Shot put three-dimensional space trajectory analysis and score prediction method based on 2D image

Abstract

The invention discloses a shot-put three-dimensional space trajectory analysis and score prediction method based on a 2D image, and relates to the technical field of computer vision. Establishing a 3D space curve model for performing multi-angle condition fitting, proposing a projection cost function to calculate a difference value between a 3D space curve and a point on a two-dimensional image plane, and correcting the 3D space curve for the difference value by using an optimization algorithm so as to enable a 2D projection of the 3D space curve to coincide with a multi-frame shot position; in addition, the position of a landing point is calculated in a self-adaptive mode by means of fusion of a landing first rebound curve and a rolling track, and the relation between the physical distance and pixels is automatically calculated by means of a gravitation formula, so that the actual landing physical distance is automatically calculated. According to the shot three-dimensional space trajectory analysis and score prediction method based on the 2D image provided by the invention, space curve fitting and high-precision landing point distance calculation are realized based on the 2D image, and the shot tracking cost is greatly reduced.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a three-dimensional spatial trajectory analysis and performance prediction method of a shot put based on a 2D image. Background Art

[0002] In the context of the development of AI sports, the use of AI algorithms to calculate sports scores and evaluate comprehensive performance has become a development trend. For shot put sports, traditional manual distance measurement and performance analysis methods are difficult to achieve objectively, accurately, and efficiently. In this context, a low-cost, easy-to-deploy solution is urgently needed to achieve efficient and high-precision shot put distance measurement.

[0003] The existing methods of processing digital images based on computer vision algorithms to achieve 2D image shot put detection and trajectory prediction rely on target detection, target tracking, quadratic curve fitting and other schemes to predict the landing point in the image; they rely on pre-calibration schemes to measure the distance at multiple locations on the image in advance and then place markers for the image processing algorithm to estimate the relationship between pixels and real physical distances, thereby realizing the physical displacement calculation of the shot put landing. The scheme that relies on 3D positioning requires the use of three active beacons for spatial positioning based on algorithms such as RSSI and TDOA or spatial positioning by image processing algorithms, thereby achieving accurate spatial position tracking and landing point prediction.

[0004] Existing methods that rely on 2D images cannot model three-dimensional space. When the plane of the ball thrown by the athlete is not perpendicular to the camera normal vector, the motion trajectory captured by the camera is a non-standard quadratic function curve on the two-dimensional image. When using higher-order functions or complex multi-layer perceptron models (MLPs) to fit, although the sampling points can be effectively fitted, it is easy to cause underfitting and prediction errors when there are insufficient detection points in complex scenes. Methods based on high-precision spatial positioning usually rely on triangulation positioning algorithms such as received signal strength (RSSI) or time difference of arrival (TDOA). Although this method has high accuracy, it has high requirements for the coordinate configuration of the beacon, and the actual deployment cost is high and too complicated. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a three-dimensional spatial trajectory analysis and performance prediction method for shot put based on 2D images, realize spatial curve fitting and high-precision landing point distance calculation based on 2D images, and greatly reduce the tracking cost of shot put.

[0006] The present invention provides a 2D image-based three-dimensional trajectory analysis and performance prediction method for shot put, comprising:

[0007] Target detection process: Make the normal plane of the camera perpendicular to the ground, then obtain a two-dimensional image of the process from the start of the throw to the stop of the shot put, detect the position of the ball through the target detection model, and synthesize a discrete trajectory with a timestamp;

[0008] Spatial curve model fitting process: establishing a 3D spatial curve model, wherein the 3D spatial curve model is a cosine angle side-throw model or a spatial oblique-throw curve model, calculating the difference between the 3D spatial curve and the points on the two-dimensional image plane through a projection cost function, and then using an optimization algorithm to correct the difference, so that the 2D projection of the 3D spatial curve coincides with the position of the multi-frame shot put;

[0009] Landing point prediction process: determine the two inflection points in the discrete trajectory to obtain the oblique curve trajectory, the first rebound trajectory and the rolling trajectory; adaptively calculate the landing point position by integrating the first rebound trajectory and the rolling trajectory;

[0010] Actual distance calculation process: The relationship between the actual height of the highest point and the pixel is automatically calculated based on the gravity formula, thereby automatically calculating the actual physical distance to the ground.

