Athletic event judging intelligent AI construction method and system

Abstract

The application discloses a kind of for track and field sports event judge intelligent AI construction method and system, specifically related to image recognition and behavior determination technical field, for solving the problem of multiple track and field event judge misjudgment dispute;The application carries out the time synchronization of starting area image sequence and starting pistol signal by image acquisition device and sound perception equipment, constructs standardization starting response sequence, extracts lower limb key point trajectory from the sequence, analyzes starting action starting time by combining displacement and speed change, judges whether there is sprinting behavior, and the foot point extracted is mapped to unified reference plane, judges shielding level, constructs penalty confidence in conjunction with image definition, penalty confidence and penalty result jointly form judge interactive frame set, present in auxiliary determination interface, for judge to review and confirm, and then judge modification behavior is used as difference sample and is fed back, is optimized by incremental training mechanism, improve the controllability and continuous optimization ability of penalty process.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition and behavior determination, and more specifically, to a method and system for constructing an intelligent AI for refereeing track and field events. Background Art

[0002] Track and field events place extremely high demands on referee accuracy, response speed, and consistency of penalties, especially in events such as sprinting, relays, and long jump. Key penalty nodes such as false start determination, relay zone handovers, and the legality of landing points directly affect athletes' performance and competition fairness. The existing manual refereeing mechanism is highly dependent on the referee's subjective experience and manual replay and review of slow-motion videos. Although some events have introduced high-frame-rate cameras and laser trigger systems to assist in penalty judging, these systems are mostly based on single-modal video information and lack structured data processing and automatic behavior recognition capabilities. The penalty results still require manual interpretation and confirmation, and there are problems such as response delays, incomplete evidence chains, and low dispute resolution efficiency.

[0003] Deficiencies in existing technologies: In the construction of AI referees for track and field, especially in typical application scenarios such as false start determination, due to the boundary jitter of the image frame time granularity and the recognition accuracy of key points on the human body, coupled with the nonlinear gradient and posture reconstruction characteristics of the athletes' starting behavior, it is difficult to establish a robust "gunshot to start" causal sequence recognition path. In addition, in relay events, the baton handover area is action-intensive and there is severe double occlusion. The existing recognition network cannot accurately distinguish between the handover hand shape and the boundary of the spatial relay zone, resulting in ambiguous handover penalties and reducing the trust in the AI ​​system's penalties. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and system for constructing an intelligent AI for refereeing track and field events, so as to solve the problem of many disputes caused by misjudgments of referees in track and field events in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for constructing an intelligent AI for refereeing track and field events, comprising the following steps:

[0007] Real-time acquisition and time synchronization calibration of the sprint starting area image sequence and the starting gun signal to generate a standardized starting response sequence;

[0008] Extract the key points of the athlete's lower limb trajectory and force trend characteristics from the starting response sequence, calculate the starting instant of the movement to identify the starting signal;

[0009] Determine whether a false start occurs based on the time difference between the start signal and the gunshot signal, and generate a candidate penalty result;

[0010] Collect image sequences from two preset angles at the starting area and perform keypoint spatial reconstruction. The confidence level of the penalty is assessed based on foot occlusion and image change areas.

[0011] The penalty candidate results and penalty confidence levels are jointly mapped to generate a referee interaction frame set, and an auxiliary judgment interface is constructed;

[0012] Collect manual review results and optimize key point detection weights and start timing judgment parameters based on the difference samples between the candidate penalty results and the review results.

[0013] In a preferred embodiment, the image sequence of the sprint starting area is collected in real time and time-synchronized with the starting gun signal to generate a standardized starting response sequence. The specific process is as follows:

[0014] A sequence of image frames of the starting area is collected by a preset image acquisition device at a fixed sampling period, wherein the image frames cover the area between the starting blocks and the athlete's feet;

[0015] The sound sensing module collects the starting gunshot signal in real time, and uses the first valid rising edge of the gunshot signal as the time synchronization benchmark;

[0016] Using the timestamp of the gunshot signal as the reference point, the image frame sequence is uniformly timed, and image segments before and after the start are intercepted according to a fixed time window to construct a standardized start response sequence.

[0017] In a preferred embodiment, the key point trajectories and force generation trend characteristics of the athlete's lower limbs are extracted from the starting response sequence, and the starting instant of the movement is calculated to identify the starting signal. The specific process is as follows:

[0018] Extracting the position information of the key points of the lower limbs in the image frames from the standardized start response sequence, the key points of the lower limbs include the toes, ankles, and knees;

[0019] Calculate the spatial displacement and velocity changes of the lower limb key points between adjacent image frames;

[0020] Identify the sudden change point of foot acceleration as the starting instant of the running action and mark the corresponding image frame number.

[0021] In a preferred embodiment, the time difference between the start signal and the gunshot signal is used to determine whether a false start event has occurred and generate a candidate penalty result. The specific process is as follows:

[0022] Calculate the time difference between the timestamp of the start action and the timestamp of the gunshot;

[0023] The obtained time difference is compared with the set false start penalty threshold. If the time difference is less than the penalty threshold, it is identified as a false start event;

[0024] Output the candidate results of false start penalties and associate the starting frame of the starting action and its key point change information.

[0025] In a preferred embodiment, multi-view images of the starting area are collected and key point space reconstruction is performed, and the penalty confidence is evaluated based on foot occlusion and image motion areas. The specific process is as follows:

[0026] capturing image frames generated at the same time at a first camera position and a second camera position, respectively, and extracting coordinate information of key points of the foot in the image frames;

[0027] Perform spatial mapping on the key points of the foot extracted from the first camera position and the second camera position, and calculate the spatial overlap to determine the occlusion level;

[0028] Combined with the areas where contour clarity changes in the image frame, the penalty confidence results are constructed based on overlap and clarity to assist in penalty decision-making.

