A visual-driven golf trajectory reconstruction method and system for training scenarios

Abstract

The application discloses a kind of visual driving golf trajectory reconstruction method and system for training scene, method includes the continuous image sequence of the ball hitting area is collected, key time window is determined by combining audio or motion signal detection ball hitting event;Weak observation information allowing missing is extracted in window and quality score is calculated;Parameterized trajectory model containing physical priori constraint is established;Confidence weighted fusion is carried out based on observation score, optimal trajectory parameter is solved and flight trajectory is reconstructed;Output includes trajectory, ball course trend, landing area and reliability information Training-oriented feedback;Support round binding and group round statistical correction simultaneously, realize closed-loop training optimization.The application only needs ordinary visual equipment, trajectory can still be reconstructed stably under conditions such as shielding, blur, complex background, has the advantages such as low cost, strong robustness, output-oriented training, support continuous optimization, is suitable for golf practice court, indoor simulation and field training and other various scenes.

Description

Technical Field

[0001] This invention relates to the field of golf ball trajectory reconstruction technology, specifically to a vision-driven golf ball trajectory reconstruction method and system for training scenarios. Background Technology

[0002] In golf training, players need to accurately understand the ball's trajectory, landing point, flight trend, and its relationship with their swing after impact in order to conduct technical analysis and adjust their training. Currently, common methods for obtaining golf ball trajectories and calculating parameters mainly include: Millimeter-wave radar solution: It acquires the ball's flight data through continuous ranging and velocity measurement, which has high accuracy, but the equipment is expensive, the deployment is complex, and the power consumption is high, making it difficult to popularize in consumer-grade training scenarios; High-speed camera solution: This solution uses a high frame rate camera to image frame by frame and reconstruct the 3D trajectory. Although it is widely used in competition broadcasts or professional training, it also has the problems of high system cost and heavy data processing burden. Conventional vision solutions rely on conventional cameras for target detection and tracking. However, they typically assume that the sphere is clearly visible in consecutive frames. In real-world training environments, these solutions are prone to tracking failures due to motion blur, rapid movement away from the target, background interference, occlusion, or excessively small target size, resulting in insufficient robustness.

[0003] Furthermore, most existing methods lack systematic design for training scenarios: they fail to fully utilize the information-dense characteristics within a short time window after the shot, resulting in low computational efficiency and high latency; they lack physical prior constraints and uncertainty modeling, making it easy for reconstruction results to have multiple solutions or drift, and the output is a single numerical value, lacking credibility assessment; at the same time, existing solutions are mostly single shot analysis, not combined with training rounds, action records, and historical data, making it difficult to form a closed loop of "evaluation-correction-retraining". Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the background art mentioned above, and to provide a vision-driven golf ball trajectory reconstruction method and system for training scenarios that can stably and reliably reconstruct the golf ball trajectory in complex training environments without relying on expensive radar or high-speed cameras and using ordinary vision equipment.

[0005] This invention achieves the above objective through the following technical solution: a vision-driven method for reconstructing the trajectory of a golf ball in a training scenario, comprising the following steps: S1: Acquire a continuous image sequence containing the golf swing area; S2: Detect the moment t0 when the ball-hitting event occurs, based on the continuous image sequence or in combination with the synchronously acquired audio signal; S3: Taking the time t0 when the ball is struck as the starting point, determine a key time window [t0, t0+ΔT], where ΔT is a preset or adaptive duration parameter; S4: Within the key time window, weak observation information of the golf ball is extracted from at least some image frames, and an observation quality score is calculated for the weak observation information of each frame; wherein, the weak observation information includes at least one of the following: ball candidate region, probability heatmap, motion blur trailing direction or optical flow direction clue; it is permissible for some image frames to have missing weak observation information within the key time window; S5: Based on a parameterized trajectory model containing physical prior constraints, establish a trajectory function describing the three-dimensional spatial motion of a golf ball. The parameters of the trajectory function include at least the initial velocity, takeoff angle, and takeoff direction. S6: Based on the observation quality score, the weak observation information of each frame within the key time window is weighted with confidence, the optimal parameters of the trajectory function are solved by an optimization algorithm, and the flight trajectory of the golf ball within the key time window is reconstructed. S7: Output training-oriented feedback results, which include at least the reconstructed trajectory, ball path trend, landing area, and credibility information of the reliability of the representation results.