[0011] The technical solution provided in the embodiments of the present invention has at least the following technical effects:

[0012] By establishing a 3D space curve model and proposing a projection cost function to calculate the difference between the space curve and the points on the two-dimensional image plane, the 2D projection of the 3D space curve is made to coincide with the position of the shot put in multiple frames, thus solving the problem that the existing methods relying on 2D images cannot model the three-dimensional space. An auxiliary method based on the fusion of a rebound trajectory and a rolling trajectory is proposed to model the position of the landing point on the curve and adaptively calculate the landing point position. By realizing space curve fitting and high-precision landing point distance calculation based on 2D images, the cost of shot put tracking is greatly reduced.

[0013] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The present invention will be further described below in conjunction with embodiments with reference to the accompanying drawings.

[0015] Figure 1 Schematic diagram of the relationship between the actual flight trajectory of the cosine angle side-jet model and its projection onto a 2D image in an embodiment of the present invention;

[0016] FIG. 2( a ) is a graph showing the random initialization parameters a, b, c of the spatial oblique parabola model according to an embodiment of the present invention and the curve l after passing through p x, p y , p z Schematic diagram of the curve l' after parameter rotation;

[0017] FIG2( b ) is a schematic diagram of the projection of FIG2( a ) on the xoy plane;

[0018] Figure 3 This is an example diagram of inflection point detection in an embodiment of the present invention;

[0019] FIG4( a ) is a schematic diagram showing the calculation principle of the linear mapping relationship of Pix-Real coordinate mapping pixel distance in an embodiment of the present invention;

[0020] FIG4( b ) is a schematic diagram showing the calculation principle of the actual distance d between the landing point and the starting point of the Pix-Real coordinate mapping in an embodiment of the present invention;

[0021] Figure 5 It is a schematic diagram of the overall process of the method of the embodiment of the present invention;

[0022] FIG6( a ) is a schematic diagram showing the results of fitting using a traditional quadratic curve model according to an embodiment of the present invention;

[0023] FIG6( b ) is a schematic diagram of the results of using a spatial oblique parabola model fitting according to an embodiment of the present invention;

[0024] FIG. 7( a ) is a schematic diagram of the prediction effect of an embodiment of the present invention on a hard ground;

[0025] FIG7( b ) is a schematic diagram of the prediction effect of an embodiment of the present invention when the soft ground cannot rebound;

[0026] FIG7( c ) is a schematic diagram of the prediction effect when the application scene of the embodiment of the present invention is blocked;

[0027] FIG. 7( d ) is a schematic diagram of the predicted effect when the rolling trajectory is offset due to a pothole in the ground according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The embodiment of the present invention provides a three-dimensional spatial trajectory analysis and performance prediction method for shot put based on 2D images, realizes spatial curve fitting and high-precision landing point distance calculation based on 2D images, and greatly reduces the tracking cost of shot put.

[0029] The technical solution in the embodiment of the present invention has the following general idea:

[0030] Based on the target detection model trained on the shot put dataset, the algorithm design and application of the multi-frame shot put position obtained by the model are carried out to achieve high-precision landing point distance calculation. In order to fit the multi-angle situation, a 3D space curve model is established, and a projection cost function is proposed to calculate the difference between the space curve and the points on the two-dimensional image plane. The Levenberg-Marquardt (LM) optimization algorithm is used to correct the 3D space curve for the difference value, so that the 2D projection of the 3D space curve coincides with the multi-frame shot put position; in addition, the landing point position is adaptively calculated by relying on the fusion of its first rebound curve and rolling trajectory, and the gravity formula is used to automatically calculate the relationship between physical distance and pixels, thereby automatically calculating the actual physical distance of landing.

