the ball's seam line as a feature and perform feature matching. However, due to non-uniform lighting, artifacts, logos, occlusions, and different types of balls, it is difficult to segment and extract the entire seam line. Also, performing feature matching is very time-consuming. Different from these existing ball monitoring applications, the disclosed embodiments use a machine learning model (e.g., a deep learning network) trained on synthetic and / or augmented ball images to directly measure the seam norm from a ball image in order to ensure that all possible challenging situations are included in the training data for the machine learning model. Further, the disclosed machine learning model has separate spin rate and spin axis regression nodes, which make the prediction of ball parameters robust to outliers.

Brief Description of the Drawings

[0016] [Figure 1A] This is a perspective view of cricketball. [Figure 1B] Figure 1A shows the seam norm of a cricket ball. [Figure 2] This is a three-dimensional (3D) diagram showing how the seam norm of a ball in flight or motion moves over time in three-dimensional (3D) space. [Figure 3] This is a block diagram of a deep learning network for predicting the spin rate and spin axis of a ball in flight or motion, according to one or more embodiments. [Figure 4] One or more embodiments of a seam-norm regression neural network are shown. [Figure 5] One or more embodiments of a spin rate regression neural network are shown. [Figure 6] One or more embodiments of a spin-axis regression neural network are shown. [Figure 7A] This document demonstrates the generation of synthetic seam norms according to one or more embodiments. [Figure 7B] This document demonstrates the generation of synthetic seam norms according to one or more embodiments. [Figure 8] This is a flowchart of a process for determining the spin rate and spin axis of a ball in flight, according to one or more embodiments. [Figure 9] This is a block diagram of a ball launch monitoring system including a measuring device that utilizes the process described with reference to Figures 3 to 8, according to one or more embodiments. [Figure 10] This is a block diagram of the measuring device shown in Figure 9, according to one or more embodiments. [Modes for carrying out the invention]

[0017] Bowling in cricket is the action of propelling a ball towards a wicket defended by the batter. Understanding proper bowling mechanics is crucial for bowlers to master various bowling techniques, such as fast bowling or spin bowling. Distinguishing between different bowling techniques relies on factors such as the ball's trajectory and velocity, which are determined by the ball's angular velocity and initial velocity at the release point.

[0018] Furthermore, factors such as drift and dip also play a significant role in the direction of the ball in flight. When the ball drifts, it deviates laterally during flight, resulting in a horizontal deviation. On the other hand, dip causes the ball to descend closer to the ground than expected, resulting in a non-linear change in the ball's vertical movement. The ball's spin motion in flight disrupts the airflow, causing it to move upward and generating a downward force that causes the ball to dip. This phenomenon, known as the Magnus effect, is closely related to spin. While radar guns can measure the speed of a cricket ball, they are insufficient to provide insight into the ball's spin characteristics. Therefore, the ability to measure the ball's spin characteristics is highly insightful. Such measurements enable coaches to quantitatively evaluate key spin bowling performance variables and provide players with objective feedback on their performance.

[0019] In baseball, understanding the spin axis and spin rate can provide pitchers with information about whether a pitch is a fastball, knuckleball, curveball, etc. This information helps pitchers improve and develop their "perfect pitch."

[0020] This disclosure provides the use of seam norms from ball images using a trained deep learning neural network to estimate spin characteristics, including spin axis and spin rate. This information can be used to measure the spin characteristics of a ball, such as a cricket ball. The vector model of seam norms described herein is advantageously more efficient and effective because it does not segment and extract the seam line, and therefore does not require feature matching. It has been recognized that segmenting and extracting the entire seam line is difficult due to non-uniform lighting artifacts, logos, occlusion, and different types of balls. Furthermore, segmentation methods are time-consuming because they require feature matching. The method disclosed herein advantageously measures the seam norm directly from ball images using a neural network. Furthermore, the spin rate and spin axis regression modes are advantageously robust to outliers. For balls that surround the entire ball and therefore do not have seams defining a seam plane (e.g., baseballs, golf balls, soccer balls, footballs, basketballs), a sticker can be wrapped around the ball to serve as a substitute for a stitched seam, and the disclosed embodiments can be applied to any ball, whether or not it has a stitched seam. Physics of Spin Rate and Spin Axis