[0006] Furthermore, in step S2, detecting the occurrence time t0 of the ball-hitting event is specifically achieved through one or a combination of the following methods: Transient peak detection is performed on the synchronously acquired audio signal to determine the moment corresponding to the ball's impact sound; The continuous image sequence is analyzed to detect the moment of visual change when the cue stick comes into contact with the ball; Based on visual recognition of the swing motion, the timing of the impact phase is determined.

[0007] Further, in step S4, the observation quality score is calculated based on at least one of the following factors: Contrast between candidate targets and image background; The degree to which the morphological features of the candidate region match the sphere model; The observed location in this frame is consistent with the trajectory location predicted based on historical observations or physical models; The evaluation results of the degree of occlusion or motion blur in the image frame.

[0008] Furthermore, in step S5, the physical prior constraints include restrictions on the range of values ​​for the trajectory function parameters, which are set based on the physical common sense of golf.

[0009] Further, in step S6, the optimization algorithm is to minimize an objective function, which includes a deviation term between the weak observation information of each frame weighted by the observation quality score and the projection position predicted according to the trajectory function, and a regularization term based on the physical prior constraints.

[0010] Furthermore, in step S7, the output credibility information is a comprehensive credibility level, and the method further includes: When the overall confidence level is lower than the first threshold, a downgraded output strategy is triggered. The downgraded output strategy includes: only outputting the ball path trend and / or landing area, without outputting the specific trajectory function parameter values.

[0011] Furthermore, it also includes step S8: training round management and closed-loop correction; S8a: Assign a round identifier to each shot reconstruction process and associate and store the trajectory reconstruction results, parameters and confidence information of that round; S8b: Based on the statistical characteristics of multiple consecutive rounds, the prior distribution or systematic bias used in the trajectory reconstruction process of subsequent rounds is corrected and updated.

[0012] Furthermore, in step S8b, when a systematic offset is detected in the landing area of ​​multiple consecutive rounds, a system calibration prompt is triggered or parameter offset correction is automatically performed.

[0013] Furthermore, in step S5, the parameters of the parameterized trajectory model also include curvature parameters for describing the curvature trend of the ball path and / or drag parameters for approximating the effect of air resistance.

[0014] A vision-driven golf ball trajectory reconstruction system for training scenarios, used to implement a vision-driven golf ball trajectory reconstruction method for training scenarios, the system comprising: The image acquisition module is used to acquire continuous image sequences; The ball-hitting event detection module is used to detect the time t0 when a ball-hitting event occurs. The processing and control module is used to determine critical time windows and coordinate the work of various modules in the system; The weak observation extraction and scoring module is used to extract weak observation information and calculate quality scores within key time windows; The trajectory modeling and solving module is used to build a parameterized trajectory model with physical constraints, and solve for the optimal parameters based on confidence-weighted fusion to reconstruct the trajectory; The output and feedback module is used to output training-guided feedback results containing credibility information. It also includes a round management module for round binding and group round statistics correction.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. Requiring only a standard camera, eliminating the need for millimeter-wave radar or high-speed cameras, it significantly reduces hardware costs and deployment complexity, making it suitable for general training scenarios. Through a "weak observation extraction + missing data tolerance" mechanism, it allows reconstruction to continue even with motion blur, occlusion, background interference, or temporary target invisibility, avoiding the overall failure of traditional continuous tracking methods due to a single frame failure. Employing a ball-hitting event triggering and critical time window limiting strategy, it concentrates computation within a short period after the ball hits, significantly reducing the number of frames processed and computational load, thus improving system real-time performance.

[0016] 2. The output not only includes the reconstructed trajectory but also provides the ball's trajectory trend (straight / bend / bend), landing area (with an uncertainty range), and parameter confidence intervals, better meeting the actual needs of "trend judgment and area assessment" during training. By comprehensively evaluating the observation quality and fit consistency, a confidence level is output, and the output content is automatically downgraded when the confidence level is low, ensuring that the feedback information is always usable and not misleading. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the training scene trajectory reconstruction system structure in this invention; Figure 2 This is a flowchart of the method in this invention; Figure 3 This is a schematic diagram of the credibility classification and degradation output in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0020] Example 1: Implementation method in a portable training scenario at a practice field This embodiment is applicable to environments such as outdoor driving ranges, and is deployed using ordinary consumer-grade camera equipment (such as mobile phones or webcams with a frame rate of 30-60fps), aiming to provide a low-cost, portable golf ball trajectory reconstruction solution.