[0029] In a preferred embodiment, spatial mapping is performed on the key points of the foot extracted from the first camera position and the second camera position, and the spatial overlap is calculated to determine the occlusion level, including:

[0030] The image coordinates of the key points of the foot extracted from the first camera position are converted into unified reference plane coordinates through perspective projection;

[0031] Mapping the image coordinates of the foot key points extracted from the second camera position to a unified reference plane through the same projection method;

[0032] The average Euclidean distance of the two groups of foot key points in the reference plane is calculated, and the area is divided into clear area, slightly occluded area and severely occluded area according to the relationship between the average Euclidean distance and the preset occlusion distance threshold.

[0033] In a preferred embodiment, the image is divided into a clear area, a slightly blocked area, and a severely blocked area according to the relationship between the average Euclidean distance and a preset occlusion distance threshold, including:

[0034] The preset occlusion distance threshold includes a first occlusion distance threshold and a second occlusion distance threshold, and the first occlusion distance threshold is smaller than the second occlusion distance threshold;

[0035] When the average Euclidean distance is less than or equal to the first occlusion distance threshold, it is determined to be a clear area;

[0036] When the average value of the Euclidean distance is greater than the first occlusion distance threshold and less than or equal to the second occlusion distance threshold, it is determined to be a slightly occluded area;

[0037] When the average Euclidean distance is greater than the second occlusion distance threshold, it is determined to be a severe occlusion area.

[0038] In a preferred embodiment, the penalty candidate results and the penalty confidence are jointly mapped to generate a referee interaction frame set, and an auxiliary decision interface is constructed. The specific process is as follows:

[0039] Extract key image frames adjacent to the starting action start frame and the gunshot frame from the starting response sequence;

[0040] Combine the key frame sequence with the penalty confidence result to form a referee interaction frame set;

[0041] The auxiliary judgment interface displays frame images, key point trajectories and penalty confidence information for the referee to confirm, modify or reject.

[0042] In a preferred embodiment, manual review results are collected, and the key point detection weights and start timing judgment parameters are optimized based on the difference samples between the candidate penalty results and the review results. The specific steps are as follows:

[0043] Record the referee's operation information and corresponding image frame number for correcting the AI ​​output penalty results in the auxiliary judgment interface;

[0044] The frame data corrected by the referee is constructed into a difference sample set, which is used as training input for updating the key point detection weights and adjusting the start timing judgment parameters;

[0045] The incremental training method is used to update the parameters in the two processing flows and conduct learning adaptation of referee correction cases.

[0046] A system for constructing an intelligent AI for refereeing track and field events, for implementing the aforementioned method for constructing an intelligent AI for refereeing track and field events, comprising:

[0047] The starting data acquisition module is used to collect and synchronize the image sequence of the sprint starting area with the starting gun signal in real time to generate a standardized starting response sequence;

[0048] The starting feature module is used to extract the key point trajectory and force trend characteristics of the athlete's lower limbs from the starting response sequence, calculate the starting instant of the movement to identify the starting signal;

[0049] A penalty generation module is used to determine whether a false start event occurs based on the time difference between the start signal and the gunshot signal, and generate a candidate penalty result;

[0050] The penalty determination module is used to collect image sequences from two preset angles of the starting area and perform key point spatial reconstruction. It also evaluates the confidence level of the penalty based on the foot occlusion and the image change area.

[0051] The interface display module is used to jointly map the penalty candidate results and the penalty confidence to generate a referee interaction frame set and build an auxiliary judgment interface;

[0052] The penalty review module is used to collect manual review results and optimize the key point detection weights and start timing judgment parameters based on the difference samples between the candidate penalty results and the review results.

[0053] The technical effects and advantages of the present invention are as follows:

[0054] The present invention synchronizes the time of the starting area image sequence with the starting gun signal to construct a standardized starting response sequence. This ensures the consistency of the image data with the penalty reference signal, improves the accuracy of subsequent analysis, and combines key point recognition with acceleration mutation detection to effectively capture the true starting moment of the starting action, significantly enhancing the sensitivity of false start detection.

[0055] By collecting multi-angle images and performing spatial reconstruction of key points, the system further addresses the interference caused by foot occlusion on referee judgments. A confidence scoring mechanism is constructed based on image clarity and key point overlap, enhancing the credibility and interpretability of the calls. Furthermore, the system integrates a referee-assisted judgment interface to enable visual interaction and manual confirmation of AI-generated decision results, improving the collaborative efficiency between the system and referees. Finally, by collecting referee correction information and constructing a differential sample set, the system optimizes detection parameters based on an incremental training mechanism, enabling AI to continuously learn and adapt to actual case studies. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of a method for constructing an intelligent AI for refereeing track and field events according to the present invention.

[0057] Figure 2 This is a structural diagram of an intelligent AI construction system for refereeing track and field events according to the present invention. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0059] Example 1: Figure 1 As shown, a method for constructing an intelligent AI for refereeing track and field events includes the following steps:

[0060] Real-time acquisition and time synchronization calibration of the sprint starting area image sequence and the starting gun signal to generate a standardized starting response sequence;

[0061] Extract the key points of the athlete's lower limb trajectory and force trend characteristics from the starting response sequence, calculate the starting instant of the movement to identify the starting signal;

[0062] Determine whether a false start occurs based on the time difference between the start signal and the gunshot signal, and generate a candidate penalty result;

[0063] Collect image sequences from two preset angles at the starting area and perform keypoint spatial reconstruction. The confidence level of the penalty is assessed based on foot occlusion and image change areas.