[0021] Figure 1A is a perspective view of a cricket ball 100. The cricket ball 100 has a stable and constant seam line 101. Two feature points used to train a neural network are seam normal vectors 105a(n) and 105b(n’). These vectors are perpendicular to the seam plane 104 (hereinafter also referred to as “seam norm”) as shown in Figure 1B. As will be described below, the seam norm can be used to measure the spin rate and spin axis 103 of the cricket ball 100 during flight or movement. Although the following discussion relates to a cricket ball, the process described below is applicable to any ball having a seam (e.g., a golf ball having markings, a softball, or a baseball).

[0022] Figure 2 is a 3D diagram showing how the seam norm moves over time within a 3D space. As shown in Figure 2, the seam norm of the spinning ball 100 forms two circles over time along the spin axis 103 and perpendicular to the spin axis 103. If there are N consecutive camera images capturing the ball 100, the i-th ball is captured at time t i and its seam norm in 3D space is n i =(x i ,y i ,z i ) and n i ’=(-x i ,-y i ,-z i ), where x i 2 +y i 2 +z i 2 =1 and z i >0. The spin axis 103 (V=(a,b,c)) of the ball 100 is obtained by fitting a plane to the 3D n i points according to Equation [1]. a*x i +b*y + c*z i +d = 0, i = 1,2,...N, [1] In the formula, d is the distance of the fitted plane from the origin.

[0023] The direction of the spin axis 103 can be determined using the right-hand rule from point n1 to point n2 within the circle, as shown in Figure 2. Once the spin axis 103 is determined, the spin rate is given by n relative to the spin axis 103. i from n i+1 Angle 201 and t i+1 from t i This can be determined by determining the ratio of the corresponding time difference as shown in equation [2].

[0024]

number

[0025] Figure 3 is a block diagram of a deep learning network 300 for predicting the spin rate and spin axis of a ball in flight or motion, according to one or more embodiments. In this exemplary embodiment, the input 301 to the network 300 (e.g., a deep learning network) includes eight ball images and corresponding image capture times, predicting the spin rate and spin axis of the ball, each ball image containing a spin around a different axis. In other embodiments, three or more ball images can be used. At least three points (seam norms) can determine a circle in 3D, and the spin axis is perpendicular to the circle. Once the spin axis is determined, the spin rate can be determined using any two points on the circle as shown in equation [2].

[0026] The exemplary network 300 shown in Figure 3 includes three deep learning neural networks (nodes), namely, a seam-norm regression neural network 302, a spin-rate regression neural network 304, and a spin-axis regression neural network 306. In some embodiments, the seam-norm regression neural network 302 uses batch processing to regress eight seam-norm vectors 303 from an input 301. The spin-rate regression neural network 304 and the spin-axis regression neural network 306 regress the spin rate 307 and spin axis 308, respectively, from the seam-norm vectors 303 and the corresponding image capture time 305 inputs.

[0027] Figure 3 shows a deep learning network, but please note that other machine learning models are also possible.

[0028] In some embodiments, a specific method for baseballs can be used without using stickers or other seam substitutes. Cricket balls are 2-fold symmetric, while baseballs are 4-fold symmetric. Therefore, two seam planes can be used to ensure that any image of a baseball captured by a camera is unique. In some embodiments, the same seam-norm regression neural networks, spin-axis regression neural networks, and spin-quantity regression neural networks used for cricket balls can be used for baseballs, but with the following differences: For the seam-norm regression neural network, the input is an image of a baseball, and the output is two seam-norm vectors. One seam-norm vector is (x,y,z), and the vector perpendicular to it is (x',y',z'). For the spin-axis regression neural network and rate regression neural network, the input is a combination of all seam-norm vectors and time differences. The outputs are the spin axis and spin rate, respectively.

