An AI-based long jump motion anomaly recognition method, system and device

Abstract

The application provides a long jump motion abnormality recognition method, system and device based on AI, relates to the field of artificial intelligence, and through collecting full-process video images of long jump motion, performing feature enhancement and edge enhancement on the images, detecting small part targets in the enhanced images based on an improved target detection algorithm, acquiring real-time coordinates and motion trajectories of the small part targets, and acquiring lower limb joint information in combination with human body posture estimation, core features of a take-off moment and a landing moment are extracted from continuous frame images, specific moments of the take-off moment and the landing moment are determined according to a judgment condition, abnormal behaviors are recognized according to the key moments, the real-time coordinates and the motion trajectories of the small part targets, and long jump rules are combined, and a recognition result is output. Small part target motion trajectories can be accurately captured, specific moments of take-off and landing can be determined, various long jump abnormal behaviors can be accurately recognized, and the problem of determination deviation caused by missed detection and misdetected small part targets can be solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more particularly to an AI-based method, system, and device for identifying abnormalities in long jump. Background Technology

[0002] In the field of anomaly detection in the approach long jump in track and field, artificial intelligence image processing technology has gradually replaced traditional manual judgment. However, existing technologies still have many key shortcomings, making it difficult to meet the accuracy requirements of practical applications. Current mainstream recognition solutions are easily affected by background interference and limb occlusion, and cannot accurately capture the real-time coordinates and movement trajectory of small parts. They also lack dedicated feature enhancement and detection mechanisms for small targets, resulting in insufficient detection accuracy for small parts such as ankles and toes, and widespread false negatives and missed detections. At the same time, existing technologies mostly rely on single features to determine the take-off and landing time, lacking the linkage analysis of small part movement features and human posture. In addition, the lack of multi-camera cross-validation and dynamic threshold adaptation mechanisms leads to large errors in time determination, resulting in biased judgments of core rule-based anomalies such as stepping on the line during take-off and invalid landing. Furthermore, some solutions have poor environmental robustness, with detection performance significantly decreasing in complex scenes such as backlighting and shadows. Moreover, most systems can only implement simple rule-based judgments and cannot quantitatively identify technical anomalies such as chaotic approach running points and insufficient take-off force.

[0003] The aforementioned problems limit the practicality of existing AI recognition systems in scenarios such as school physical tests, professional training, and small-scale competitions, making it difficult to guarantee the reliability and objectivity of the judgment results. Therefore, developing a robust and comprehensive long jump anomaly recognition solution that can accurately detect small targets such as the ankle and toes, accurately determine the take-off / landing moment, and is highly effective has become an urgent technical problem to be solved in this field. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides an AI-based method, system, and device for identifying abnormal long jump movements. By constructing a high-precision and robust long jump abnormality identification system, it can accurately capture the movement trajectory of small targets such as the ankle and toes, accurately determine the critical moments of take-off and landing, and thus accurately identify various abnormal long jump behaviors. This solves the problem of judgment bias caused by missed or false detections of small targets, improves the accuracy and reliability of abnormality identification, and is suitable for deployment in various scenarios such as campus physical tests, professional training, and small-scale competitions.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect of the present invention, an AI-based method for identifying abnormal long jump movements is provided, comprising the following steps: S1: Acquire video images of the entire long jump process to focus on capturing the motion status of small targets during the approach run, take-off, and landing phases, including the ankle and toes; S2: After performing feature enhancement and edge enhancement processing on the image, an enhanced image is obtained. Small targets in the enhanced image are detected based on an improved target detection algorithm to obtain the real-time coordinates and motion trajectory of the small targets. The lower limb joint information is obtained by combining human posture estimation. S3: Based on the real-time coordinates and motion trajectory of the small target, as well as the information of the lower limb joints, extract the core features of the take-off time and landing time from the continuous frame images, and determine the specific time of the take-off time and landing time according to the judgment conditions. S4: Based on the real-time coordinates and movement trajectory of the key moment and small target, combined with the long jump rules, identify abnormal behavior and output the identification results.

[0006] As a preferred method, when acquiring images, three cameras are deployed respectively on the side of the runway, directly in front of the take-off line, and directly in front of the sandpit to form a three-dimensional acquisition perspective. The camera on the side of the runway is aimed at the runway trajectory, focusing on capturing the continuous movement of the ankle and toes; the camera directly in front of the take-off line is aimed at the take-off line and the take-off foot, focusing on capturing the positional relationship between the ankle, toes and the take-off line at the moment of take-off; and the camera directly in front of the sandpit is aimed at the landing area of ​​the sandpit, focusing on capturing the contact state of the ankle and toes at the moment of landing.