[0021] S1: Data Acquisition The camera is fixed to the side or rear of the hitting area to ensure coverage of the moment of impact and the initial stage of the ball's flight. A continuous sequence of images I(t) and an audio signal A(t) (e.g., via the device's microphone) are acquired simultaneously.

[0022] S2: Ball Hitting Event Detection The detection of the moment t0 of the ball-hitting event is achieved using a combination of audio and visual signals: • Perform short-time energy calculation and transient peak detection on audio signal A(t). When a transient peak exceeding the threshold is detected, it is initially marked as the moment of the ball being hit. • Simultaneously analyze image frames before and after this moment, and detect abrupt changes in motion in the contact area between the cue and the ball using inter-frame difference or optical flow methods for visual verification; • If the audio and visual detection results are consistent in time (e.g., the difference is no more than 3 frames), then t0 is confirmed as the moment the ball was struck.

[0023] S3: Determining the Key Time Window Starting from t0, the critical time window length is set to ΔT = 0.8s (which can be adaptively adjusted based on empirical ball speed values). Subsequent trajectory reconstruction is performed only within this window to avoid the computational burden and delays caused by processing the entire process.

[0024] S4: Weak Observation Extraction and Quality Scoring Within the key time window, weak observation information of the golf ball is extracted frame by frame: • For each frame of image, a lightweight object detection network (such as YOLOTiny or MobileNetSSD) is used to generate sphere candidate regions and corresponding confidence heatmaps; Simultaneously, the direction of motion blur and the direction of sparse optical flow in the image are calculated as auxiliary observation clues; • If no reliable candidate is detected in a frame (e.g., confidence level is below 0.3), then the observation in that frame is marked as missing (o_k=∅). For non-missing frames, calculate the observation quality score c_k, which comprehensively considers: the contrast between the candidate region and the background, the circularity of the region, the consistency with the predicted position of the previous frame, and the evaluation result of the blur degree of the image in this frame.

[0025] S5: Establish a parametric trajectory model Establish a three-dimensional trajectory function x(t;θ), where the parameter vector... include: Initial velocity v0 (typically 30-80 m / s); takeoff angle α (range 10°-45°); takeoff direction (Horizontal direction angle); curvature trend parameter s (negative for left curvature, positive for right curvature, range [-0.1, 0.1] rad / m). The 3D trajectory is projected into a 2D pixel trajectory using camera calibration parameters. .

[0026] S6: Confidence-weighted fusion solution Construct the objective function: Where o_k is the observation position of the k-th frame (not included if missing), R(θ) is a regularization term based on physical priors (such as penalizing parameters that exceed a reasonable range), and λ is the regularization coefficient.

[0027] The Levenberg-Marquardt algorithm was used to solve for the optimal parameters θ*, and the 3D trajectory and its projected pixel trajectory within the key time window were reconstructed.

[0028] S7: Output Training-Guided Feedback • Reconstructed trajectory: Outputs the projected trajectory on the image plane as a continuous curve; • Ball trajectory trend: Determine whether it is a "straight ball", "left curve ball", or "right curve ball" based on the curvature parameter s; • Landing point area: Output an ellipse in the ground projection area of ​​the image to represent the range of uncertainty of the landing point; • Credibility information: Based on the comprehensive observation completeness, mean quality score, and fitting residuals, the overall credibility Conf is calculated and divided into three levels: "High" (Conf≥0.7), "Medium" (0.4≤Conf<0.7), and "Low" (Conf<0.4). • Degraded output: If the confidence level is "low", only the ball trajectory trend and landing area will be output, without outputting specific parameter values.

[0029] S8: Round Binding and Group Round Correction • Assign a unique round identifier R_i to each shot and store all output information for that round accordingly; • After recording multiple rounds (e.g., 10 rounds), statistically analyze the characteristics such as the center position shift and initial velocity distribution of the landing area; If a systematic offset is detected in the landing area (e.g., the landing points of 5 consecutive balls are all off to the right), the user is prompted to recalibrate the camera position, or the bias correction term is automatically introduced into the subsequent solution to update the prior distribution P(θ).