[0029] Figure 4 shows a seam-norm regression neural network 400 in one or more embodiments. Instead of seam segmentation, the seam-norm regression neural network 400 is trained directly. In this exemplary embodiment, the input 401 is a 64x64 ball image, and the output 402 is its seam-norm vector. In this exemplary configuration, the seam-norm regression neural network 400 comprises five blocks, each block having one transition unit for downsampling feature size followed by two residual units. Each block has three subblocks.

number

number

[0030] In some embodiments, the blocks of the neural network 400 may be arranged in series, the blocks may be daisy-chained, and the last block may be connected to a fully connected layer. In some embodiments, the blocks of the neural network 400 may be arranged in parallel, the blocks may be daisy-chained, and the last block may be connected to a fully connected layer. In some embodiments, real images of balls are used along with ground truth data to train the neural network 400. In other embodiments, synthetic data is used along with real image data, or instead, to train the neural network 400. For example, synthetic ball images can be generated using non-uniform lighting, complex background occlusion, artifacts, and different types of balls, resulting in additional synthetic ball images in which the seam norm is uniformly distributed on a unit sphere.

[0031] In some embodiments, real and / or composite ball images can be augmented during training by modifying the color, contrast, brightness, sharpness, shift, and noise levels of the ball images. The augmentation levels of color, contrast, brightness, and sharpness may be random, for example, in the range of about 0.8 to about 1.2. In this exemplary embodiment, the maximum shift is 4 pixels, which is the maximum error caused by ball detection. For noise, Gaussian blur or other suitable image processing techniques can be applied, for example, using sigma values ​​that vary from about 0.1 to about 9.0.

[0032] In some embodiments, augmentation level 1.0 corresponds to the original image, i.e., no data augmentation is applied. Data augmentation is used to minimize the gap between training data and test data. For example, if the training image has a brightness level of 10 and the test image has a brightness level of 8, there is a difference of 2 = 10⁻⁸ in brightness levels. This data shift causes lower test accuracy for the pre-trained model. For example, by applying data augmentation using a random ratio between 0.8 and 1.2, the pre-trained model can have good accuracy with brightness levels between 8 (0.8 * 10) and 12 (1.2 * 10). In some embodiments, this range can be fine-tuned to have better accuracy for a set of test images. Thus, the brightness range may be wider, but it cannot be less than 0, which corresponds to a completely black image. Note that the brightness range should not be too high, as the maximum intensity is 255. Therefore, in this example, the top ratio of brightness is obtained by 255 / (current training brightness).

[0033] Figure 5 shows a spin rate regression neural network 500 in one or more embodiments, and Figure 6 shows a spin axis regression neural network 600 in one or more embodiments. In this example, both the spin rate regression neural network 500 and the spin axis regression neural network 600 share the same architecture. It should be understood that different architectures of the regression neural network may also be used. Inputs 501, 601 are 1D arrays containing 16 seam norms and their corresponding time differences. Output 502 for neural network 500 contains one node for regression of spin rate, and output 602 for neural network 600 contains three nodes for regression of spin axes in 3D space.

[0034] Both neural networks 500 and 600 each have five blocks, each block including one fully connected layer for increasing the number of inputs and four residual units. Each residual unit includes a first fully connected layer coupled to a first activation function (e.g., a rectified linear unit (ReLU) activation function), followed by a second fully connected layer, followed by a second activation function (e.g., a ReLU activation function). Similar to the seam-norm neural network 400, in this exemplary embodiment, neural networks 500 and 600 have a batch size of 2560, a Nesterov momentum of 0.9 without damping, and a weight decay of 10. -4 The model is trained over 1000 epochs using batch processing with stochastic gradient descent. The learning rate is cosine-shaped, decreasing from 0.1 to 0.

[0035] Figures 7A and 7B illustrate the generation of a composite seam norm vector according to one or more embodiments. Referring to Figure 7A, inputs 501 and 601 for a spin rate regression neural network 500 and a spin axis regression neural network 600 are shown, respectively. In this exemplary embodiment, inputs 501 and 601 are 1D arrays each containing 16 seam norms, and n i =(x i ,y i ,z i ) and n i '=(-x i ,-y i ,-z i ), including, where i=1,2,...8. Furthermore, the 1D array of input 501 has a time difference ΔT j The 1D array of input 601 includes the time difference ratio ΔT j / ΔT j+1 The formula includes j = 1, 2, 3, 4...7.