[0064] The penalty candidate results and penalty confidence levels are jointly mapped to generate a referee interaction frame set, and an auxiliary judgment interface is constructed;

[0065] Collect manual review results and optimize key point detection weights and start timing judgment parameters based on the difference samples between the candidate penalty results and the review results.

[0066] To accurately extract the key timing behaviors during the sprint start, it is necessary to first complete the synchronous calibration of image acquisition and starting signal to construct a standardized start response sequence;

[0067] Step 1: Real-time acquisition and time synchronization calibration of the sprint starting area image sequence and the starting gun signal to generate a standardized starting response sequence, including the following:

[0068] Image acquisition devices are selected and placed in front of and on both sides of the starting area to ensure that their shooting images can cover the entire area including the starting blocks, athletes' feet and calves. A high-speed camera with a frame rate of not less than 120 frames per second and a resolution of not less than 1920×1080 is used to meet the requirements of clear image recording of fast movements. The image acquisition device operates continuously with a fixed sampling period (for example, 1 frame is acquired every 8.33 milliseconds), continuously acquiring a sequence of image frames several seconds before and after the start, and attaching acquisition timestamp information accurate to the millisecond level to each frame of the image, forming a time-stamped original image stream.

[0069] A sound sensing device is deployed and positioned no more than 1 meter from the starter. This module, with an audio sampling rate of at least 44.1kHz, is used to capture the acoustic changes of the starting gunshot signal in real time. Signal processing methods are used to extract the first significant rising edge of the gunshot—the inflection point where the sound pressure signal experiences a sudden increase in amplitude from the static background noise—as a reference point for determining the start of the gunshot. The timestamp of this inflection point is set as the start time, and all subsequent analyses are calibrated based on this time.

[0070] With the start reference time as 0 seconds, the timestamps of all original image frames were shifted and converted to a relative time coordinate system. For example, if the original timestamp of an image frame was t = 3.524 seconds and the start reference time was 3.500 seconds, the image frame was recalibrated to +0.024 seconds. Based on this, image frames were captured within the time window from 0.5 seconds before to 2.5 seconds after the start. That is, at a setting of 120 frames per second, 60 frames before the start and 300 frames after the start were selected to form the start segment, a total of 360 frames, forming a complete standardized start response sequence.

[0071] Each frame in this response sequence contains complete image information and relative time tags, which will be used to extract the athlete's foot key point trajectory, muscle force trends, and displacement mutation characteristics. This data is the core data source for identifying "starting moment," "false start," and "effective start" in the subsequent penalty algorithm. For example, in an actual competition, the starting gun was calibrated at 3.000 seconds, and the image sequence was collected from 2.500 seconds to 5.500 seconds, with a total of 360 frames. Frame 87 corresponds to a time of +0.225 seconds. After key point analysis, it was determined to be the starting frame where the foot first left the starting blocks.

[0072] By aligning image acquisition with the gunshot signal, a standard timeline for the starting event is established, effectively eliminating the time deviation between the image frame and the starting signal at the acquisition end, ensuring the temporal consistency and comparability of subsequent starting behavior recognition; the constructed standardized starting response sequence provides a high-precision, strong temporal raw data foundation for full-process AI modeling, improving the accuracy and stability of recognition.

[0073] Step 2: Extract the key points of the athlete's lower limb trajectory and force trend characteristics from the starting response sequence, calculate the starting instant of the movement to identify the starting signal, including the following:

[0074] For each image frame in the standardized start response sequence constructed in Step 1, keypoint detection is performed sequentially to extract the positional coordinates of key lower limb components, such as the athlete's toes, ankles, and knees. Keypoint detection utilizes a publicly trained posture estimation model, which, provided the image resolution is at least 1920×1080, can reliably extract the pixel coordinates of key points. To ensure the accuracy of the extracted region, the region of interest (ROI) is limited to the area below the starting blocks and the athlete's calf in the image, thus avoiding non-target interference.

[0075] Next, we perform spatial displacement calculations on the same key point in consecutive image frames, recording the relative displacement between each frame (e.g., in pixels). This displacement, combined with the time interval, is converted into actual velocity change. For example, at a frame rate of 120 frames per second, the time interval between adjacent frames is approximately 8.33 milliseconds. If the ankle point shifts 12 pixels horizontally between two frames, then, considering the physical pixel size of the image acquisition device (e.g., each pixel represents 1.2 mm), we can deduce that the movement velocity is approximately 1.73 m / s. This process continues in this manner, forming a complete time-position-velocity trajectory sequence for each key point.

[0076] Next, key point accelerations are calculated based on the speed trend. Approximate acceleration values ​​are obtained by dividing the inter-frame speed difference by the time interval. To eliminate jitter interference caused by minor errors, a threshold for acceleration significance is set. When the acceleration of a key point jumps in amplitude (for example, from 0.2 to over 1.5) in consecutive frames, and the magnitude of the change exceeds a preset mutation amplitude (for example, set to 1.0 meter), that moment is identified as the athlete's starting point. The mutation point at the toes or ankle is generally used as the basis for this determination.