[0007] Preferably, a modified target detection algorithm is used to detect small targets in the feature-enhanced image to obtain the real-time coordinates and motion trajectory of the small targets, and the whole-body joint information is obtained by combining human pose estimation, including the following steps: A multi-scale feature pyramid is constructed to perform multi-scale scaling on the feature-enhanced image, and channel attention and spatial attention mechanisms are introduced to perform feature weighting on the processed feature-enhanced image. Based on the feature-weighted image, the Sobel operator is used to enhance the edges of small targets to obtain an edge-enhanced image. The improved YOLOv8-nano is then used to perform target detection on the edge-enhanced image, and the detection results of small targets are output. Based on MediaPipe Pose, the joints of the lower limbs of the human body are extracted. Combined with the detection results of small targets, the coordinate information of the joints of the lower limbs of the human body is obtained. A dynamic human motion posture model is constructed based on the joint coordinate information to help determine the motion state of small targets. The joints of the lower limbs of the human body include the hip joint, knee joint, ankle joint, and toes.

[0008] Furthermore, the improved YOLOv8-nano adds a small target detection branch to the neck of YOLOv8-nano to increase the number of layers for small target feature extraction and improve the detection accuracy of small parts; it introduces the Focal Loss function, adjusts the weights of positive and negative samples to reduce the interference of background samples on small part samples, and adds coordinate loss weights to improve the accuracy of small part coordinate detection; and it redesigns the anchor frame size to take into account the size characteristics of ankles and toes to improve the matching degree between small part targets and anchor frames.

[0009] As a preferred method, after target detection is completed, a weighted average algorithm is used to fuse the coordinates of the same small target detected by multiple cameras. The weights are set according to the camera angle and detection confidence to obtain the real-time coordinates of the small target and eliminate the bias of single-camera detection. A Kalman filter algorithm is used to track the coordinates of the small target in real time and predict the motion trajectory of the small target.

[0010] Preferably, the core features at the takeoff moment include velocity change characteristics, contact state characteristics, attitude linkage characteristics, and positional relationship characteristics; the core features at the landing moment include velocity drop characteristics, contact occurrence characteristics, center of gravity change characteristics, and attitude stability characteristics.

[0011] Preferably, the conditions for determining the take-off time include: vertical speed of the toes ≥ 2m / s, contact area between the toes and the ground is 0, and the knee angle decreases rapidly from the maximum bending value; the conditions for determining the landing time include: vertical speed of the toes ≤ 0.5m / s, contact area between the toes and the sandpit ground > 0, and the vertical acceleration of the body's center of gravity changes from negative to positive.

[0012] Preferably, the abnormal behaviors include abnormal take-off stepping on / crossing the line, abnormal landing, abnormal approach run, and technical abnormalities; the abnormal approach run includes crossing the boundary, pausing, and exceeding the time limit; the technical abnormalities include disordered steps, abnormal take-off angle, and falling forward after landing.

[0013] In a second aspect of the invention, an AI-based long jump motion anomaly recognition system is provided, comprising: Image acquisition module: used to acquire video images of the entire long jump process, focusing on capturing the motion state of small targets during the approach run, take-off, and landing phases, including the ankle and toes; Target enhancement module: Used to enhance the features and edges of small targets in an image to obtain an enhanced image; Target detection module: Detects small targets in the enhanced image based on an improved target detection algorithm to obtain the real-time coordinates and motion trajectory of the small targets, and obtains lower limb joint information by combining human pose estimation; The timing determination module is used to extract the core features of the take-off time and landing time from continuous frame images based on the real-time coordinates, movement trajectory, and lower limb joint information of the small target, and to determine the specific time of the take-off time and landing time according to the judgment conditions. Anomaly identification module: used to identify abnormal behavior based on the real-time coordinates and movement trajectory of the target at the key moment and in small parts, combined with the long jump rules, and output the identification results.

[0014] In a third aspect of the invention, a long jump anomaly identification device is provided, the device comprising: a memory, a processor, and a long jump anomaly identification program stored in the memory and executable on the processor, the long jump anomaly identification program being configured to implement the steps of the AI-based long jump anomaly identification method as described in any one of claims 1 to 8.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: To address the issues of missed and false detections in small areas and improve target detection accuracy, YOLOv8-nano is improved from three dimensions: model structure, anchor box design, and loss function. Simultaneously, feature enhancement through multi-scale feature pyramids and dual attention mechanisms, along with Sobel edge enhancement, effectively solves the problem of missed and false detections in small areas such as ankles and toes in existing technologies. Furthermore, through multi-camera coordinate fusion and Kalman filter trajectory tracking, the accuracy rate for small target detection is ≥95%, and the false detection rate is ≤3%, laying a precise data foundation for subsequent time-lapse judgment and anomaly identification.

[0016] Accurately determine the critical moment of takeoff / landing and eliminate judgment bias: This invention integrates the speed, contact state, and displacement characteristics of small parts (ankle / toe) with the posture linkage characteristics of human lower limb joints. It uses LSTM temporal modeling, Kalman filter trajectory smoothing, and multi-feature cross-validation to determine the moment of takeoff / landing. At the same time, it combines multi-camera weighted average calibration and dynamic threshold adaptation to different sports groups to achieve a takeoff / landing moment judgment error of ≤40ms (1 frame). This completely solves the problem of moment judgment bias caused by inaccurate detection of small parts in existing technologies and provides a reliable time benchmark for accurate judgment of rule-based anomalies.

