An artificial intelligence-based highlight picture real-time capturing method

Abstract

The present application relates to the technical field of real-time snapshot of highlight pictures, and particularly relates to a real-time snapshot method of highlight pictures based on artificial intelligence. The method comprises the following steps: collecting physiological data and sports data of athletes; obtaining explosive power score, concentration score and attention score of athletes; setting a threshold range, defining the moment when the explosive power score, concentration score and attention score exceed the threshold range as a highlight moment; and taking a snapshot of the highlight picture. The present application can more accurately evaluate the explosive power performance of athletes by collecting their physiological data and sports data. The concentration score of athletes can be accurately evaluated by collecting their facial expression data in real time. The attention of the audience can be analyzed by pattern recognition algorithm through real-time collection of the sound and body data of the audience, thereby adding a new dimension to the identification of highlight moments. The intelligences camera technologies such as phase detection auto focus and pan-tilt automatic adjustment are integrated to ensure the clarity and continuity of the highlight pictures.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of real-time capture of highlight frames, and particularly relates to a real-time capture method of highlight frames based on artificial intelligence. BACKGROUND

[0002] Sports, as an important part of human history, has always been driving the progress of human civilization. Sports events are not only the stage of various sports competitions, but also the important window to reflect the physical, intellectual and spiritual outlook of human beings. From the ancient Greek Olympics to today's various top events around the world, sports events have always attracted the attention and enthusiasm of countless spectators. The reason for this is that sports events can show the infinite potential and great passion of human beings. On the field, athletes constantly break the limits of human physical ability with their amazing explosive power, high concentration and strong competitive desire. These moments of brilliance often evoke the cheers of the audience and resonate with countless people. It can be said that these precious "highlight moments" not only fully demonstrate the charm of sports, but also are an important factor in the lasting appeal of sports events.

[0003] However, in today's information age, it is difficult to fully capture these highlight moments by relying solely on existing on-site shooting techniques. Traditional event shooting is mostly controlled by manual camera operation, which is difficult to track the rapid movements of athletes in real time, resulting in blurred images or missed highlight moments. Relying solely on observation and shooting of athletes' external movements cannot accurately determine when the highlight moment is. Focusing only on the performance of athletes themselves ignores the feelings and feedback of the audience on the highlights, and cannot fully assess the true brilliance of a moment. Therefore, there is an urgent need for a real-time capture method of highlight frames based on artificial intelligence, which uses advanced perception technology, intelligent analysis algorithms and intelligent control technology to achieve comprehensive perception, accurate assessment and intelligent capture of athlete's state, in order to break through the limitations of existing technology and bring a more spectacular visual feast to sports event spectators. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the present application provides a real-time capture method of highlight frames based on artificial intelligence.

[0005] To solve the above technical problems, the present application provides the following technical scheme: a real-time capture method of highlight frames based on artificial intelligence, comprising:

[0006] S1, collecting physiological data and movement data of the athlete, the physiological data including the height, weight, heart rate and blood pressure of the athlete. The movement data includes the speed, movement amplitude, movement frequency and joint angle of the athlete.

[0007] S2. Extract features from the motion data to obtain motion features, extract features from the physiological data to obtain physiological features, construct an explosive power evaluation model based on a recurrent neural network, input the motion features and physiological features, and output the athlete's explosive power score.

[0008] S3. Collect athletes' facial expression data in real time, extract expression features from the facial expression data, establish a focus assessment index, and obtain athletes' focus scores based on expression features.

[0009] S4. Real-time collection of external attention data, including the voice and body data of the audience, feature extraction of the voice and body data, and calculation of the athlete's attention score through pattern recognition algorithm.

[0010] S5. Set the explosiveness score threshold, focus score threshold, and attention score threshold. When the explosiveness score exceeds the explosiveness score threshold, the focus score exceeds the focus score threshold, and the attention score exceeds the attention score threshold, it is judged as a highlight moment.

[0011] S6. Focus on the athlete's face and automatically adjust the shooting angle to capture exciting moments.

[0012] According to the present invention, a method for real-time capture of exciting scenes based on artificial intelligence, in step S2, the process of feature extraction of motion data includes:

[0013] S21. Using a deep learning-based target detection algorithm, locate the contours and coordinates of the athlete's key body parts.