[0077] After detecting the acceleration mutation point, the corresponding image frame number is recorded and marked as the action start frame, which serves as the starting point for subsequent false start timing determination. For example, during a certain start, the speed of the ankle key point in the 92nd frame image suddenly increases from 0.3 m / s to 2.0 m / s, the acceleration exceeds 2.0, and the toe point also shows a synchronous mutation. At this time, the system records the 92nd frame as the instant of the start action;

[0078] It should be noted that the identification of starting behavior does not rely on the referee's naked eye judgment. Instead, it realizes standardized starting signal recognition through continuous spatiotemporal trajectory and dynamic threshold judgment mechanism. In terms of parameter setting, the acceleration threshold, speed conversion ratio and image frame interval are all based on the hardware configuration of the image acquisition device and the actual application environment. They can also be adjusted according to the starting characteristics of athletes of different age groups, thereby enhancing the adaptability and engineering deployability of the method.

[0079] This step achieves accurate recognition of the starting action by capturing the displacement and speed mutations of key points such as the toes, ankles, and knee joints. This avoids the traditional coarse-grained judgment based on the overall displacement of the body and enhances the system's sensitivity to tiny starting movements. The accurate recognition of the starting signal provides a highly reliable event trigger point for false start penalties.

[0080] Step 3: Determine whether a false start has occurred based on the time difference between the start signal and the gunshot signal, and generate a candidate penalty result, including the following:

[0081] The image frame number corresponding to the start of the running movement identified in step 2 is used, and combined with the frame rate parameter of the image acquisition device (e.g., 120 frames per second), the timestamp of the image frame is calculated. For example, if the start movement is identified at frame 94, the corresponding timestamp is 94 ÷ 120 = 0.783 seconds. This timestamp is the start time of the running movement.

[0082] Next, obtain the timestamp of the starting gun signal recorded by the sound sensing module in step 1. This timestamp is recorded by the system immediately when the first valid rising edge of the sound wave signal is detected. For example, the timestamp of the signal recording is 0.800 seconds.

[0083] Next, the time difference between the start of the start and the gunshot is calculated: the start time is subtracted from the gunshot time. In the above example, the time difference is 0.800–0.783=0.017 seconds. To ensure the system's versatility across different types of events, a false start penalty threshold can be preset in the system. This is generally set according to IAAF standards, with a recommended threshold of 0.100 seconds (100 milliseconds). This threshold indicates that if an athlete starts within 100 milliseconds of hearing the gunshot, it is generally considered an impossible time for the human nervous system to complete, and therefore will be considered a false start.

[0084] The system compares the time difference (0.017 seconds) with the preset threshold (0.100 seconds). If the time difference is less than the threshold, the system determines that a false start has occurred. In this example, 0.017 seconds is less than 0.100 seconds, so it is identified as a false start.

[0085] After identifying a false start, candidate false start penalty results are automatically generated and bound to the contextual information of the start event, including but not limited to: the starting action start image frame number, the position coordinates of the athlete's foot key points in the frame image, the key point speed value, the take-off direction and other data, to form a complete candidate penalty data structure. This candidate penalty data structure can be used for subsequent referee auxiliary interface display and manual review.

[0086] In addition, it also supports personalized adjustment of relevant threshold settings. For example, in youth competitions, the threshold can be set to 0.120 seconds based on actual reaction ability, and in high-level competitive competitions it is strictly set to below 0.100 seconds, thereby improving the accuracy and adaptability of the penalty.

[0087] This step realizes the automatic association between starting behavior and rule penalties. The system can calculate the starting time difference with millisecond-level accuracy, and automatically identify whether there is a violation of the rules based on this, and form a candidate penalty result; it realizes an intelligent closed loop from sports behavior to referee basis, greatly improving the penalty response speed and fairness.

[0088] Step 4: Collect image sequences from two preset angles in the starting area and perform keypoint spatial reconstruction. Based on the foot occlusion and image change areas, evaluate the penalty confidence, including the following:

[0089] Based on multi-view image acquisition and key point space reconstruction methods, a joint evaluation model for occlusion level judgment and contour clarity is constructed;

[0090] A primary camera position and a secondary camera position are set up in the starting area of ​​the competition venue. These two camera positions are positioned opposite each other on either side of the starting line, with varying viewing angles (e.g., ±30 degrees). Each camera captures a sequence of image frames in real time at a fixed frame rate (e.g., 120 frames per second) and selects the corresponding image frame at the same time (synchronized with the starting gun signal).

[0091] Subsequently, the system extracts the coordinates of the key points of the athlete's feet in the first-perspective and second-perspective image frames respectively. The extracted key points include the toes, ankles and heels. The extraction method is based on the trained posture recognition model (such as the skeleton point detection model) and is marked in the form of image pixel coordinates.

[0092] Next, the coordinates of the key points from the two camera perspectives are spatially mapped. This mapping uses a three-dimensional geometric model after the camera's internal and external parameters have been calibrated. The three-dimensional coordinates of the key points in physical space are reconstructed based on the principle of dual-view triangulation. The spatial overlap is calculated based on the area of ​​the overlapping area of ​​the key points from the two perspectives and the degree of projection overlap. The spatial overlap is expressed as a value between 0 and 1, which is used to quantify whether the same foot key point is fully displayed from different perspectives. For example, when the distance between the ankle points reconstructed from the two perspectives in three-dimensional space is less than 5 mm, and the overlapping area of ​​the projection contours exceeds 80%, the spatial overlap can be assigned a value of 0.9, indicating that the point is not severely occluded.

[0093] After determining the spatial overlap, the system further analyzes the changes in the outline clarity of the foot region within the image frame. Outline clarity is assessed by extracting the foot outline using an edge operator and calculating edge strength metrics, such as the grayscale gradient change rate or edge density. If the degree of edge blur exceeds a preset threshold (e.g., an edge gradient value below 25), the image is considered to have significant motion blur or occlusion.