[0030] Example 2: Implementation method in simulator / indoor training scenario This embodiment is applicable to indoor simulated training environments where the background is relatively stable but may be subject to interference such as reflections and specular reflections.

[0031] S2: Ball Hitting Event Detection In indoor environments, the sound of the ball being struck can be affected by echoes, so visual detection is the primary method of detection. • Detect the instant of contact between the clubhead and the ball using the high-speed differential method; • Optionally, the swing motion recognition model (based on posture estimation) can be combined to determine the impact phase.

[0032] S4: Weak Observation Extraction and Quality Scoring • Due to the stable background, background subtraction and morphological operations are used to extract motion candidate regions; • To address glare interference, a "color consistency" check has been added to the quality score: a real golf ball should appear white or highlighted in the image; if the color distribution of the candidate area is abnormal, its score will be reduced. • It is permissible for some frames to be missing due to excessive reflection, which the system can compensate for through subsequent fusion processes.

[0033] S6: Confidence-weighted fusion solution • In a windless indoor environment, the curvature parameter s can be ignored, and the model can be simplified as follows: ; • Improve the consistency of repeated training by gradually correcting camera calibration errors or system biases through round-by-round statistics (e.g., 20 consecutive balls).

[0034] Example 3: Implementation in Complex Playing Environments (such as Real Playing Fields) This embodiment is applicable to real court training, where the environment is complex, the background is cluttered, and the target may move away quickly or be temporarily obscured.

[0035] S3: Determining the Key Time Window Adaptive duration ΔT is used: • Initially set to 1.0s; If the percentage of valid observation frames within the window is less than 50%, the window will be automatically shortened to 0.6 seconds to reduce noise interference.

[0036] S4: Weak Observation Extraction and Quality Scoring • Lightweight optical flow method is used to extract motion-salient regions, and candidates are screened by combining color and shape priors; • For severely occluded or blurry frames, mark them as missing directly and do not force the extraction of observations; • The quality score focuses on evaluating “motion continuity” and “background complexity”.

[0037] S7: Output Training-Guided Feedback • In complex environments, the system outputs "medium" or "low" confidence results more frequently; • When the downgraded output strategy is activated, it still ensures the output of "ball trajectory trend (straight / left / right)" and "approximate landing area" to provide users with basic training reference.

[0038] System hardware and module composition (combined) Figure 1 ) The vision-driven golf ball trajectory reconstruction system includes: • Image acquisition module (1): Ordinary camera, frame rate 3060fps; • Ball-hitting event detection module (2): Implements multimodal (audio / visual) event detection; • Processing control module (3): Used to determine critical time windows and schedule the work of each module; • Weak observation extraction and scoring module (4): Extract candidates and calculate quality scores; • Trajectory modeling and solution module (5): Establish physical model and perform confidence-weighted fusion solution; • Output and Feedback Module (6): Generates credible training feedback; • Round Management Module (7): Implements round binding and group round correction.

[0039] Process description (in conjunction with) Figure 2 ) The entire method is executed according to steps S101 and S502, sequentially completing event detection, weak observation extraction within the window, physical modeling, confidence fusion, output feedback, and round correction, forming a closed loop of "detection, reconstruction, feedback, and correction".

[0040] Trustworthy grading and degradation output mechanism (combined) Figure 3 ) The system scores based on weak observation completeness and average quality. The overall confidence level (Conf) is calculated based on the fitting residuals and output in tiers according to preset thresholds. • Conf≥0.7: Output complete trajectory, parameter range, and landing point area; • 0.4≤Conf<0.7: Output trajectory trend, landing area, and coarse-grained parameter range; Conf<0.4: Only outputs the trend and landing area, and marks it as "low confidence", without outputting specific parameters.

[0041] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0042] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for visually driven golf ball trajectory reconstruction in a training scenario, characterized in that: Includes the following steps: S1: Acquire a continuous image sequence containing the golf swing area; S2: Detect the moment t0 when the ball is struck, based on the continuous image sequence or in combination with the synchronously acquired audio signal; S3: Taking the time t0 when the ball is struck as the starting point, determine a key time window [t0, t0+ΔT], where ΔT is a preset or adaptive duration parameter; S4: Within the key time window, weak observation information of the golf ball is extracted from at least some image frames, and an observation quality score is calculated for the weak observation information of each frame; wherein, the weak observation information includes at least one of the following: ball candidate region, probability heatmap, motion blur trailing direction or optical flow direction clue; it is permissible for some image frames to have missing weak observation information within the key time window; S5: Based on a parameterized trajectory model containing physical prior constraints, establish a trajectory function describing the three-dimensional spatial motion of a golf ball. The parameters of the trajectory function include at least the initial velocity, takeoff angle, and takeoff direction. S6: Based on the observation quality score, the weak observation information of each frame within the key time window is weighted with confidence, the optimal parameters of the trajectory function are solved by an optimization algorithm, and the flight trajectory of the golf ball within the key time window is reconstructed. S7: Output training-oriented feedback results, which include at least the reconstructed trajectory, ball path trend, landing area, and credibility information of the reliability of the representation results.