[0014] S22. Construct a human dynamics model based on motion data to assess muscle strength.

[0015] S23. Calculate the motion velocity and acceleration of key parts based on the streamer method.

[0016] S24. Evaluate the mass distribution of an athlete's limbs based on motion data, and calculate the moment of inertia by combining the moment of inertia.

[0017] According to the present invention, a method for real-time capture of exciting scenes based on artificial intelligence, in step S2, the process of constructing an explosive power evaluation model based on a recurrent neural network includes:

[0018] The motion and physiological data are cleaned and standardized to obtain standardized data.

[0019] Construct a basic recurrent neural network model, and set the number of layers and the number of neurons in each layer.

[0020] Historical data is collected, and the athlete's historical athletic and physiological characteristics are extracted from the historical data. The historical athletic and physiological characteristics are used as inputs, and the explosive power score is used as the output to train a basic recurrent neural network model. The model parameters that meet the test accuracy are retained to obtain an explosive power evaluation model.

[0021] According to the present invention, a method for real-time capture of exciting scenes based on artificial intelligence, in step S3, the process of extracting facial expression features from facial expression data includes:

[0022] The Viola-Jones face detection algorithm is used to detect face regions, and N key feature points are set within the face region using an active shape model.

[0023] Based on the location of key feature points, the face region is divided into facial feature regions, which include the eyes, eyebrows, nose, lips, and face area.

[0024] Calculate the geometric quantification index between key feature points, which includes distance, angle, and area.

[0025] Texture analysis was performed on each facial feature region to extract Gabor and LBP texture features.

[0026] Track the motion trajectory of key feature points in consecutive frames, calculate the temporal features of each key feature point, and use them as dynamic features of the face region.

[0027] An expression vector is constructed using geometric quantization metrics, Gabor and LBP texture features, and dynamic features, and then the expression vector is used as the expression feature.

[0028] According to the present invention, a real-time capture method for capturing exciting moments based on artificial intelligence includes the following steps in step S3: establishing a focus assessment model; collecting facial expression data of athletes at different levels of focus and labeling them with corresponding focus score tags; training the focus assessment model and retaining model parameters that meet a preset accuracy rate; and outputting focus scores by taking facial expression features as input.

[0029] According to the present invention, a method for real-time capture of exciting scenes based on artificial intelligence, in step S4, the process of feature extraction of sound data includes:

[0030] The audio data is denoised, and the continuous audio data is segmented to obtain short-frequency audio signals.

[0031] Spectral analysis of short-frequency sound signals yields short-frequency sound characteristics, including volume characteristics, frequency characteristics, and prosodic characteristics.

[0032] A recursive feature elimination method was used to filter short-frequency sound features to obtain relevant sound features.

[0033] According to the present invention, a method for real-time capture of exciting scenes based on artificial intelligence includes the following steps in step S4:

[0034] The limb data is calibrated and filtered, and continuous limb movements are segmented into independent action units, including clapping and waving movements.

[0035] Extract spatial and temporal features from the action unit. Spatial features include the amplitude, direction, and displacement distance of the action, while temporal features include the duration and frequency of the action.

[0036] Spatial and temporal features are standardized to obtain relevant limb features.

[0037] According to the present invention, a method for real-time capture of exciting moments based on artificial intelligence, in step S4, the process of calculating the athlete's attention score using a pattern recognition algorithm includes: quantifying relevant sound features and relevant body features, including the sound range of the relevant sound features and the movement range of the relevant body features; setting an index score for each sound range and each movement range; and weighted summing the sound range index score and the movement range index score to obtain the attention score.

[0038] According to the present invention, a method for real-time capture of exciting scenes based on artificial intelligence, in step S6, the process of focusing on the athlete's face includes: using phase detection autofocus technology, determining the position of the focal plane by detecting high-frequency signals in the image, and combining face detection and eye detection technologies to achieve focus on the athlete's face.

[0039] According to the present invention, a method for real-time capture of exciting scenes based on artificial intelligence, in step S6, the process of automatically adjusting the shooting angle includes:

[0040] The motion state of the camera body is detected using a three-axis accelerometer and a gyroscope sensor.