[0094] Finally, spatial overlap and outline clarity are mapped together to a penalty confidence score, which is a continuous value between 0 and 1. For example, when spatial overlap is 0.9 and outline clarity is 0.85, the penalty confidence score is 0.88 according to the set weighting rules. If the spatial overlap is only 0.4 and the clarity score is 0.6, the comprehensive score drops below 0.5, indicating that the current frame image information is insufficient to independently support the penalty conclusion.

[0095] Perform spatial mapping on the key points of the foot extracted from the first and second camera positions, and calculate the spatial overlap to determine the occlusion level. Specifically, the following steps are performed:

[0096] A first camera position and a second camera position are set up in the starting line area of ​​the track, with a certain angle between the two positions (such as a horizontal angle of 30 degrees) to collect images of the athlete's feet at the same time from different directions. The image plane coordinates of the key points of the feet in the two sets of images are extracted separately through a posture recognition algorithm, including the pixel positions of key structural points such as the toes, ankles and heels. The two sets of image coordinates are projected onto a unified reference plane using a unified perspective projection transformation model. This reference plane is the plane of the track, which is set as a two-dimensional coordinate system in the system. The origin is the midpoint of the starting line, the horizontal axis is consistent with the direction of the runway, and the vertical axis is perpendicular to the direction of the runway. The perspective projection model is based on the pre-completed camera intrinsic parameters (focal length, principal point position) and extrinsic parameters (camera spatial position and posture) calibration results, combined with the pinhole projection principle, to restore the image pixel coordinates to the corresponding physical space position, and then convert them into reference plane coordinates. The same projection model is used for both camera positions to ensure mapping consistency and avoid spatial errors caused by perspective differences.

[0097] On a unified reference plane, the spatial coordinates of each key point in the first camera position and the second camera position are obtained respectively, the Euclidean distance of the same key point in the two perspectives is calculated, and the average Euclidean distance of all key points is obtained. The Euclidean distance is used to measure the consistency of the reconstruction results of the same target point from two perspectives, reflecting the degree of occlusion or visual distortion. For example, if the ankle point is mapped to the coordinates (120mm, 50mm) in the first perspective and (122mm, 53mm) in the second perspective, the distance between the two is approximately 3.6mm, indicating that the difference in perspective is small and there is no obvious occlusion.

[0098] The partitioning is determined based on the obtained average Euclidean distance and the preset occlusion distance threshold. The preset occlusion distance threshold is obtained by statistics of the system training samples, for example, it is set to two dividing values ​​of 10mm and 30mm. When the average distance is less than 10mm, it is considered that the two perspective mappings are highly consistent and divided into a clear area; when the distance is between 10mm and 30mm, it means that there is slight occlusion or image difference in some areas, and it is divided into a slightly occluded area; if the distance exceeds 30mm, it means that the reconstruction results under the two perspectives are quite different, there is significant occlusion or structural misjudgment, and it is divided into a severely occluded area. For example, if the current average Euclidean distance is 8.2mm, it is marked as a clear area; if it is 25.5mm, it is marked as a slightly occluded area; if it is 32.8mm, it is marked as a severely occluded area.

[0099] Based on the actual site layout and equipment resolution, a two-level occlusion judgment threshold is set in a unified reference plane. That is, the occlusion distance threshold includes a first occlusion distance threshold and a second occlusion distance threshold. The first occlusion distance threshold is used to determine the unobstructed state, and the second occlusion distance threshold is used to distinguish between slight and severe occlusion states. The first occlusion distance threshold must be less than the second occlusion distance threshold. For example, in the starting area of ​​a standard 100-meter runway, referring to the view reconstruction error control range, the system sets the first occlusion distance threshold to 15 mm and the second occlusion distance threshold to 30 mm.

[0100] The foot key points extracted from the first and second camera positions are spatially reconstructed and mapped to a unified reference plane. The average Euclidean distance between all key point pairs is calculated as a quantitative indicator to determine the spatial consistency of the current frame.

[0101] If the average value of the Euclidean distance is less than or equal to the first occlusion distance threshold, it means that the foot key points of the two perspectives almost completely overlap, indicating that the field of view is unobstructed and the image quality is reliable, and the image frame is marked as a clear area;

[0102] If the average Euclidean distance is greater than the first threshold but less than or equal to the second occlusion distance threshold, it indicates that there is a certain reconstruction deviation or partial occlusion between the two perspectives, and it is marked as a slightly occluded area;

[0103] If the average Euclidean distance exceeds the second occlusion distance threshold, it means that the key point matching error is significant, there may be obvious occlusion or image anomaly, and the current image frame is marked as a severe occlusion area.

[0104] Through plane mapping and distance analysis, the occlusion level can be quickly judged, which can be applied to the evaluation of image validity and filtering of abnormal situations in the real-time penalty system, ensuring the reliability of subsequent action recognition and penalty results.

[0105] The confidence level of penalty decisions is constructed based on overlap and clarity, including:

[0106] Based on the calculated spatial overlap of the foot key points, it is defined as the normalized reflection of the average Euclidean distance of multiple key points after they are mapped to a unified reference plane under the first and second camera positions. Specifically, the average value of the Euclidean distance is classified into the interval [0, 1] to represent the degree of overlap consistency. For example, if the preset maximum tolerable error distance is 50mm, if the current average distance is 10mm, then the overlap index is calculated as 1-10 / 50=0.8, indicating good perspective reconstruction consistency;

[0107] The image clarity of the foot area in the current image frame is calculated. Image clarity is used to reflect the clarity of the image edge texture. The intensity of the local gradient change of the image is usually used as a quantitative indicator. Specifically, the grayscale image in the area surrounded by the foot is extracted from the image, and the gradient map of the area is calculated using the Sobel edge operator. The percentage of pixels with edge gradient values ​​greater than the set threshold is counted as the clarity score. For example, when the edge threshold is set to 40, if there are 800 pixels in the foot area, of which 480 pixels have edge gradients exceeding the threshold, the clarity index is 480 / 800=0.6. The index ranges from 0 to 1. The larger the value, the clearer the image edge and the smaller the motion blur.