[0031] 1. Model building

[0032] 1.1 Shot Put Detection Model

[0033] Use the YOLOv8 detection open source algorithm as the shot put target detection model.

[0034] 1.2 Space curve model

[0035] Under ideal conditions, ignoring external interference factors such as air resistance, the ball's motion state is standard oblique projection motion, and its trajectory is a standard quadratic curve, which can be modeled as formula (1).

[0036] y=a·x 2 +b·x+c (1)

[0037] It is known that the diameter of a standard shot put is between 95 mm and 130 mm, the air density is a common value of 1.225 kg / m3, the ball resistance coefficient is about 0.4 to 0.5, and the speed of a normal person throwing a shot put is about 10 m / s to 14 m / s. From the traditional quadratic curve model of formula (2), it can be known that the actual air resistance of the shot put is about 0.2 N to 0.7 N, so the effect of air resistance is ignored in the modeling of the present invention. d is the air resistance, C d is the drag coefficient, ρ is the air density, A is the cross-sectional area of ​​the shot put, and v is the velocity of the shot put.

[0038]

[0039] The actual throwing venue of the athlete is a fan-shaped area. The angle between the plane where the shot put flight trajectory is located and the camera normal plane (the plane perpendicular to the camera normal vector) is recorded as θ. When the camera normal plane is perpendicular to the ground, the present invention proposes to model it as a cosine angle side throw model, as shown in formula (3).

[0040] y=a·(x·cos(θ)) 2 +b·x·cos(θ)+c (3)

[0041] A three-dimensional Cartesian coordinate system (x, y, z) is constructed with the starting point as the origin. The angle between the plane woy where the throwing direction vector w is located and the plane xoy parallel to the camera normal plane is θ. The actual flight trajectory is the oe curve, and the image obtained on the 2D image of the camera is the oe' curve. The relationship between the actual flight trajectory and the projection onto the 2D image can be obtained from formula (3), as follows: Figure 1 shown.

[0042] The cosine angle side throw model can theoretically fit the actual shot put flight trajectory to a very high degree, but it is necessary to ensure that the starting point is on the y-axis. Therefore, this solution has better fitting effect and speed, but it has extremely high requirements on the detection accuracy of the target detection algorithm and the video frame rate to ensure the detection of the shot put release frame.

[0043] In order to reduce the demand for shot frame detection, a spatial oblique throw curve model is proposed. First, the spatial initialization curve is established as shown in formula (4), and the spatial rotation matrix R is defined. The direction of the rotation axis is known to be parallel to the y-axis, so there is a rotation axis normal vector, as shown in formula (5). The point through which the rotation axis passes is used as the parameter to be optimized, and the detected shot put throwing position is used as the initialization value. The spatial curve can be calculated through the rotation matrix, as shown in formula (6).

[0044]

[0045] According to the above formula, randomly initialize the parameters a, b, c, p x , p y , p z , the obtained curve is shown in Figure 2. Figure 2 (a) shows the curve l with random initialization parameters a, b, c and the curve l after p x , p y , p z Figure 2(b) shows the projection of the curve l' after parameter rotation on the xoy plane.

[0046] 1.3 Landing point prediction modeling

[0047] In order to model the landing point, ground information needs to be provided as reference to the model. The present invention proposes an auxiliary method based on the fusion of a rebound trajectory and a rolling trajectory to model the landing point on the curve. Here, the spatial position sequence of the target detection from the start to the stop is defined as Where n is the position of the ball in a total of n frames detected from the start of the toss to the stop of the ball, and 3 is a total of three data in each frame: the timestamp of the position, the horizontal position, and the vertical position.

[0048] Since the trajectory of the flight in the air is a parabola, the inflection point on the trajectory is calculated using discrete second-order differentials. The Taylor formula gives us formula (7) and formula (8). Combining the two formulas, ignoring the higher-order terms (i.e., the Lagrange remainder O(h 3 ))Calculate the second-order differential as shown in formula (9).