[0017] Improving the accuracy and reliability of anomaly identification: This invention uses precise small-part coordinates, movement trajectories, and take-off / landing times as a basis, combined with long jump competition rules, to achieve intelligent identification of all types of anomalies, including take-off line stepping / crossing the line, invalid landing, abnormal run-up (crossing the boundary / pausing / exceeding the time limit), and technical anomalies (disordered steps / abnormal take-off angle / landing with a forward fall). Through a multi-judgment logic of "coordinate comparison - temporal verification - linkage feature judgment", the anomaly identification accuracy rate is ≥98%. Compared with the subjectivity of traditional manual judgment and the mechanical nature of existing AI recognition, this invention significantly improves the objectivity, accuracy, and reliability of anomaly identification.

[0018] Enhanced system environmental robustness and adaptability to multiple practical applications: This invention adopts a three-camera three-dimensional perspective collaborative acquisition, combined with supplementary lighting and image preprocessing, which effectively solves the interference of complex environments such as backlight, shadow, venue reflection, and limb self-occlusion on detection. It can be stably deployed in multiple scenarios such as campus physical testing, professional athlete training, and small-scale event judging, breaking through the limitation of some existing recognition systems that are only suitable for fixed laboratory / training hall scenarios. Attached Figure Description

[0019] Figure 1 This is a flowchart of an AI-based method for identifying abnormal long jump movements according to the present invention. Figure 2 This is a block diagram of an AI-based long jump motion anomaly recognition system according to the present invention; Figure 3 This is a partial flowchart of step S2 of the present invention. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0021] The long jump mentioned in this invention specifically refers to the approach long jump in track and field, which is the standard long jump event in which the athlete accelerates through the approach track, takes off on one foot from the take-off board, and lands in the sandpit after an aerial maneuver, rather than the standing long jump, triple jump, or other jumping events.

[0022] Please see Figure 1 As shown, this invention provides an AI-based method for identifying abnormal long jump movements, comprising the following steps: S1: Acquire video images of the entire long jump process to focus on capturing the motion status of small targets during the approach run, take-off, and landing phases, including the ankle and toes; A high-definition industrial camera (resolution ≥ 1920×1080) with a frame rate ≥ 30FPS (ensuring detailed capture of small, fast-moving targets) is selected. A fixed-focus lens (focal length 25-50mm) is used, supporting autofocus to reduce motion blur. A supplementary light (adjustable brightness) is also provided to handle backlighting and shadow scenarios, preventing unclear capture of small targets due to insufficient light. This invention deploys a 3-camera collaborative acquisition system, ensuring simultaneous acquisition by all three cameras through a synchronous triggering mechanism. The acquisition time error is ≤ 10ms, preventing misalignment of small target movement trajectories due to asynchronous acquisition time and ensuring the continuity of ankle and toe movement sequences.

[0023] When acquiring images, three cameras were deployed at the side of the runway, directly in front of the take-off line, and directly in front of the sandpit, respectively, to form a three-dimensional acquisition perspective. The camera at the side of the runway (1.5-2m from the runway, 1.2-1.5m above the runway) was aimed at the runway trajectory, focusing on capturing the continuous movement of the ankle and toes. The camera at the direct front of the take-off line (2-3m from the take-off line, 0.8-1m above the take-off line) was aimed at the take-off line and the take-off foot, focusing on capturing the positional relationship between the ankle, toes, and the take-off line at the moment of take-off. The camera at the direct front of the sandpit (2-3m from the edge of the sandpit, 1-1.2m above the sandpit) was aimed at the landing area of ​​the sandpit, focusing on capturing the touchdown state of the ankle and toes at the moment of landing.

[0024] S2: After performing feature enhancement and edge enhancement processing on the image, an enhanced image is obtained. Small targets in the enhanced image are detected based on an improved target detection algorithm to obtain the real-time coordinates and motion trajectory of the small targets. The lower limb joint information is obtained by combining human posture estimation. Before any operation, the original image needs to be preprocessed. A combination of Gaussian filtering and median filtering is used to remove salt-and-pepper noise and Gaussian noise from the image, preserving the detailed features of small targets and avoiding misjudgments caused by noise. Histogram equalization algorithm is used to adjust the brightness and contrast of the image to maximize the grayscale difference between the ankle and toe areas and the background, avoiding blurring of small target features due to insufficient light. Based on camera calibration parameters, distortion correction is performed on the image to correct the positional shift of small targets caused by lens distortion, ensuring the accuracy of the coordinates of small targets.