[0108] The spatial overlap and image clarity are weighted and fused using an equal-weight strategy to generate the penalty confidence score for the current image frame. This score is also normalized to the range of 0 to 1. For example, if the overlap index of the current frame is 0.8 and the clarity index is 0.6, the final penalty confidence score is (0.8 + 0.6) / 2 = 0.7.

[0109] Based on the set confidence level threshold, the confidence level of the penalty is divided into the following levels: when the score is above 0.75, it is marked as "high confidence" penalty basis; between 0.5 and 0.75, it is marked as "medium confidence"; and when it is below 0.5, it is marked as "low confidence" and prompts manual review. For example, during a penalty process, if the starting action starts before the starter gun is fired and the confidence score of the corresponding image frame is 0.82, the system can directly output a high-confidence false start penalty result; if the confidence score is only 0.43, only a false start candidate result is generated and marked as "low confidence", requiring manual confirmation.

[0110] This penalty confidence score is internally bound to the false start candidate penalty structure, assisting referees in determining whether the candidate result requires further manual verification or whether the final decision should be withheld during the review phase. This demonstrates the system's adaptive judgment capabilities for occlusion and image quality issues. Scoring model parameters (such as overlap threshold, clarity scoring criteria, and weighting coefficients) are determined through training with labeled data from historical competition samples to ensure stable and repeatable judgments.

[0111] By introducing a multi-angle image reconstruction and occlusion analysis mechanism, the problem of single-view recognition being easily obscured and having a high probability of misjudgment is effectively solved; by jointly constructing a penalty confidence index using spatial overlap and contour clarity, it can not only output the penalty conclusion but also measure its reliability, thereby enhancing the interpretability of the results and the referee's trust in the results.

[0112] Conduct manual review and confirmation of false start penalty results, improve the transparency and reliability of the system's penalty decisions, map automatically identified penalty candidate results with their corresponding penalty confidence levels, generate a visual referee interaction frame set, and build an image-data collaborative auxiliary decision interface for referees to confirm, modify, or reject decisions.

[0113] Step 5: Combine the candidate penalty results and the penalty confidence to generate a referee interaction frame set and build an auxiliary decision interface, including the following content:

[0114] After identifying the false start frame and calibrating the gunshot frame, the system extracts multiple image frames adjacent to the false start frame and the gunshot frame from the standardized start response sequence as key visual references. Typically, the system selects three frames forward and three frames backward for supplementary framing. For example, with a sampling time of 2 milliseconds per frame, the system extracts seven frames from 6 milliseconds before the gunshot to 6 milliseconds after the start of the action, forming a continuous frame window. For example, if the gunshot occurs at frame 100 and the corresponding action starts at frame 103, the system will extract images from frames 97 to 103 as the key frame sequence.

[0115] Next, the key frame sequence is bound to the corresponding frame's confidence score, forming a referee interaction frame set. The confidence score is derived from the aforementioned occlusion level and image clarity, typically ranging from 0 to 1. A higher value indicates a higher degree of confidence in the decision for that frame. For example, if frame 101 has high spatial overlap and no significant blur, its confidence level can reach 0.92. If frame 99 has significant occlusion, the confidence level drops to 0.65.

[0116] This frame set is then passed as input to the auxiliary judgment interface, which simultaneously displays each frame, the extracted foot key point trajectory information, the AI-generated penalty candidate conclusion (such as suspected false start), the false start time difference (such as -0.021 seconds), and the confidence value associated with each frame. The interface layout can adopt a multi-window partition format: the left side plays the frame images sequentially, the right side displays a dynamic rendering of the key point trajectory in the current frame, and the information bar below is set to simultaneously display comprehensive judgment information such as confidence level, time difference, and occlusion level.

[0117] In addition, the assisted judgment interface provides three referee operation options: confirm, modify, and reject, allowing referees to manually intervene in the system-generated false start candidate results. If the referee believes the image basis for the penalty is clear and the confidence level is higher than 0.85, they can click Confirm. If the recognition frame is unclear or the penalty result is ambiguous, they can choose to modify and manually specify the action start frame. If they believe there is no false start, they can choose to reject and generate a manual correction record.

[0118] By displaying penalty key frame images, action trajectories and scoring results through a visual interface, referees can quickly review key events and improve the efficiency of penalty review; the interface supports "confirm-modify-reject" operations, giving the AI ​​penalty mechanism a manual supervision interface, enhancing the applicability and fault tolerance of the AI ​​system.

[0119] Step 6: Collect manual review results and optimize key point detection weights and start timing judgment parameters based on the difference samples between the candidate penalty results and the review results, including the following:

[0120] Correction information from the manual review process needs to be collected, and the referee's corrected penalty samples are used as training data and fed back into the model to jointly update the key point detection weights and start timing judgment parameters;

[0121] First, in the auxiliary judgment interface, when the referee modifies or rejects the false start candidate result output by the system, the system will automatically record the image frame number corresponding to the operation, the content of the penalty conclusion correction, and the timestamp of the modification. This information will be used as feedback sample metadata and marked as a manually corrected case. For example, if the AI ​​system determines that the start is on the 104th frame and the referee modifies it to the 106th frame, the record will include the original judgment of the 104th frame, the manual correction to the 106th frame, and the corresponding frame number and timestamp.