2. The method for visually driven golf ball trajectory reconstruction in a training scenario according to claim 1, characterized in that: In step S2, detecting the time t0 of the ball-hitting event is specifically achieved through one or a combination of the following methods: Transient peak detection is performed on the synchronously acquired audio signal to determine the moment corresponding to the ball's impact sound; The continuous image sequence is analyzed to detect the moment of visual change when the cue stick comes into contact with the ball; Based on visual recognition of the swing motion, the timing of the impact phase is determined.

3. The method for visually driven golf ball trajectory reconstruction in a training scenario according to claim 1, characterized in that: In step S4, the observation quality score is calculated based on at least one of the following factors: Contrast between candidate targets and image background; The degree to which the morphological features of the candidate region match the sphere model; The observed location in this frame is consistent with the trajectory location predicted based on historical observations or physical models; The evaluation results of the degree of occlusion or motion blur in the image frame.

4. The method for visually driven golf ball trajectory reconstruction in a training scenario according to claim 1, characterized in that: In step S5, the physical prior constraints include restrictions on the range of values ​​for the trajectory function parameters, which are set based on the physical common sense of golf.

5. The method for visually driven golf ball trajectory reconstruction in a training scenario according to claim 1, characterized in that: In step S6, the optimization algorithm is to minimize an objective function, which includes a deviation term between the weak observation information of each frame weighted by the observation quality score and the projection position predicted according to the trajectory function, and a regularization term based on the physical prior constraints.

6. The method for visually driven golf ball trajectory reconstruction in a training scenario according to claim 1, characterized in that: In step S7, the output credibility information is a comprehensive credibility level, and the method further includes: When the overall confidence level is lower than the first threshold, a downgraded output strategy is triggered. The downgraded output strategy includes: only outputting the ball path trend and / or landing area, without outputting the specific trajectory function parameter values.

7. The method for visually driven golf ball trajectory reconstruction in a training scenario according to claim 1, characterized in that: It also includes step S8: training round management and closed-loop correction; S8a: Assign a round identifier to each shot reconstruction process and associate and store the trajectory reconstruction results, parameters and confidence information of that round; S8b: Based on the statistical characteristics of multiple consecutive rounds, the prior distribution or systematic bias used in the trajectory reconstruction process of subsequent rounds is corrected and updated.

8. The method for visually driven golf ball trajectory reconstruction in a training scenario according to claim 7, characterized in that: In step S8b, when a systematic offset is detected in the landing area of ​​multiple consecutive rounds, a system calibration prompt is triggered or parameter offset correction is performed automatically.

9. The method for visually driven golf ball trajectory reconstruction in a training scenario according to claim 1, characterized in that: In step S5, the parameters of the parameterized trajectory model also include curvature parameters for describing the curvature trend of the ball path and / or drag parameters for approximating the effect of air resistance.

10. A vision-driven golf ball trajectory reconstruction system for training scenarios, characterized in that: The system for implementing the vision-driven golf ball trajectory reconstruction method for training scenarios as described in any one of claims 1-9, the system comprising: The image acquisition module is used to acquire continuous image sequences; The ball-hitting event detection module is used to detect the time t0 when a ball-hitting event occurs. The processing and control module is used to determine critical time windows and coordinate the work of various modules in the system; The weak observation extraction and scoring module is used to extract weak observation information and calculate quality scores within key time windows; The trajectory modeling and solving module is used to build a parameterized trajectory model with physical constraints, and solve for the optimal parameters based on confidence-weighted fusion to reconstruct the trajectory; The output and feedback module is used to output training-guided feedback results containing credibility information. It also includes a round management module for round binding and group round statistics correction.

Images