[0025] Please see Figure 3 As shown, a modified target detection algorithm is used to detect small targets in the feature-enhanced image to obtain their real-time coordinates and motion trajectories. This is combined with human pose estimation to obtain full-body joint information. The process includes the following steps: A multi-scale feature pyramid is constructed to perform multi-scale scaling on the feature-enhanced image, and channel attention mechanism (SE-Net) and spatial attention mechanism (CBAM) are introduced to perform feature weighting on the processed feature-enhanced image; By constructing a multi-scale feature pyramid, the feature-enhanced image is scaled at multiple scales to form a multi-level feature map from coarse to fine, thereby improving the feature response capability for small-scale targets (such as ankles and toes) and preventing small targets from being obscured by background or large target features. Based on this, a channel attention mechanism (SE-Net) and a spatial attention mechanism (CBAM) are introduced to weight the processed feature-enhanced image. Specifically, the channel attention mechanism (SE-Net) learns the importance of different feature channels through operations such as global average pooling and fully connected layers, automatically assigning higher weights to feature channels relevant to small target recognition, suppressing redundant channels, and improving feature representation capability. The spatial attention mechanism (CBAM) learns the importance of spatial location through convolution operations, generating a spatial attention map that focuses on the spatial region where the small target is located, suppressing background interference, and making the model more focused on key locations such as ankles and toes. The results of channel attention and spatial attention are fused to comprehensively weight the feature-enhanced image, further strengthening the features of small targets and providing clearer and more discriminative feature inputs for subsequent target detection.

[0026] Based on the feature-weighted image, the Sobel operator is used to enhance the edges of small targets to obtain an edge-enhanced image. The improved YOLOv8-nano is then used to perform target detection on the edge-enhanced image, and the detection results of small targets are output. By calculating the gradient value of each pixel in the feature-weighted image, regions with drastic pixel value changes (i.e., edges) are identified and enhanced. These regions are then reinforced to highlight the edge contours of small targets (such as toes and ankle joints), making the shape and boundaries of small targets clearer, reducing detection difficulty, and decreasing false positives and false negatives. Subsequently, an improved YOLOv8-nano model is used to perform target detection on the edge-enhanced image. Specifically, multi-camera edge-enhanced images are used as input to the detection model. The improved YOLOv8-nano model extracts the features of small targets (ankles and toes) from the edge-enhanced image, and outputs the bounding box, confidence score (threshold set to 0.7 to avoid false positives), and coordinate information of the small targets.

[0027] The improved YOLOv8-nano adds a small target detection branch to the neck of YOLOv8-nano to increase the number of layers for small target feature extraction and improve the detection accuracy of small parts. It introduces the Focal Loss function to adjust the weights of positive and negative samples, reduce the interference of background samples on small part samples, and solve the problem of small part samples being few in number and easily overlooked. At the same time, it adds coordinate loss weights to improve the accuracy of small part coordinate detection. The anchor frame size is redesigned for the characteristics of ankle and toe sizes (such as 10×15, 15×20, 20×25) to improve the matching degree between small part targets and anchor frames.

[0028] Specifically, to address the issues of insufficient sample size, susceptibility to background interference, and inadequate coordinate detection accuracy in detecting small targets (ankles, toes), this invention improves YOLOv8-nano from three dimensions: model structure, anchor frame design, and loss function, forming a complete and accurate small target detection model.

[0029] Model Structure: A small object detection branch is added to the neck of YOLOv8-nano. That is, on the basis of the original feature fusion structure, a new small object feature extraction path is added. Through 1×1 convolution and upsampling operations, deep semantic features and shallow fine-grained features are fused to form a feature map specifically for small object detection. The number of small object feature extraction layers is increased, so that the model can obtain stronger feature expression ability at small-scale feature layers and avoid small parts (ankles, toes) being compressed and lost in deep features. The small object detection branch works in parallel with the main detection branch, and the final detection result is jointly decided by the two branches, which improves the recall rate of small part detection.

[0030] Anchor frame design: Based on dataset statistics, the pixel size of the ankle and toe in the image is mainly concentrated in the range of 10×15, 15×20, and 20×25. Considering the size characteristics of the ankle and toe, the anchor frame size is redesigned to [10×15, 15×20, 20×25]. By replacing the original general anchor frame of YOLOv8-nano, the anchor frame size is made to better match the actual size of small targets, improving the matching degree between small targets and anchor frames and reducing missed detections and false detections caused by mismatch between anchor frames and targets. The anchor frame ratio is maintained at 1:1, 1:2, and 2:1 to take into account different postures of small targets (such as toes pointing forward and ankles turning to the side).

[0031] Loss Function: To further improve the detection accuracy and coordinate localization accuracy of small targets, the loss function is improved by introducing Focal Loss as the classification loss. The weights of positive and negative samples are adjusted to reduce the interference of background samples on small target samples, addressing the problem of small target samples being few in number and easily overlooked. The calculation formula is as follows: ; In the formula, This represents the model's predicted probability for the target category, with a value range of [0,1]. It is used when the true label is a positive sample (a small part of the target, such as the ankle or toe). When the true label is the sample (background), ; This represents the class balance factor, used to balance the importance of positive and negative samples. For positive samples... When negative samples ,in γ represents the focusing parameter, which is used to reduce the loss weight of easily distinguishable samples (such as clear backgrounds) and make the model pay more attention to difficult samples (such as small targets). In this invention, the value is 2.

[0032] CIoU Loss is used as the coordinate loss, while also considering the intersection-union ratio, center point distance, and aspect ratio consistency, and their weights are increased to improve the accuracy of bounding box coordinate detection for small targets. The calculation formula is as follows: ; ; In the formula, IoU represents the distance between the predicted bounding box b and the ground truth bounding box b. The intersection and union ratio, This represents the center point (x, y) of the predicted bounding box and the center point of the ground truth bounding box. The squared Euclidean distance between them; c represents the diagonal length of the smallest bounding rectangle that simultaneously contains the predicted and ground truth boxes; v measures the consistency of the aspect ratio between the predicted and ground truth boxes. This represents the weighting coefficient.