[0041] Based on the athlete's position and movement in the frame, the camera angle is automatically adjusted using a gimbal control algorithm.

[0042] By combining image recognition technology, the system tracks and locks onto the position of athletes in the frame, capturing exciting moments.

[0043] This invention provides a real-time capture method for highlight moments based on artificial intelligence. By collecting athletes' physiological and kinematic data, it constructs a three-dimensional description of the athlete's state. This includes not only the athlete's basic physical characteristics but also key indicators such as their motor abilities and physical condition. Based on this comprehensive perceptual data, it can more accurately assess the athlete's explosive performance, laying a solid foundation for the identification of highlight moments. Secondly, this scheme collects athletes' facial expression data in real time and uses computer vision technology to extract facial features, accurately assessing the athlete's focus score. Compared to relying solely on manual observation, this focus analysis based on facial micro-expressions is more objective and accurate, providing a comprehensive understanding of the athlete's psychological state and increasing the accuracy of highlight moment identification. Simultaneously, it can also collect the sound and body data of the audience in real time, analyzing the audience's attention through pattern recognition algorithms, adding a new dimension to highlight moment identification. Based on a comprehensive evaluation of the athlete's explosive power, focus, and audience attention, this scheme sets corresponding scoring thresholds. Only when all three indicators simultaneously meet the threshold is it determined to be a highlight moment. This strict logical relationship ensures that the identification of highlight moments is more reliable and accurate, avoiding false alarms or omissions. Furthermore, it possesses the ability to flexibly adjust thresholds, allowing for optimization based on the characteristics of different sports and continuously improving recognition accuracy. Finally, the solution integrates intelligent camera technologies such as phase detection autofocus and gimbal auto-adjustment, enabling real-time tracking and aiming at athletes to ensure the clarity and continuity of exciting footage. Combined with the previously identified highlights, it can automatically trigger continuous shooting, capturing the best moments and presenting viewers with a superior visual experience. In summary, this AI-based real-time highlight capture solution integrates advanced perception technology, intelligent analysis algorithms, and intelligent control technology, achieving comprehensive perception, accurate assessment, and intelligent capture of athlete status, bringing a completely new viewing experience to sporting events. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating a method for real-time capture of exciting scenes based on artificial intelligence, provided in an embodiment of the present invention.

[0046] Figure 2 This is a flowchart of feature extraction of motion data in a real-time capture method for exciting scenes based on artificial intelligence provided in an embodiment of the present invention;

[0047] Figure 3 This is a flowchart illustrating a method for extracting facial features in real-time capture of exciting scenes based on artificial intelligence, as provided in an embodiment of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0049] The following is combined Figures 1-3 This invention describes a method for real-time capture of exciting scenes based on artificial intelligence.

[0050] Figure 1 This is a schematic diagram of a real-time image capture method based on artificial intelligence provided in an embodiment of the present invention.

[0051] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for real-time capture of exciting scenes based on artificial intelligence, including:

[0052] S1. Collect the athlete's physiological and kinematic data. Physiological data includes the athlete's height, weight, heart rate, and blood pressure. Kinematic data includes the athlete's speed, range of motion, frequency of motion, and joint angles.

[0053] In this embodiment, physiological data such as an athlete's height and weight help to better understand the athlete's physical characteristics and potential. Kinematic data such as an athlete's speed, range of motion, and joint angles can reflect the athlete's movement technique, explosive power, and coordination. By collecting this data, basic data support can be provided for subsequent explosive power assessments and focus analysis. These physiological indicators and kinematic data can more accurately depict the athlete's overall condition and provide a basis for judging key moments.

[0054] S2. Extract features from the motion data to obtain motion features, extract features from the physiological data to obtain physiological features, construct an explosive power evaluation model based on a recurrent neural network, input the motion features and physiological features, and output the athlete's explosive power score.

[0055] Figure 2 This is a flowchart illustrating the feature extraction of motion data in a real-time capture method for capturing exciting scenes based on artificial intelligence, as provided in an embodiment of the present invention. Figure 2 As shown, the process of feature extraction from motion data includes:

[0056] S21. Using a deep learning-based target detection algorithm, locate the contours and coordinates of the athlete's key body parts.