[0122] Subsequently, based on the recorded content, the modified image frame and several frames before and after it (e.g., 3 frames before and after, for a total of 7 frames) are constructed into a difference sample set. This set contains multiple pieces of information, including image frames, corresponding key point annotation information (AI prediction and manual correction comparison), time difference penalty marks (e.g., non-false start is corrected to false start), etc., and serves as training input for subsequent model parameter updates. Among them, the key point annotation error can be automatically extracted by the system through the pixel distance difference between the AI ​​prediction point and the manual confirmation point. For example, if the ankle point offset is 12 pixels, the error vector is constructed accordingly;

[0123] Then, the incremental training mechanism is started to optimize the parameters respectively. The specific process is as follows:

[0124] The optimization of key point detection weights is to fine-tune the weight coefficients in the original key point detection model based on the corrected annotation results in the difference samples, thereby enhancing the model's accuracy in identifying error-prone areas (such as blurred and occluded feet);

[0125] The optimization of the start timing judgment parameters is to correct the dynamic thresholds and feature extraction parameters (such as the sensitivity of acceleration mutation judgment) in the start judgment rules based on the corrected start frame timestamp. Taking actual values ​​as an example, if the system's original acceleration mutation recognition threshold is set at 0.45, and after multiple referee feedback it is found that this value is prone to early start judgment under low-definition conditions, the system will automatically lower the threshold to 0.38 to improve robustness.

[0126] This step realizes human-machine collaborative learning based on referee behavior, can continuously optimize detection and timing parameters according to historical correction records, and has the ability of self-evolution; through differential sample training and incremental update mechanism, it can be more adaptable to complex scenarios and reduce the risk of performance degradation during long-term use.

[0127] It should be noted that the thresholds involved in the embodiments can be determined according to specific scenarios and requirements.

[0128] The present invention synchronizes the time of the starting area image sequence with the starting gun signal to construct a standardized starting response sequence. This ensures the consistency of the image data with the penalty reference signal, improves the accuracy of subsequent analysis, and combines key point recognition with acceleration mutation detection to effectively capture the true starting moment of the starting action, significantly enhancing the sensitivity of false start detection.

[0129] By collecting multi-angle images and performing spatial reconstruction of key points, the system further addresses the interference caused by foot occlusion on referee judgments. A confidence scoring mechanism is constructed based on image clarity and key point overlap, enhancing the credibility and interpretability of the calls. Furthermore, the system integrates a referee-assisted judgment interface to enable visual interaction and manual confirmation of AI-generated decision results, improving the collaborative efficiency between the system and referees. Finally, by collecting referee correction information and constructing a differential sample set, the system optimizes detection parameters based on an incremental training mechanism, enabling AI to continuously learn and adapt to actual case studies.

[0130] Example 2: An intelligent AI construction system for track and field event referees, such as Figure 2 As shown, specifically including:

[0131] The starting data acquisition module is used to collect and synchronize the image sequence of the sprint starting area with the starting gun signal in real time to generate a standardized starting response sequence;

[0132] The starting feature module is used to extract the key point trajectory and force trend characteristics of the athlete's lower limbs from the starting response sequence, calculate the starting instant of the movement to identify the starting signal;

[0133] A penalty generation module is used to determine whether a false start event occurs based on the time difference between the start signal and the gunshot signal, and generate a candidate penalty result;

[0134] The penalty determination module is used to collect image sequences from two preset angles of the starting area and perform key point spatial reconstruction. It also evaluates the confidence level of the penalty based on the foot occlusion and the image change area.

[0135] The interface display module is used to jointly map the penalty candidate results and the penalty confidence to generate a referee interaction frame set and build an auxiliary judgment interface;

[0136] The penalty review module is used to collect manual review results and optimize the key point detection weights and start timing judgment parameters based on the difference samples between the candidate penalty results and the review results.

[0137] The above formulas are all dimensionless and calculated numerically. Specific dimension removal can be achieved by various means such as standardization, which will not be elaborated here. The formula is a formula obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0138] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, ATA hard drives, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state ATA hard drive.

[0139] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0140] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0141] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0142] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0143] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0144] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A method for constructing an intelligent AI for refereeing track and field events, characterized in that: The steps include: Real-time acquisition and time synchronization calibration of the sprint starting area image sequence and the starting gun signal to generate a standardized starting response sequence; Extract the key points of the athlete's lower limb trajectory and force trend characteristics from the starting response sequence, calculate the starting instant of the movement to identify the starting signal; Determine whether a false start occurs based on the time difference between the start signal and the gunshot signal, and generate a candidate penalty result; Collect image sequences from two preset angles at the starting area and perform keypoint spatial reconstruction. The confidence level of the penalty is assessed based on foot occlusion and image change areas. The penalty candidate results and penalty confidence levels are jointly mapped to generate a referee interaction frame set, and an auxiliary judgment interface is constructed; Collect manual review results and optimize key point detection weights and start timing judgment parameters based on the difference samples between the candidate penalty results and the review results; Extract the key points of the athlete's lower limb trajectory and force trend characteristics from the starting response sequence, calculate the starting instant of the movement to identify the starting signal. The specific process is as follows: Extracting the key position information of the lower limbs in the image frame from the standardized start response sequence, where the key position information of the lower limbs includes the toes, ankles, and knees; Calculate the spatial displacement and velocity changes of the lower limb key points between adjacent image frames; Identify the sudden change point of foot acceleration as the starting moment of the running action and mark the corresponding image frame number; Collect multi-view images of the starting area and perform keypoint spatial reconstruction. The penalty confidence is evaluated based on foot occlusion and image motion areas. The specific process is as follows: capturing image frames generated at the same time at a first camera position and a second camera position, respectively, and extracting coordinate information of key points of the foot in the image frames; Perform spatial mapping on the key points of the foot extracted from the first camera position and the second camera position, and calculate the spatial overlap to determine the occlusion level; Combined with the areas where contour clarity changes in the image frame, the penalty confidence results are constructed based on overlap and clarity to assist in penalty decision-making.