[0033] The Focal Loss and CIoU Loss are weighted and fused to form the total loss function, calculated as follows: ; In the formula, This represents the Focal Loss weight, with a value of 1.0. This represents the CIoU Loss weight, with a value of 2.0, which enhances the learning of the accuracy of small part coordinate detection.

[0034] After target detection is completed, a weighted average algorithm is used to fuse the coordinates of the same small target detected by multiple cameras. The weights are set according to the camera angle and detection confidence to obtain the real-time coordinates of the small target and eliminate the bias of single-camera detection. A Kalman filter algorithm is used to track the coordinates of the small target in real time, predict the motion trajectory of the small target, fill in the coordinate missing caused by momentary occlusion, and avoid missing small targets. For example, when the ankle is occluded by the lower leg at the moment of take-off, the coordinates are supplemented by trajectory prediction.

[0035] The final detection results include the bounding box, confidence score, real-time coordinates, and motion trajectory of small targets.

[0036] Based on MediaPipe Pose, the joints of the lower limbs of the human body are extracted. Combined with the detection results of small targets, the coordinate information of the joints of the lower limbs of the human body is obtained. A dynamic human motion posture model is constructed based on the joint coordinate information to help determine the motion state of small targets. The joints of the lower limbs of the human body include the hip joint, knee joint, ankle joint, and toes.

[0037] The specific implementation process is as follows: First, a lightweight MediaPipe Pose model is deployed and basic joints are extracted. This model supports 2D / 3D pose estimation. Long jump video frames, after small target enhancement and edge enhancement processing, are used as input. For the detection requirements of long jump, the lower limb joint detection branch is activated, extracting only eight lower limb joints: left and right hip joints, left and right knee joints, left and right ankle joints, and left and right toes. This effectively reduces the computational overhead of non-core joints and ensures the real-time performance of the model. The MediaPipe Pose model extracts feature information from the input image through the BlazePose backbone network and combines it with a keypoint regression algorithm to output the pixel coordinates (x, y) or 3D world coordinates (x, y, z) of the above eight lower limb joints, where the z-axis represents depth information. It also outputs the detection confidence score for each joint, with a confidence threshold of 0.8. When the confidence score is ≥0.8, the joint is considered valid; when the confidence score is <0.8, the joint is marked as "to be completed".

[0038] Then, the detection results of small targets are fused with the joint coordinates extracted by MediaPipe to obtain the final coordinate information of the joints. Considering that the MediaPipe Pose model has insufficient detection accuracy for extremely small targets such as toes, the joint coordinates are calibrated and completed by combining the small target detection results of the aforementioned improved YOLOv8-nano to improve coordinate accuracy.

[0039] For the two core small parts, the center point coordinates of the bounding box of the small part detected by YOLOv8-nano are weighted and fused with the corresponding joint point coordinates output by MediaPipe Pose to obtain the final joint point coordinates. The weight coefficients are dynamically adjusted according to the detection confidence of the two to ensure that the fused coordinates are closer to the actual position of the small part.

[0040] For hip and knee joints, since the pixel area of ​​these joints is relatively large, MediaPipe Pose has high detection accuracy, so the joint coordinates output by this model are directly used.

[0041] If the confidence score of a lower limb joint point in MediaPipe Pose is <0.8 due to factors such as limb self-occlusion and environmental interference, the coordinates of the missing joint point are derived by reverse inversely using the effective ankle and toe coordinates detected by YOLOv8-nano, based on the physiological constraints of human movement (such as the linkage between the knee and ankle joints and the proportions of the human skeleton). The derivation formula is: P 补全 =P 已知 +k·Δ; where P 补全 P represents the coordinates of the missing joints after completion. 已知 This represents the effective small-scale coordinates detected by YOLOv8-nano, k represents the physiological constraint coefficient (e.g., the vertical distance coefficient between the knee and ankle joints is preset to 1.2), and Δ represents the standard offset between known joints and missing joints, which is preset based on the proportions of the human skeleton.

[0042] Subsequently, a dynamic human motion posture model is constructed: based on the fused lower limb joint coordinates, the joints are associated in the order of "hip joint → knee joint → ankle joint → toe" according to the human physiological structure, forming two lower limb skeletal chains. The relative angles and Euclidean distances between adjacent joints are calculated to quantify lower limb posture features. For example, the knee flexion angle is calculated using the vector dot product formula: knee flexion angle = arccos[(hip joint - knee joint vector)] (knee-ankle joint vector) / (|hip-knee joint|×|knee-ankle joint|)). Simultaneously, the joint coordinates, angles, and distances of consecutive video frames are temporally sampled at a frequency consistent with the video frame rate (30 FPS) to form a temporal pose sequence. A Kalman filter is then used to smooth this temporal sequence, eliminating joint coordinate jumps caused by single-frame noise and ensuring the continuity and stability of the human motion pose model.