[0057] In this embodiment, a large amount of training data containing athlete movements is collected, and key joints, including the head, shoulders, elbows, wrists, hips, knees, and ankles, are labeled. A deep learning model based on convolutional neural networks is constructed, and an object detection algorithm is trained. In new video frames, the trained model is used to quickly locate the coordinates and contours of the athlete's key parts, and Kalman filtering is used to track the motion trajectory of the key parts in consecutive frames to improve detection stability.

[0058] S22. Construct a human dynamics model based on motion data to assess muscle strength.

[0059] In this embodiment, an anthropometric data of an athlete, including height, weight, and limb length, is used to establish a multi-rigid-body human model. Based on anatomical characteristics such as muscle attachment points and joint range of motion, a musculoskeletal dynamics model is established. The motion trajectory data of key parts are input into the model, and the torques of each joint are solved inversely using the Newton-Euler equations. Combined with the physiological characteristics of muscles, the force output of major muscle groups is estimated.

[0060] S23. Calculate the motion velocity and acceleration of key parts based on the streamer method.

[0061] In this embodiment, the instantaneous velocity and acceleration are calculated using the central difference method by utilizing the position coordinates of key parts in the video. In order to suppress the influence of noise on the differential operation, the coordinate data can be smoothed and filtered first. By comparing the coordinate changes at different time points, the linear velocity and angular velocity of each joint are obtained. Further differentiation can yield the linear acceleration and angular acceleration, providing basic data for the next step of analysis.

[0062] S24. Evaluate the mass distribution of an athlete's limbs based on motion data, and calculate the moment of inertia by combining the moment of inertia.

[0063] In this embodiment, the athlete's body measurement data is used to predict the mass and inertial characteristics of each limb segment using a statistical model. The motion trajectory data of key parts are substituted into the model, and the torque of each joint is calculated according to Newton's second law. The torque is multiplied by the angular acceleration to obtain the inertial torque of each joint.

[0064] The process of constructing a burst power assessment model based on recurrent neural networks includes:

[0065] The motion and physiological data are cleaned and standardized to obtain standardized data.

[0066] Construct a basic recurrent neural network model, and set the number of layers and the number of neurons in each layer.

[0067] Historical data is collected, and the athlete's historical athletic and physiological characteristics are extracted from the historical data. The historical athletic and physiological characteristics are used as inputs, and the explosive power score is used as the output to train a basic recurrent neural network model. The model parameters that meet the test accuracy are retained to obtain an explosive power evaluation model.

[0068] In this embodiment, deep learning modeling of motion and physiological data allows for a better capture of the complex mechanisms underlying an athlete's explosive power. Recurrent neural networks can establish a nonlinear mapping between motion and physiological characteristics and explosive power, resulting in higher prediction accuracy compared to traditional statistical models. Furthermore, this model can continuously learn and update, leading to continuous improvement in the accuracy of explosive power assessment as data accumulates.

[0069] This model can take into account athletes' athletic and physiological data in real time and quickly output explosive power scores. Combined with preset explosive power thresholds, it can promptly detect when an athlete's explosive power reaches a critical state, providing coaches with real-time alerts. This helps coaches adjust training plans in a timely manner, enabling athletes to compete at their best and improve performance. It also provides coaches with objective feedback on athletes' training status; based on the explosive power scores output by the model, coaches can understand the athlete's training progress and potential, and adjust training plans accordingly. During competition, this evaluation model can also provide coaches with real-time physiological monitoring, facilitating more precise pre-competition preparation and in-competition adjustments.

[0070] S3. Collect athletes' facial expression data in real time, extract expression features from the facial expression data, establish a focus assessment index, and obtain athletes' focus scores based on expression features.

[0071] Figure 3 This is a flowchart illustrating the extraction of facial expression features in a real-time capture method for capturing exciting scenes based on artificial intelligence, provided by an embodiment of the present invention. For example... Figure 3 As shown, the process of extracting facial expression features from facial expression data includes:

[0072] The Viola-Jones face detection algorithm is used to detect face regions, and N key feature points are set within the face region using an active shape model.