[0049]

[0050] In order to obtain the inflection point where the discrete curve changes from an upward convex curve, the three consecutive frame positions are judged by formula (10) to determine whether they are inflection points. If the x of formula (10) is satisfied, i is the inflection point. In order to avoid the misjudgment of the inflection point caused by the slight error in the image positioning of the target detection model, Δ=0.01 is taken. This method is used to detect the inflection point on the discrete curve, such as Figure 3 shown.

[0051] f″(x i )>0+Δ (10)

[0052] The discrete curve is divided into three parts: the oblique parabolic curve trajectory before the first inflection point; the first rebound trajectory from one frame after the first inflection point to the second inflection point; and the rolling trajectory from the last inflection point to the stop. The oblique parabolic curve trajectory and the first rebound trajectory are fitted using the cosine angle side-projection model or the spatial oblique parabolic curve model, and the two curves are combined to calculate the position C1 of the intersection projected on the 2D image; linear regression is used for the rolling trajectory, and the intersection position C2 is obtained by combining the regression line and the oblique parabolic curve.

[0053] In order to adapt to different ground materials, the second norm of the C1 and C2 coordinates is calculated, and the Euclidean distance value d c When it is less than the threshold, C1 is taken as the predicted landing point, otherwise C2 is taken as the landing point.

[0054] 1.4Pix-Real coordinate system mapping relationship modeling

[0055] In order to calculate the real space distance using the distance between two points in the image, it is necessary to model the relationship between the image coordinate system and the real coordinate system. The cosine angle side-projection model and the space oblique-projection curve model proposed in the present invention can calculate the quadratic term coefficient, linear term coefficient and constant term value of the oblique-projection curve on the plane where it is located, and can calculate the precise position p of the highest point in its oblique-projection trajectory. t And the exact location of the landing point p e .

[0056] Since the horizontal component of w along the flight direction is uniform linear motion, the relationship between displacement, velocity and time can be used to calculate p t The corresponding timestamp t t and p e The corresponding timestamp te , as shown in Figure 4(a). The highest point p in the real environment can be calculated by the gravity formula t The height value h (unit: meter) relative to the ground is as shown in formula (11), and g is the gravitational acceleration, which is 9.81m / s. The linear mapping relationship between the actual height and the pixel distance is obtained through this formula, so the actual distance d (unit: meter) between the landing point and the starting point can be calculated through formula (12), as shown in Figure 4(b).

[0057]

[0058]

[0059] 2. Cost Function

[0060] The data obtained by target detection is a two-dimensional point T on the plane. 2d , the i-th point is recorded as [x ti ,f(x ti )]. The fitting curve is a three-dimensional curve, and the points on the curve are denoted by P 3d , the i-th point is recorded as [x pi ,z pi ,f(x pi ,z pi )]. Since traditional least squares methods cannot directly calculate the average error, this embodiment proposes a projection cost function cost proj (T 2d ,P 3d ), as shown in formula (13). Since the plane where the image is located is parallel to the plane of xoy during modeling, let z pi =0P 3d The predicted key points are projected onto T 2d The plane where the mean square error is calculated.

[0061]

[0062] The Levenberg-Marquardt (LM) optimization algorithm is used to optimize the cosine angle side-throw model and the space oblique curve model proposed in the present invention, so as to fit the curve in space. For the cosine angle side-throw model, there are learnable parameters a, b, c, θ. After LM optimization, the projection cost function cost is obtained. proj The minimum parameters; for the spatial parabolic curve model, there are learnable parameters a, b, c, θ, p x , p y , p z , after LM optimization, the projection cost function cost is obtained proj The minimum parameter.