[0043] Finally, based on the constructed human motion posture model, the system assists in judging the motion state of small parts of the target. Combining the two key stages of the long jump—takeoff and landing—the system achieves accurate judgment of the motion state of small parts from three dimensions: changes in joint angles, trends in coordinate displacement, and the linkage relationship of lower limb joints.

[0044] During the take-off phase, the movement state of the ankle and toes is judged by judging the changes in the angle of the ankle and knee joints (at the moment of take-off, the angle between the ankle and knee joints rapidly increases from 120°±10° to 170°±5°, i.e., the ankle completes the push-off action), the vertical displacement of the toe coordinate (vertical displacement ≥5cm at the moment of take-off), and the contact state between the toes and the ground (from "contact" to "suspended"). At the same time, the effectiveness of the take-off action is judged by verifying whether the changes in the angles of the hip, knee, and ankle joints conform to the "push-off force" sequence (knee joint bends first and then extends → ankle joint extends synchronously → toes leave the ground). If the sequence does not conform, it is marked as "abnormal take-off posture".

[0045] During the landing phase, the movement state of the ankle and toes is determined by judging the changes in ankle angle (the ankle angle rapidly decreases from 170°±5° to 120°±10° at the moment of landing, i.e., the ankle completes the cushioning action), the vertical displacement of the toe coordinate (vertical displacement ≤-5cm at the moment of landing), and the contact state between the toes and the sandpit ground (from "suspended" to "in contact"). At the same time, the effectiveness of the landing action is determined by verifying whether the changes in the angles of the hip, knee, and ankle joints conform to the "landing cushioning" sequence (toes touch the ground first → ankle joint flexion → knee joint flexion simultaneously). If the sequence is not conformed to, it is marked as "abnormal landing posture (such as sitting back or falling forward)".

[0046] In addition, the normal threshold range of lower limb joint angles and displacements is preset (e.g., an ankle extension angle ≥160° is normal, and <150° is abnormal "insufficient take-off force"). The real-time characteristics of the human motion posture model are compared with the preset thresholds to output the quantitative judgment results of the motion state of small parts.

[0047] S3: Based on the real-time coordinates and motion trajectory of the small target, as well as the information of the lower limb joints, extract the core features of the take-off time and landing time from the continuous frame images, and determine the specific time of the take-off time and landing time according to the judgment conditions. The core features of the takeoff moment (the instant from the athlete's run-up push-off to complete takeoff) include: velocity abrupt change (the vertical velocity of the toes / ankles increases abruptly from near 0 to ≥2 m / s within one frame, while the horizontal velocity increases significantly, indicating the completion of the push-off force); contact state features (the contact area between the toes and the ground changes from present to absent, switching from a contact state to a suspended state, indicating that the body is completely off the ground); posture linkage features (the knee joint flexion angle reaches its maximum value and then rapidly decreases (≥10° / frame), while the hip joint angle opens synchronously, forming a linkage posture for the push-off force); and positional relationship features (at the moment of takeoff, the distance between the coordinates of the toes / ankles and the takeoff line reaches its minimum or just crosses it); and the landing moment... The core features of the moment (the instant the athlete first contacts the sandpit ground from the air) include the following characteristics: a sudden drop in speed (the vertical velocity of the toes / ankles drops sharply from a high speed of ≥3m / s during hang time to ≤0.5m / s within 1 frame, indicating the start of the ground contact buffer), the appearance of contact (the contact area increases from 0 when the toes / ankles first contact the sandpit ground, indicating that the body begins to bear the ground reaction force), the change in center of gravity (the vertical acceleration of the body's center of gravity calculated from the coordinates of the hip, knee, and ankle joints changes from negative to positive, indicating that the center of gravity begins to sink), and the stability of posture (after landing, the position of both feet or one foot no longer undergoes significant displacement within 3 frames, and the body posture tends to stabilize, indicating that the landing buffer is complete).

[0048] The coordinates, velocity, and acceleration of the toe / ankle, as well as the angle changes of the knee and hip joints, are extracted from consecutive frames. LSTM or a sliding window (e.g., 5 frames) is used to smooth and temporally model the features, capturing the continuous changes of "approach run → push-off → take-off" and eliminating single-frame noise interference. A weighted average is applied to the take-off moments detected by cameras positioned to the side of the runway and directly in front of the take-off line (weights are set according to detection confidence) to eliminate single-camera perspective bias and ensure that the take-off moment determination error is ≤40ms (1 frame). A take-off moment is determined when the following conditions are met simultaneously: vertical toe velocity ≥2m / s, toe-to-ground contact area is 0, and the knee angle rapidly decreases from its maximum bending value (≥10° / frame).

[0049] The coordinates, velocity, and acceleration of the toes / ankles, as well as the trajectory of the body's center of gravity, are extracted from consecutive frames. A Kalman filter is used to smooth the trajectory, eliminating jumps caused by occlusion or noise, and accurately capturing the "touch → cushioning → stabilization" process. A landing is determined when the following conditions are met simultaneously: vertical velocity of the toes ≤ 0.5 m / s, contact area between the toes and the sandpit surface > 0, and vertical acceleration of the body's center of gravity changes from negative to positive. If the first point of contact at the moment of landing is outside the sandpit and closer to the takeoff line than the nearest landing point inside the sandpit, the landing is considered invalid, and the coordinates at that moment are recorded as an anomaly detection criterion.