[0073] Based on the location of key feature points, the face region is divided into facial feature regions, including the eyes, eyebrows, nose, lips, and face area. These key feature points can accurately describe the geometry of the face. This region division facilitates the subsequent targeted extraction of features from each facial region.

[0074] Calculate geometric quantification metrics between key feature points, including distance, angle, and area. These metrics quantify changes in facial geometry, such as eyebrow height, eye opening / closing, and mouth opening.

[0075] Texture analysis was performed on each facial feature region to extract Gabor and LBP texture features. Gabor features reflect the direction and frequency information of facial muscle textures, while LBP features describe the microscopic patterns of texture. These texture features can characterize subtle changes in facial expressions.

[0076] By tracking the motion trajectories of key feature points in consecutive frames and calculating the temporal features of each key feature point, the dynamic features of the face region can be obtained. Through face tracking algorithms, the motion trajectories of key feature points in consecutive frames are tracked, and the temporal dynamic features such as position, velocity, and acceleration of each key feature point are calculated, reflecting the rate and rhythm of facial expression changes.

[0077] By using geometric quantization metrics, Gabor and LBP texture features, and dynamic features, facial expression vectors are constructed and used as facial expression features. This comprehensively depicts the changing states of facial expressions.

[0078] The process of establishing a focus assessment index includes: building a focus assessment model; collecting facial expression data of athletes at different levels of focus and labeling them with corresponding focus score tags; training the focus assessment model and retaining model parameters that meet the preset accuracy; and using facial expression features as input to output focus scores.

[0079] In this embodiment, focus is one of the key indicators for judging highlight moments. Only when athletes are highly focused can they unleash their optimal explosive power and technical skill. The focus assessment model allows for real-time monitoring of athletes' focus levels, providing a reliable basis for identifying highlight moments. Combining focus scores with explosive power and attention scores allows for more accurate capture of truly spectacular moments.

[0080] Meanwhile, the focus assessment model can provide coaches with real-time feedback on athletes' focus levels. Coaches can adjust training plans promptly based on changes in focus scores, helping athletes develop good focus habits. During competitions, coaches can also use focus data to choose appropriate times to deploy athletes, improving the overall execution of tactics.

[0081] When the system detects athletes entering a state of high concentration, it automatically captures these moments to provide viewers with a better viewing experience. These highlights of focused performance can be viewed not only live but also used in broadcasts, significantly enhancing the appeal of the broadcast content. High-quality broadcast content attracts more sponsors, thereby increasing the commercial value of the event.

[0082] S4. Real-time collection of external attention data, including the voice and body data of the audience, feature extraction of the voice and body data, and calculation of the athlete's attention score through pattern recognition algorithm.

[0083] A microphone array is used to collect real-time audio data from the audience, including cheers and applause. Simultaneously, infrared sensors are used to collect data on audience body language, such as clapping and waving. This data reflects the audience's current level of attention and excitement.

[0084] The process of feature extraction from sound data includes:

[0085] The audio data is denoised, and the continuous audio data is segmented to obtain short-frequency audio signals.

[0086] Spectral analysis of short-frequency sound signals yields short-frequency sound characteristics, including volume characteristics, frequency characteristics, and prosodic characteristics.

[0087] A recursive feature elimination method was used to filter short-frequency sound features to obtain relevant sound features.

[0088] The process of feature extraction from limb data includes:

[0089] The limb data is calibrated and filtered, and continuous limb movements are segmented into independent action units, including clapping and waving movements.

[0090] Extract spatial and temporal features from the action unit. Spatial features include the amplitude, direction, and displacement distance of the action, while temporal features include the duration and frequency of the action.

[0091] Spatial and temporal features are standardized to obtain relevant limb features.

[0092] The process of calculating an athlete's attention score using a pattern recognition algorithm includes: quantifying relevant vocal and physical features, including the vocal range of the vocal features and the range of motion of the physical features; assigning a score to each vocal range and each range of motion; and finally, weighted summing the vocal range score and the range of motion score to obtain the attention score.