[0063] According to the general idea of ​​the invention described above, this embodiment provides a method for analyzing the trajectory of a shot put in three dimensions and predicting the performance based on a 2D image. Figure 5 As shown, the following steps may be included:

[0064] S1, target detection process: make the normal plane of the camera perpendicular to the ground, then obtain a two-dimensional image of the process from the start of throwing the ball to the stop of the shot put, detect the position of the ball through the target detection model, and synthesize a discrete trajectory with a timestamp;

[0065] S2, space curve model fitting process: establish a 3D space curve model, the 3D space curve model is a cosine angle side throw model or a space oblique throw curve model, calculate the difference between the 3D space curve and the points on the two-dimensional image plane through a projection cost function, and then use an optimization algorithm to correct the difference value, so that the 2D projection of the 3D space curve coincides with the position of the multi-frame shot put;

[0066] S3, landing point prediction process: determine the two inflection points in the discrete trajectory, so as to obtain the oblique curve trajectory, the first rebound trajectory and the rolling trajectory; adaptively calculate the landing point position by integrating the first rebound trajectory and the rolling trajectory;

[0067] S4. Actual distance calculation process: The relationship between the actual height of the highest point and the pixel is automatically calculated according to the gravity formula, thereby automatically calculating the actual physical distance to the ground.

[0068] In a possible implementation, the 3D space curve model is a cosine angle side-throwing model, and the formula is as follows:

[0069] y=a·(x·cos(θ)) 2 +b·x·cos(θ)+c

[0070] Among them, θ is the angle between the plane where the shot put flight trajectory is located and the camera normal plane, and a, b, and c are learnable parameters.

[0071] However, although the cosine angle side throw model can theoretically fit the actual shot put flight trajectory to a very high degree, it is necessary to ensure that the starting point is on the y-axis. Therefore, this solution has better fitting effect and speed, but it has extremely high requirements on the detection accuracy of the target detection algorithm and the video frame rate to ensure the detection of the shot put release frame.

[0072] In another possible implementation, in order to reduce the demand for hand frame detection, the 3D space curve model is a space oblique parabola curve model, and the formula is as follows:

[0073]

[0074] Where R is the spatial rotation matrix, p = [p x ,p y ,p z ] are learnable parameters, and (x, y, z) represent the coordinates of the point in the coordinate system.

[0075] In this embodiment, the formula of the projection cost function is:

[0076]

[0077] Among them, T 2d Represents a point in a discrete trajectory, and the i-th point is recorded as [x ti ,f(x ti )],P 3d Represents the point on the fitted three-dimensional curve, and the i-th point is recorded as [x pi ,z pi ,f(x pi ,z pi )], m represents the number of discrete trajectories.

[0078] The optimization algorithm used for the difference value calculated by the projection cost function is the Levenberg-Marquardt optimization algorithm.

[0079] In a possible implementation, determining two inflection points in the discrete trajectory specifically includes calculating the inflection points using discrete second-order differentials for three consecutive frames, and the i-th point x i The inflection point is when the following formula is satisfied:

[0080] f″(x i )>0+Δ

[0081] in,

[0082] Among them, f(x i ) is the height of the i-th point, Δ is the error threshold, which is 0.01 in this embodiment, and Δx is x i With x i-1 The difference between .

[0083] In this embodiment, the landing point prediction process is to adaptively calculate the landing point position according to the fusion method of the first rebound trajectory and the rolling trajectory, specifically including: fitting the oblique parabola curve trajectory and the first rebound trajectory using a 3D space curve model, and calculating the intersection of the two curves as the first intersection position C1; using linear regression on the rolling trajectory, and combining the regression line and the oblique parabola curve to obtain the second intersection position C2; calculating the second norm of the first intersection position C1 and the second intersection position C2, and when the Euclidean distance value d of the second norm is cWhen the distance is less than the threshold (the threshold is set to 40 pixels in this embodiment), the first intersection position C1 is taken as the landing position, otherwise the second intersection position C2 is taken as the landing position (usually the intersection of the two curves is more accurate, but the Euclidean distance value d c If it is greater than or equal to the threshold, it means that there is almost no rebound trajectory or the rebound trajectory is wrong. It is more accurate to get the intersection point by combining the regression line and the oblique throwing curve).