[0050] S4: Based on the real-time coordinates and movement trajectory of the key moment and small target, combined with the long jump rules, identify abnormal behavior and output the identification results.

[0051] The abnormal behaviors include abnormal take-off stepping / crossing the line, abnormal landing, abnormal approach run, and technical abnormalities; the abnormal approach run includes crossing the boundary, pausing, and exceeding the time limit; the technical abnormalities include disordered steps, abnormal take-off angle, and falling forward after landing.

[0052] Take-off line stepping / crossing abnormality: Based on the real-time coordinates of the toes and ankles at the moment of take-off, determine whether the toes / ankles touch or cross the take-off line (by comparing with the calibrated take-off line coordinates). If the coordinates of the toes / ankles at the moment of take-off exceed the range of the take-off line (deviation ≥ 2mm), it is judged as take-off line stepping / crossing abnormality.

[0053] Invalid landing anomaly: Based on the real-time coordinates of the toes and ankles at the moment of landing, determine whether the toes / ankles touched the ground outside the sandpit, and whether the contact point is closer to the take-off line than the nearest landing point inside the sandpit; at the same time, determine whether the athlete leaves the field from the front of the sandpit after landing. If the athlete leaves from the side / rear and the first contact point is closer to the take-off line than the landing point inside the sandpit, the landing is determined to be invalid.

[0054] Approach run abnormality: Based on the movement trajectory of the ankle and toes, determine whether the approach run has crossed the boundary or stopped / turned back midway (the trajectory reverses or stops); combined with the approach run sequence, determine whether the attempt jump exceeds the time limit.

[0055] Technical anomalies: Based on the movement trajectory and posture characteristics of the ankle and toes, determine whether the approach point is chaotic (excessive fluctuation in ankle and toe displacement), whether the take-off angle is abnormal (excessive deviation in ankle and knee joint angle), and whether the landing is backward / forward (abnormal force direction of the ankle and toes at the moment of landing).

[0056] During the long jump, the system displays the real-time trajectory of small target areas (ankle, toes) and the takeoff / landing time. If an anomaly is detected, an alarm is immediately issued (audio and visual cues), indicating the anomaly type (e.g., stepping on the takeoff line) and the judgment criteria (e.g., toe coordinates exceeding the takeoff line by 3mm at the takeoff moment). After the long jump, a complete anomaly identification report is output, including the anomaly type, the time of occurrence, the judgment criteria (coordinates of the small target area, screenshot of motion characteristics), and a trajectory diagram of the small target area, facilitating user traceability and analysis. This invention supports integration with external devices (such as physical testing systems and training management systems) and can output standardized anomaly identification data (JSON format) for easy data statistics and analysis.

[0057] In a second aspect of the invention, please refer to Figure 2As shown, this invention provides an AI-based long jump motion anomaly recognition system for implementing the above-described method. The system includes the following modules: Image acquisition module: used to acquire video images of the entire long jump process, focusing on capturing the motion state of small targets during the approach run, take-off, and landing phases, including the ankle and toes; Target enhancement module: Used to enhance the features and edges of small targets in an image to obtain an enhanced image; Target detection module: Detects small targets in the enhanced image based on an improved target detection algorithm to obtain the real-time coordinates and motion trajectory of the small targets, and obtains lower limb joint information by combining human pose estimation; The timing determination module is used to extract the core features of the take-off time and landing time from continuous frame images based on the real-time coordinates, movement trajectory, and lower limb joint information of the small target, and to determine the specific time of the take-off time and landing time according to the judgment conditions. Anomaly identification module: used to identify abnormal behavior based on the real-time coordinates and movement trajectory of the target at the key moment and in small parts, combined with the long jump rules, and output the identification results.

[0058] In a third aspect of the present invention, a long jump anomaly identification device is provided, the device comprising: a memory, a processor, and a long jump anomaly identification program stored in the memory and executable on the processor, the long jump anomaly identification program being configured to implement the steps of the AI-based long jump anomaly identification method described above.

[0059] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An AI-based method for identifying abnormal long jump movements, characterized in that, Includes the following steps: S1: Acquire video images of the entire long jump process to focus on capturing the motion status of small targets during the approach run, take-off, and landing phases, including the ankle and toes; S2: After performing feature enhancement and edge enhancement processing on the image, an enhanced image is obtained. Small targets in the enhanced image are detected based on an improved target detection algorithm to obtain the real-time coordinates and motion trajectory of the small targets. The lower limb joint information is obtained by combining human posture estimation. S3: Based on the real-time coordinates and motion trajectory of the small target, as well as the information of the lower limb joints, extract the core features of the take-off time and landing time from the continuous frame images, and determine the specific time of the take-off time and landing time according to the judgment conditions. S4: Based on the real-time coordinates and movement trajectory of the key moment and small target, combined with the long jump rules, identify abnormal behavior and output the identification results.