[0093] In this embodiment, by collecting on-site data in real time, the athlete's attention score can be quickly calculated, providing timely data support for the real-time identification of exciting moments. Combining both auditory and physical data can more comprehensively reflect the audience's attention status. The fusion of multi-source data can improve the accuracy and robustness of attention assessment. Compared with subjective judgment, attention assessment based on data analysis is more objective and fair, helping to eliminate the interference of human factors and improve the accuracy of the entire system. Furthermore, this attention assessment method based on audience behavior data is not only applicable to sporting events but can also be extended to other scenarios such as concerts and lectures.

[0094] S5. Set the explosiveness score threshold, focus score threshold, and attention score threshold. When the explosiveness score exceeds the explosiveness score threshold, the focus score exceeds the focus score threshold, and the attention score exceeds the attention score threshold, it is judged as a highlight moment.

[0095] In this embodiment, the explosive power score threshold represents the minimum required level of explosive power for an athlete's movement. Movements below this threshold are not considered spectacular. The focus score threshold represents the minimum required level of focus for an athlete. Movements below this threshold may indicate a lack of necessary focus. The attention score threshold represents the minimum level of attention the audience pays to the movement. Movements below this threshold indicate that the audience may not be paying attention. A large amount of actual sports video data is collected, covering various sports. Each movement in the video is manually annotated, and explosive power, focus, and attention scores are given. An annotated dataset is constructed to provide a basis for subsequent threshold determination. Statistical analysis is performed on the movements in the annotated dataset, and probability distribution histograms of each score indicator are plotted. Based on the distribution of the histograms, appropriate quantiles are selected as thresholds for each score. For the real-time collected data, the explosive power, focus, and attention scores for each athlete are calculated. When an athlete's three scores simultaneously exceed their respective preset thresholds, the movement is determined to be a spectacular moment.

[0096] S6. Focus on the athlete's face and automatically adjust the shooting angle to capture exciting moments.

[0097] The process of focusing on the athlete's face includes: using phase detection autofocus technology to determine the position of the focal plane by detecting high-frequency signals in the image, and combining face detection and eye detection technologies to achieve focus on the athlete's face.

[0098] In this embodiment, face detection and eye detection technologies are combined: First, a deep learning model, such as MTCNN, is used to detect the athlete's facial position in the image. Then, the eye region is further located, and eye features are used as the focus point to ensure that the athlete's face remains sharp at all times. This focusing method based on face and eye detection can handle scenes with fast-moving athletes and maintain good focus.

[0099] The process of automatically adjusting the shooting angle includes:

[0100] The camera's motion is detected using a three-axis accelerometer and gyroscope sensor, including translation and rotation. Based on the athlete's position and movement in the frame, and combined with computer vision technology, the camera tracks and predicts the athlete's trajectory. A gimbal control algorithm adjusts the gimbal's translation and rotation in real time to ensure the camera remains focused on the athlete. This automatic angle adjustment method, based on multi-sensor fusion, can handle various complex motion scenarios and ensure shooting accuracy.

[0101] Combining image recognition technology, the system tracks and locks onto the athlete's position within the frame, capturing exciting moments. Utilizing target detection and tracking algorithms, it continuously locks onto the athlete's position within the frame. When a moment exceeding the thresholds for explosive power, focus, and attention is detected, continuous shooting is automatically triggered to capture the exciting instant. A high frame rate continuous shooting mode is employed to ensure the best possible footage is captured. Simultaneously, the captured images are intelligently sorted and filtered based on the scores for explosive power, focus, and attention.

[0102] Example 1: Athlete A, gender: male, age: 25, height: 180cm, weight: 75kg.

[0103] The heart rate during high-intensity exercise was 150 bpm, the blood pressure before exercise was 120 / 80 mmHg, and the blood pressure after exercise was 130 / 85 mmHg.

[0104] The maximum speed is 6.2 m / s. During the acceleration phase, the knee bend angle is 90°. When the speed reaches its peak, the movement frequency is 2.2 Hz. The elbow joint angle is 120° and the shoulder angle is 150° when exerting force.

[0105] Key body parts: Knee: coordinates (x=50, y=120), Elbow: coordinates (x=70, y=150), Shoulder: coordinates (x=60, y=160).

[0106] Construct a human dynamics model based on motion data to assess leg muscle strength: 300N, upper limb muscle strength: 200N, and average strength: (300N+200N) / 2=250N.