2. The method for constructing an intelligent AI for refereeing track and field events according to claim 1, characterized in that: The image sequence of the sprint starting area is collected in real time and time-synchronized with the starting gun signal to generate a standardized starting response sequence. The specific process is as follows: A sequence of image frames of the starting area is collected by a preset image acquisition device at a fixed sampling period, wherein the image frames cover the area between the starting blocks and the athlete's feet; The sound sensing module collects the starting gunshot signal in real time, and uses the first valid rising edge of the gunshot signal as the time synchronization benchmark; Using the timestamp of the gunshot signal as the reference point, the image frame sequence is uniformly timed, and image segments before and after the start are intercepted according to a fixed time window to construct a standardized start response sequence.

3. The method for constructing an intelligent AI for refereeing track and field events according to claim 2, characterized in that: Based on the time difference between the start signal and the gunshot, a false start is determined and a candidate penalty result is generated. The specific process is as follows: Calculate the time difference between the timestamp of the start action and the timestamp of the gunshot; The obtained time difference is compared with the set false start penalty threshold. If the time difference is less than the penalty threshold, it is identified as a false start event; Output the candidate results of false start penalties and associate the starting frame of the starting action and its key point change information.

4. The method for constructing an intelligent AI for refereeing track and field events according to claim 3, characterized in that: Perform spatial mapping on the key points of the foot extracted from the first camera position and the second camera position, and calculate the spatial overlap to determine the occlusion level, including: The image coordinates of the key points of the foot extracted from the first camera position are converted into unified reference plane coordinates through perspective projection; Mapping the image coordinates of the foot key points extracted from the second camera position to a unified reference plane through the same projection method; The average Euclidean distance of the two groups of foot key points in the reference plane is calculated, and the area is divided into clear area, slightly occluded area and severely occluded area according to the relationship between the average Euclidean distance and the preset occlusion distance threshold.

5. The method for constructing an intelligent AI for refereeing track and field events according to claim 4, characterized in that: Based on the relationship between the average Euclidean distance and the preset occlusion distance threshold, the image is divided into clear area, slightly occluded area, and severely occluded area, including: The preset occlusion distance threshold includes a first occlusion distance threshold and a second occlusion distance threshold, and the first occlusion distance threshold is smaller than the second occlusion distance threshold; When the average Euclidean distance is less than or equal to the first occlusion distance threshold, it is determined to be a clear area; When the average value of the Euclidean distance is greater than the first occlusion distance threshold and less than or equal to the second occlusion distance threshold, it is determined to be a slightly occluded area; When the average Euclidean distance is greater than the second occlusion distance threshold, it is determined to be a severe occlusion area.

6. The method for constructing an intelligent AI for refereeing track and field events according to claim 5, characterized in that: The penalty candidate results and penalty confidence levels are jointly mapped to generate a referee interaction frame set, and an auxiliary decision interface is constructed. The specific process is as follows: Extract key image frames adjacent to the starting action start frame and the gunshot frame from the starting response sequence; Combine the key frame sequence with the penalty confidence result to form a referee interaction frame set; The auxiliary judgment interface displays frame images, key point trajectories and penalty confidence information for the referee to confirm, modify or reject.

7. The method for constructing an intelligent AI for refereeing track and field events according to claim 6, characterized in that: Collect manual review results and optimize key point detection weights and start timing judgment parameters based on the difference samples between the candidate penalty results and the review results. The specific steps are as follows: Record the referee's operation information and corresponding image frame number for correcting the AI ​​output penalty results in the auxiliary judgment interface; The frame data corrected by the referee is constructed into a difference sample set, which is used as training input for updating the key point detection weights and adjusting the start timing judgment parameters; The incremental training method is used to update the two processing processes of key point detection weight update and start timing judgment parameter adjustment, and the learning adaptation of referee correction case is carried out.

8. A system for constructing an intelligent AI for refereeing track and field events, for implementing a method for constructing an intelligent AI for refereeing track and field events as claimed in any one of claims 1 to 7, characterized in that: include: The starting data acquisition module is used to collect and synchronize the image sequence of the sprint starting area with the starting gun signal in real time to generate a standardized starting response sequence; The starting feature module is used to extract the key point trajectory and force trend characteristics of the athlete's lower limbs from the starting response sequence, calculate the starting instant of the movement to identify the starting signal; A penalty generation module is used to determine whether a false start event occurs based on the time difference between the start signal and the gunshot signal, and generate a candidate penalty result; The penalty determination module is used to collect image sequences from two preset angles of the starting area and perform key point spatial reconstruction. It also evaluates the confidence level of the penalty based on the foot occlusion and the image change area. The interface display module is used to jointly map the penalty candidate results and the penalty confidence to generate a referee interaction frame set and build an auxiliary judgment interface; The penalty review module is used to collect manual review results and optimize the key point detection weights and start timing judgment parameters based on the difference samples between the candidate penalty results and the review results.

Images