[0084] Preferably, the actual distance calculation process specifically includes: calculating the precise position p of the highest point in the oblique parabola trajectory through a 3D space curve model t And the exact location of the landing point p e , using the relationship between displacement, velocity and time to calculate p t The corresponding timestamp t t and p e The corresponding timestamp t e , calculate the highest point p in the real environment through the gravity formula t The height value h relative to the ground is used to obtain the linear mapping relationship between the actual height and the pixel distance. Then, the actual distance d between the landing point and the starting point is calculated based on the linear mapping relationship. The formula is as follows:

[0085]

[0086] Among them, w e is the pixel height of the highest point of the curve, w t is the pixel distance from the landing point to the starting point.

[0087] In order to verify the effectiveness of the 3D space curve model, the traditional quadratic curve model of formula (2) and the space oblique parabola curve model of formula (6) were fitted and verified using the same data, as shown in Figure 6. Figure 6(a) shows the fitting result using the traditional quadratic curve model. Although it is relatively accurate in predicting the landing point, the fitting effect is poor at the highest point and near the throwing position, which will cause errors in the modeling of the Pix-Real coordinate system mapping relationship and affect the final actual distance; Figure 6(b) shows the fitting result using the space oblique parabola curve model. It can be observed that its fitting effect is better than the traditional solution. The red curve in the figure is the oblique parabola curve, the yellow curve is the first rebound curve, the blue straight line is the rolling trajectory fitting line, the purple point is the final landing point predicted by the oblique parabola curve and the first rebound curve, the cyan point is the final landing point predicted by the rolling trajectory and the oblique parabola curve, and the arrow color and the point pointed by the arrow are dependent on d c The final landing point for judging the relationship with the threshold.

[0088] In order to ensure that the established model can be applied in complex terrain and complex actual environments, different types of terrain and different actual scenes are selected for effect verification.

[0089] As shown in Figure 7(a), it shows the prediction effect under hard ground, and Figure 7(b) shows the prediction effect under soft ground that cannot rebound. The verification of the environments in which Figures 7(a) and 7(b) are located proves the generalization ability of the model proposed in the present invention for different ground scenes. Figure 7(c) shows the prediction effect when the application scene is occluded. This figure proves that the model can still predict the landing position with high accuracy when the target detection effect is lost due to complex background or blocked by obstacles. Figure 7(d) shows the prediction effect when the rolling trajectory is offset due to potholes on the ground. This figure proves that the solution of the present invention can still maintain high-precision prediction in the face of complex ground environments.

[0090] 100 sets of data were actually sampled in different scenarios, and the verification results are shown in Table 1. The image resolution is 1920×1080, and the accuracy of the trained YOLOv8 shot put detection model is 99.31% and the recall is 98.62%. The recording frame rate uses a high refresh rate of 120FPS to ensure that the pixel error of the actual image is captured at the landing point, and the normal frame rate of the actual application scenario is simulated by sampling at 25FPS to ensure that the actual detection effect is met. The actual ranging error is the difference between the distance measured by the ruler after landing and the distance predicted by the algorithm. The final data is the average error after 100 sets of actual sampling.

[0091] Table 1 Shot Put Distance Measurement Evaluation Table

[0092] Average Error Traditional quadratic curve model Space oblique parabolic curve model Image pixel average error (unit: pixel) 32.51 0.82 Average error of real distance measurement (unit: mm) 81.28 2.05

[0093] The present invention establishes a 3D space curve model and proposes a projection cost function to calculate the difference between the space curve and the point on the two-dimensional image plane, so that the 2D projection of the 3D space curve coincides with the position of the shot put in multiple frames, thereby solving the problem that the existing method relying on 2D images cannot model the three-dimensional space; an auxiliary method based on the fusion of a rebound trajectory and a rolling trajectory is also proposed to model the position of the landing point on the curve and adaptively calculate the landing point position; by realizing space curve fitting and high-precision landing point distance calculation based on 2D images, the shot put tracking cost is greatly reduced.