[0107] The acceleration is estimated to be 3.5 m / s² based on the streamer method.

[0108] Upper body mass: 30 kg (40% of body weight), lower body mass: 45 kg (60% of body weight). The distance from the body's center of gravity to the ground is 0.9 m. The calculated moment of inertia is 36.45.

[0109] Historical data was collected, and the table is as follows:

[0110] Training number Action speed (m / s) Action frequency (Hz) Knee angle (°) Heart rate (bpm) Blood pressure (mmHg) Explosive force score 1 6.0 2.0 90 150 120 / 80 0.85 2 5.5 1.8 95 145 115 / 75 0.80 3 6.2 2.2 85 160 125 / 85 0.90 4 5.9 2.1 88 155 120 / 80 0.88 5 6.1 2.3 87 158 121 / 79 0.91 … … … … … … … 30 6.5 2.4 82 152 122 / 81 0.95

[0111] Motion and physiological features are extracted from historical data and integrated into an input feature vector.

[0112] [0.75,0.60,0.50,0.40,0.48,0.55], Explosiveness Score = 0.95.

[0113] The structure of the basic recurrent neural network model is constructed, with the input layer consisting of 6 feature nodes.

[0114] Hidden layer: 1 LSTM layer containing 64 units, with Tanh activation function.

[0115] Output layer: 1 node, activation function is Sigmoid (output focus score).

[0116] Model hyperparameters are set as follows: learning rate: 0.001; batch size: 32; training epochs: 100; loss function: mean squared error (MSE); optimizer: Adam optimizer.

[0117] The model is trained using historical data, and the model parameters that reach the preset loss function value are retained.

[0118] The Viola-Jones algorithm was used to detect the facial region of athlete A.

[0119] Face region coordinates: top left corner (x=50, y=80), bottom right corner (x=100, y=140).

[0120] N key feature points are defined. Some of the key feature points and their coordinates are as follows:

[0121] Left eye corner: (x=55, y=90), right eye corner: (x=65, y=90), left eyebrow: (x=52, y=85), right eyebrow: (x=68, y=85), tip of nose: (x=60, y=110), left corner of mouth: (x=55, y=130), right corner of mouth: (x=65, y=130).

[0122] Based on the definition of key feature points, the face region is divided into the following feature regions:

[0123] Eye area: Left ((55,90),(60,100)), Right ((60,90),(70,100)).

[0124] Eyebrow area: Left ((52,85),(55,90)), Right ((65,85),(58,90)).

[0125] Nose area: (60,100)-(60,120).

[0126] Lip area: Left mouth ((55,130),(60,135)), Right mouth ((65,130),(60,135)).

[0127] Face region: The main region contains all features.

[0128] Interocular distance: .

[0129] Distance between eyebrows: .

[0130] The angle between the left corner of the eye and the tip of the nose: .

[0131] Area of ​​the left eye region (rectangle): (70−55)×(100−90)=15×10=150cm² 2 .

[0132] Gabor characteristics = [0.3, 0.5, 0.2, 0.7].

[0133] LBP feature = [1,0,1,1,0].

[0134] Motion trajectory data (in 5 consecutive frames)

[0135] Frame 1: Tip of nose (60, 110), left corner of mouth (55, 130).

[0136] Frame 2: Tip of nose (60,112), left corner of mouth (56,132).

[0137] Frame 3: Tip of nose (60,113), left corner of mouth (57,134).

[0138] Frame 4: Tip of nose (60, 114), left corner of mouth (58, 135).

[0139] Frame 5: Tip of nose (60,115), left corner of mouth (59,136).

[0140] Dynamic features of the tip of the nose:

[0141] Movement path: from (60,110) to (60,115).

[0142] Time interval: 1 frame = 0.04 seconds.

[0143] Average velocity (calculated through inter-frame variation): .

[0144] Dynamic features of the left corner of the mouth:

[0145] Movement path: from (55,130) to (59,136).

[0146] Average speed: .

[0147] Nose tip feature vector: dynamic feature = [25, 0.04].

[0148] Left corner of the mouth feature vector: dynamic feature = [50, 0.04].