2. The AI-based method for identifying abnormal long jump movements according to claim 1, characterized in that, When acquiring images, three cameras were deployed on the side of the runway, in front of the take-off line, and in front of the sandpit to form a three-dimensional acquisition perspective. The camera on the side of the runway was aimed at the runway trajectory, focusing on capturing the continuous movement of the ankle and toes; the camera in front of the take-off line was aimed at the take-off line and the take-off foot, focusing on capturing the positional relationship between the ankle, toes and the take-off line at the moment of take-off. The camera is positioned directly in front of the sandpit, aimed at the landing area, focusing on capturing the moment the ankles and toes touch the ground.

3. The AI-based method for identifying abnormal long jump movements according to claim 1, characterized in that, The algorithm detects small targets in the enhanced image based on an improved target detection algorithm to obtain their real-time coordinates and motion trajectories. It also combines this with human pose estimation to obtain full-body joint information. The process includes the following steps: A multi-scale feature pyramid is constructed to perform multi-scale scaling on the feature-enhanced image, and channel attention and spatial attention mechanisms are introduced to perform feature weighting on the processed feature-enhanced image. Based on the feature-weighted image, the Sobel operator is used to enhance the edges of small targets to obtain an edge-enhanced image. The improved YOLOv8-nano is then used to perform target detection on the edge-enhanced image, and the detection results of small targets are output. Based on MediaPipe Pose, the joints of the lower limbs of the human body are extracted. Combined with the detection results of small targets, the coordinate information of the joints of the lower limbs of the human body is obtained. A dynamic human motion posture model is constructed based on the joint coordinate information to help determine the motion state of small targets. The joints of the lower limbs of the human body include the hip joint, knee joint, ankle joint, and toes.

4. The AI-based method for identifying abnormal long jump movements according to claim 3, characterized in that, The improved YOLOv8-nano adds a small target detection branch to the neck of YOLOv8-nano to increase the number of layers for small target feature extraction and improve the detection accuracy of small parts; it introduces the Focal Loss function, adjusts the weights of positive and negative samples to reduce the interference of background samples on small part samples, and adds coordinate loss weights to improve the accuracy of small part coordinate detection; and it redesigns the anchor frame size to take into account the size characteristics of ankles and toes to improve the matching degree between small part targets and anchor frames.

5. The AI-based method for identifying abnormal long jump movements according to claim 3, characterized in that, After target detection is completed, a weighted average algorithm is used to fuse the coordinates of the same small target detected by multiple cameras. The weights are set according to the camera angle and detection confidence to obtain the real-time coordinates of the small target and eliminate the bias of single-camera detection. A Kalman filter algorithm is used to track the coordinates of the small target in real time and predict the motion trajectory of the small target.

6. The AI-based method for identifying abnormal long jump movements according to claim 1, characterized in that, The core features of the takeoff moment include abrupt velocity change, contact state, attitude linkage, and positional relationship; the core features of the landing moment include a sudden drop in velocity, contact occurrence, center of gravity change, and attitude stability.

7. The AI-based method for identifying abnormal long jump movements according to claim 6, characterized in that, The conditions for determining the take-off time include: vertical speed of the toes ≥ 2m / s, contact area between the toes and the ground is 0, and the knee angle decreases rapidly from the maximum bending value; the conditions for determining the landing time include: vertical speed of the toes ≤ 0.5m / s, contact area between the toes and the sandpit ground > 0, and the vertical acceleration of the body's center of gravity changes from negative to positive.

8. The AI-based method for identifying abnormal long jump movements according to claim 1, characterized in that, The abnormal behaviors include abnormal take-off stepping / crossing the line, abnormal landing, abnormal approach run, and technical abnormalities; the abnormal approach run includes crossing the boundary, pausing, and exceeding the time limit; the technical abnormalities include disordered steps, abnormal take-off angle, and falling forward after landing.

9. An AI-based long jump anomaly recognition system, applied to the AI-based long jump anomaly recognition method according to any one of claims 1-8, characterized in that, The system includes: Image acquisition module: used to acquire video images of the entire long jump process, focusing on capturing the motion state of small targets during the approach run, take-off, and landing phases, including the ankle and toes; Target enhancement module: Used to enhance the features and edges of small targets in an image to obtain an enhanced image; Target detection module: Detects small targets in the enhanced image based on an improved target detection algorithm to obtain the real-time coordinates and motion trajectory of the small targets, and obtains lower limb joint information by combining human pose estimation; The timing determination module is used to extract the core features of the take-off time and landing time from continuous frame images based on the real-time coordinates, movement trajectory, and lower limb joint information of the small target, and to determine the specific time of the take-off time and landing time according to the judgment conditions. Anomaly identification module: used to identify abnormal behavior based on the real-time coordinates and movement trajectory of the target at the key moment and in small parts, combined with the long jump rules, and output the identification results.

10. A device for identifying abnormal long jump movements, characterized in that, The apparatus includes: a memory, a processor, and a long jump anomaly identification program stored in the memory and executable on the processor, the long jump anomaly identification program being configured to implement the steps of the AI-based long jump anomaly identification method as described in any one of claims 1 to 8.

Images