[0036] The spin rate of a cricket ball is typically in the range of approximately 400 to 3000 revolutions per minute (rpm), and the spin axis can be any unit vector in 3D space. The spin axes can be uniformly distributed along the entire surface of the unit sphere, and each spin axis has a random spin rate in the range of approximately 300 to 3600 rpm. As shown in Figure 7B, in some embodiments, one spin axis is selected from the surface of the unit sphere, and another point n1 is selected as the initial seam norm. In some embodiments, a random number between 300 and 3600 is selected as the spin rate, and n1 is rotated with different time differences to generate seven seam norms. In some embodiments, where the ball image is captured by two cameras (a spin camera and a speed camera), there are two different timings. For the speed camera, the time difference is constant, for example, 1.0 / 220 s. For the spin camera, the time differences are, for example, 8, 9, 10, 11, and 12 ms.

[0037] To make both spin rate regression and spin axis regression more robust, in some embodiments, three of the eight seam norms are allowed as outliers, each with a difference of 2 degrees from ground truth. During training, three seam norms are randomly selected, initialized with random values, and each seam norm is rotated, for example, by a random angle between 0 and 2 degrees in a random direction. Exemplary process

[0038] Figure 8 is a flowchart of a process for determining the spin rate and spin axis of a ball in flight, according to one or more embodiments. Process 800 may be carried out in a ball launch monitoring system 900, for example, as described with reference to Figures 9 and 10.

[0039] Process 800 includes capturing a time series of images of the ball in flight and the corresponding image capture times using an image sensor (801), inputting the image frames and image capture times into a machine learning model (802), and predicting the ball's spin axis and spin rate based on the machine learning model (803). Each of these steps is described above with reference to Figures 3 to 7.

[0040] Exemplary ball launch monitoring system Figure 9 is a block diagram of a ball launch monitoring system 900 according to one or more embodiments. A measurement device or measuring device 902 is positioned on the floor or ground behind the user 905 who hits the cricket ball 904 with a bat 903 (e.g., behind the cricket ball 904). Further data processing can be performed via a cloud service. A viewing device 901 (e.g., a computer or mobile device) is provided for the user 905 or coach to visualize the measurement results on a display screen (e.g., a computer or mobile device display screen). Alternatively or additionally, the measuring device 902 may be positioned on the user 905's side.

[0041] Figure 10 is a block diagram of the measuring device 902 of Figure 9, according to one or more embodiments. In some embodiments, the measuring device 902 may include two sets of cameras. The first set of cameras includes two high-speed, low-resolution cameras 1004A, 1004B that form a first stereo system for measuring the movement of the bat and cricket ball during a swing of the bat 903 by the user 905. The second set of cameras includes two high-resolution cameras 1006A, 1006B that form a second stereo system for measuring the ball trajectory. In some embodiments, the second set of cameras may have a lower speed than the first set of cameras. In some embodiments, the illumination system 1005 ensures properly illuminated images for 3D scanning of the bat 903, the radar unit 1003 provides timing and sequence of images captured by cameras 1004A, 1004B, 1006A, and 1006B, the computing and control system 1002 performs real-time processing of the captured images and radar signals, and the wireless communication system 1001 sends data to the viewing device 901 shown in Figure 9.

[0042] This specification includes details of many specific implementations, which should not be interpreted as limitations on the scope of any invention or claim, but rather as descriptions of features specific to particular embodiments of a particular invention. Certain features described herein in the context of a separate embodiment may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented separately in multiple embodiments or in any preferred subcombination. Furthermore, features are described above as acting in a particular combination, and may even be initially claimed as such, but in some cases one or more features of a claimed combination may be removed from that combination, and the claimed combination may cover a subcombination or a variation of a subcombination.