[0094] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0096] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0097] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0098] Although the specific implementation modes of the present invention are described above, those skilled in the art should understand that the specific implementation modes described are only illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A 2D image-based three-dimensional trajectory analysis and performance prediction method for shot put, characterized in that: include: Target detection process: Make the normal plane of the camera perpendicular to the ground, then obtain a two-dimensional image of the process from the start of the throw to the stop of the shot put, detect the position of the ball through the target detection model, and synthesize a discrete trajectory with a timestamp; Spatial curve model fitting process: establishing a 3D spatial curve model, wherein the 3D spatial curve model is a cosine angle side-throw model or a spatial oblique-throw curve model, calculating the difference between the 3D spatial curve and the points on the two-dimensional image plane through a projection cost function, and then using an optimization algorithm to correct the difference, so that the 2D projection of the 3D spatial curve coincides with the position of the multi-frame shot put; Landing point prediction process: determine the two inflection points in the discrete trajectory to obtain the oblique curve trajectory, the first rebound trajectory and the rolling trajectory; adaptively calculate the landing point position by integrating the first rebound trajectory and the rolling trajectory; Actual distance calculation process: The relationship between the actual height of the highest point and the pixel is automatically calculated based on the gravity formula, thereby automatically calculating the actual physical distance to the ground.

2. The method according to claim 1, characterized in that: The formula of the cosine angle side throw model is as follows: y=a·(x·cos(θ)) 2 +b·x·cos(θ)+c Among them, θ is the angle between the plane where the shot put flight trajectory is located and the camera normal plane, and a, b, and c are learnable parameters.

3. The method according to claim 1, characterized in that: The formula of the spatial oblique parabolic curve model is as follows: Where R is the spatial rotation matrix, p = [p x ,p y ,p z ] are learnable parameters, and (x, y, z) represent the coordinates of the point in the coordinate system.

4. The method according to claim 1, characterized in that: The formula of the projection cost function is: Among them, T 2d Represents a point in a discrete trajectory, and the i-th point is recorded as [x ti ,f(x ti )],P 3d Represents the point on the fitted three-dimensional curve, and the i-th point is recorded as [x pi ,z pi ,f(x pi ,z pi )], m represents the number of discrete trajectories.

5. The method according to claim 1 or 4, characterized in that: The optimization algorithm used for the difference value calculated by the projection cost function is the Levenberg-Marquardt optimization algorithm.

6. The method according to claim 1, characterized in that: Determining the two inflection points in the discrete trajectory specifically includes calculating the inflection points using discrete second-order differentials for three consecutive frames, and the i-th point x i The inflection point is when the following formula is satisfied: f″(x i )>0+D in, Among them, f(x i ) is the height of the i-th point, Δ is the error threshold, and Δx is x i With x i-1 The difference between .

7. The method according to claim 1, characterized in that: In the landing point prediction process, the landing point position is adaptively calculated according to the fusion method of the first rebound trajectory and the rolling trajectory after landing, specifically including: using a 3D space curve model to fit the oblique parabola curve trajectory and the first rebound trajectory, and jointly calculating the intersection of the two curves as the first intersection position; using linear regression on the rolling trajectory, and jointly calculating the regression line and the oblique parabola curve to obtain the second intersection position; calculating the second norm of the first intersection position and the second intersection position, and taking the first intersection position as the landing point position when the Euclidean distance value of the second norm is less than a threshold, otherwise taking the second intersection position as the landing point position.

8. The method according to claim 1, characterized in that: The actual distance calculation process specifically includes: calculating the precise position p of the highest point in the oblique parabola trajectory through the 3D space curve model t And the exact location of the landing point p e , using the relationship between displacement, velocity and time to calculate p t The corresponding timestamp t t and p e The corresponding timestamp t e , calculate the highest point p in the real environment through the gravity formula t The height value h relative to the ground is used to obtain the linear mapping relationship between the actual height and the pixel distance. Then, the actual distance d between the landing point and the starting point is calculated based on the linear mapping relationship. The formula is as follows: Among them, w e is the pixel height of the highest point of the curve, w t is the pixel distance from the landing point to the starting point.

Images