[0043] Similarly, although the operations are shown in a specific order in the drawings, this should not be understood as requiring that such operations be performed in a specific illustrated order or sequential order, or that all shown operations be performed, in order to achieve the desired result. In certain circumstances, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged in multiple software products. The inventions disclosed herein include the following: [Aspect 1] It is a method, Using an image sensor, capture a time series of images of a ball in flight and the corresponding image capture time. Using at least one processor, the image frame and the image capture time are input to a machine learning model, Using at least one of the aforementioned processors, predict the spin axis and spin rate of the ball based on the machine learning model, Methods that include... [Aspect 2] The method according to embodiment 1, wherein the machine learning model includes: a first neural network trained to predict the seam norm of the ball in 3D space based on a training set of ball images; a second neural network trained to predict the spin rate of the ball based on the predicted seam norm and a time difference obtained from the training set of ball images; and a third neural network trained to predict the spin axis of the ball based on the ratio of the predicted seam norm and the time difference. [Aspect 3] The method according to embodiment 2, wherein the training set of ball images includes composite ball images of different types of balls, and the ball images have uneven lighting, complex backgrounds, occlusion, or artifacts. [Aspect 4] The method according to embodiment 2, wherein the training set of ball images includes expanded ball images of different types of balls, and the ball images have at least one of random color, random contrast, random brightness, random sharpness, random shift, or random noise level. [Aspect 5] The method according to embodiment 2, wherein a randomly generated synthetic seam norm is used to train the first neural network. [Aspect 6] The method according to embodiment 2, wherein the first neural network, the second neural network, and the third neural network are trained using batch processing, stochastic gradient descent, Nesterov momentum without dampening, weight decay, or a cosine-shaped learning rate. [Aspect 7] The method according to embodiment 2, wherein the first neural network, the second neural network, and the third neural network are recurrent neural networks. [Aspect 8] The method according to embodiment 6, wherein the first neural network comprises M blocks, each block comprising a transition unit for downsampling the feature size, followed by N residual units, where N and M are integer values ​​greater than 0. [Aspect 9] The method according to embodiment 6, wherein the second neural network and the third neural network each comprise M blocks, each block having a fully connected layer for increasing the number of inputs, followed by N residual units, each residual unit comprising a first fully connected layer, followed by a first activation function, followed by a second fully connected layer, followed by a second activation function, where N and M are integer values ​​greater than 0. [Aspect 10] The method according to embodiment 1, wherein the ball is a cricket ball. [Aspect 11] It is a system, At least one camera, At least one processor, A memory for storing instructions, which, when executed by the at least one processor, the at least one processor, Using at least one of the aforementioned cameras, capture a time series of images of the ball in flight and the corresponding image capture time. Inputting the image frame and the image capture time into a machine learning model, Using the aforementioned machine learning model, predict the spin axis and spin rate of the ball, A memory that stores instructions for performing an operation including, A system equipped with these features. [Aspect 12] The system according to embodiment 11, wherein the machine learning model includes: a first neural network trained to predict the seam norm of the ball in 3D space based on a training set of ball images; a second neural network trained to predict the spin rate of the ball based on the predicted seam norm and a time difference obtained from the training set of ball images; and a third neural network trained to predict the spin axis of the ball based on the ratio of the predicted seam norm and the time difference. [Aspect 13] The system according to embodiment 12, wherein the training set of ball images includes composite ball images of different types of balls, and the ball images have uneven lighting, complex backgrounds, occlusion, or artifacts. [Aspect 14] The system according to embodiment 12, wherein the training set of ball images includes expanded ball images of different types of balls, and the ball images have at least one of random color, random contrast, random brightness, random sharpness, random shift, or random noise level. [Aspect 15] The system according to embodiment 12, wherein a randomly generated synthetic seam norm is used to train the first neural network. [Aspect 16] The system according to embodiment 12, wherein the first neural network, the second neural network, and the third neural network are trained using batch processing, stochastic gradient descent, Nesterov momentum without dampening, weight decay, or a cosine-shaped learning rate. [Aspect 17] The system according to embodiment 12, wherein the first neural network, the second neural network, and the third neural network are recurrent neural networks. [Aspect 18] The system according to embodiment 17, wherein the first neural network comprises M blocks, each block comprising a transition unit for downsampling the feature size and N residual units thereafter, where M and N are integer values ​​greater than 0. [Aspect 19] The system according to embodiment 17, wherein the second neural network and the third neural network each include M blocks, each block having one fully connected layer for increasing the number of inputs and N residual units thereafter, each residual unit including a first fully connected layer and a first activation function thereafter, a second fully connected layer thereafter and a second activation function thereafter, where M and N are integer values ​​greater than 0. [Aspect 20] The system according to embodiment 11, wherein the ball is a cricket ball.

Claims

1. It is a method, Using an image sensor, capture a time series of images of a ball in flight and the corresponding image capture time. Using at least one processor, the image frame and the image capture time are input to a machine learning model, Using the at least one processor, predict the spin axis and spin rate of the ball based on the machine learning model, Includes, The aforementioned machine learning model, A first neural network trained to predict the seam norm of the ball in 3D space based on a training set of ball images, A second neural network trained to predict the spin rate of the ball based on the predicted seam norm and the time difference obtained from the training set of ball images, A third neural network trained to predict the spin axis of the ball based on the predicted seam norm and the ratio of the time difference, Methods that include...

2. The method according to claim 1, wherein the training set of ball images includes composite ball images of different types of balls, and the ball images have uneven lighting, complex backgrounds, occlusion, or artifacts.

3. The method according to claim 1, wherein the training set of ball images includes expanded ball images of different types of balls, and the ball images have at least one of random color, random contrast, random brightness, random sharpness, random shift, or random noise level.

4. The method according to claim 1, wherein a randomly generated synthetic seam norm is used to train the first neural network.

5. The method according to claim 1, wherein the first neural network, the second neural network, and the third neural network are trained using batch processing, stochastic gradient descent, Nesterov momentum without dampening, weight decay, or a cosine-shaped learning rate.

6. The method according to claim 1, wherein the first neural network, the second neural network, and the third neural network are regressive neural networks.

7. The method according to claim 5, wherein the first neural network comprises M blocks, each block comprising a transition unit for downsampling the feature size, followed by N residual units, where N and M are integers greater than 0.

8. The method according to claim 5, wherein the second neural network and the third neural network each comprise M blocks, each block having a fully connected layer for increasing the number of inputs, followed by N residual units, each residual unit comprising a first fully connected layer, followed by a first activation function, followed by a second fully connected layer, followed by a second activation function, where N and M are integer values ​​greater than 0.

9. The method according to claim 1, wherein the ball is a cricket ball.

10. It is a system, At least one camera, At least one processor, A memory for storing instructions, which, when executed by the at least one processor, the at least one processor, Using at least one of the aforementioned cameras, capture a time series of images of the ball in flight and the corresponding image capture time. Inputting the image frame and the image capture time into a machine learning model, Using the aforementioned machine learning model, predict the spin axis and spin rate of the ball, A memory that stores instructions for performing an operation including, Equipped with, The aforementioned machine learning model, A first neural network trained to predict the seam norm of the ball in 3D space based on a training set of ball images, A second neural network trained to predict the spin rate of the ball based on the predicted seam norm and the time difference obtained from the training set of ball images, A third neural network trained to predict the spin axis of the ball based on the predicted seam norm and the ratio of the time difference, A system that includes this.

11. The system according to claim 10, wherein the training set of ball images includes composite ball images of different types of balls, and the ball images have uneven lighting, complex backgrounds, occlusion, or artifacts.

12. The system according to claim 10, wherein the training set of ball images includes expanded ball images of different types of balls, and the ball images have at least one of random color, random contrast, random brightness, random sharpness, random shift, or random noise level.

13. The system according to claim 10, wherein a randomly generated synthetic seam norm is used to train the first neural network.

14. The system according to claim 10, wherein the first neural network, the second neural network, and the third neural network are trained using batch processing, stochastic gradient descent, Nesterov momentum without dampening, weight decay, or a cosine-shaped learning rate.

15. The system according to claim 10, wherein the first neural network, the second neural network, and the third neural network are recurrent neural networks.

16. The system according to claim 15, wherein the first neural network comprises M blocks, each block comprising a transition unit for downsampling the feature size and N residual units thereafter, where M and N are integer values ​​greater than 0.

17. The system according to claim 15, wherein the second neural network and the third neural network each include M blocks, each block having one fully connected layer for increasing the number of inputs and N residual units thereafter, each residual unit including a first fully connected layer and a first activation function thereafter, a second fully connected layer thereafter and a second activation function thereafter, where M and N are integer values ​​greater than 0.

18. The system according to claim 10, wherein the ball is a cricket ball.

Images