[0149] All extracted features are integrated to form the final facial expression feature vector. The facial expression vector includes geometric quantization metrics, Gabor and LBP texture features, and dynamic features. Assume the final constructed facial expression feature vector is as follows:

[0150] Geometric quantification metrics (standardized processing of distance, angle, and area):

[0151] Geometric features = [Interocular distance standardization = 2010 = 0.5, Eyebrow distance standardization = 2016 = 0.8, Angle = 9075.96 ≈ 0.844, Left eye area standardization = 300150 = 0.5].

[0152] Gabor characteristics: [0.3,0.5,0.2,0.7][0.3,0.5,0.2,0.7].

[0153] LBP features: [1,0,1,1,0][1,0,1,1,0].

[0154] Dynamic features (considering the tip of the nose and the left corner of the mouth):

[0155] Dynamic characteristics = [nose tip speed = 25, left corner of mouth speed = 50].

[0156] Output of the facial expression vector combining all features:

[0157] Facial feature vector = [0.5, 0.8, 0.844, 0.5, 0.3, 0.5, 0.2, 0.7, 1, 0, 1, 1, 0, 25, 50].

[0158] A single-layer feedforward neural network is used as the attention assessment model. The input layer contains facial expression feature vectors, and the output layer is the attention score (between 0 and 1).

[0159] Model parameters: Input nodes: 15 (from facial expression feature vectors); Hidden layer nodes: 5; Output node: 1 (attention score); Activation functions: ReLU (hidden layer), Sigmoid (output layer); Loss function: Mean squared error.

[0160] Collect facial expression data of athletes at different levels of focus and their corresponding focus score labels. Example data is as follows:

[0161] [0.5,0.8,0.844,0.5,0.3,0.5,0.2,0.7,1,0,1,0,0,25,50],0.85。

[0162] [0.6,0.6,0.7,0.6,0.2,0.4,0.3,0.6,1,0,0,0,0,20,40],0.70。

[0163] [0.5,0.7,0.8,0.4,0.4,0.5,0.5,0.5,0,1,0,1,1,30,55],0.60。

[0164] [0.4,0.5,0.6,0.3,0.1,0.3,0.1,0.2,0,0,1,0,1,15,30],0.45。

[0165] [0.7,0.9,0.9,0.8,0.5,0.6,0.4,0.8,1,1,1,1,0,35,60],0.95。

[0166] [0.35,0.4,0.3,0.2,0.2,0.4,0.4,0.5,0,0,0,0,0,5,10],0.20。

[0167] The attention assessment model was trained using the collected data. Batch gradient descent was employed, and the following steps were followed:

[0168] Data preprocessing:

[0169] Standardized facial expression features: The input data is standardized so that its mean is 0 and its variance is 1.

[0170] The labels (focus score) are normalized to a range of 0-1.

[0171] Training steps:

[0172] Randomly divide the training and test sets, for example, 70% training set and 30% test set.

[0173] Training the model and using cross-validation saves time and resources.

[0174] Convergence condition:

[0175] Training stops when the model’s loss function remains unchanged for 10 consecutive iterations without exceeding a threshold (e.g., 0.001).

[0176] Retain model parameters that meet the preset accuracy (e.g., R²>0.85).

[0177] After completing model training and evaluating model performance using the test set, assume the following results are obtained:

[0178] Mean absolute error (MAE): 0.05.

[0179] Root mean square error (RMSE): 0.07.

[0180] Coefficient of determination (R²): 0.88 (indicating that the model has a good explanatory power for attention).

[0181] Input feature vector: [0.5,0.8,0.844,0.5,0.3,0.5,0.2,0.7,1,0,1,1,0,25,50]

[0182] The input feature vector is standardized to transform it into a distribution with a mean of 0 and a variance of 1.

[0183] The output is calculated using forward propagation of the trained neural network model:

[0184] Standardized features are input into the hidden layer, and the ReLU activation function is applied.

[0185] The output of the hidden layer is passed to the output layer, and the final focus score is obtained through the Sigmoid activation function.

[0186] The focus score obtained from the model is: Focus Score = 0.80.

[0187] Real-time acquisition of external attention data and feature extraction of sound data, including noise reduction, audio segmentation, spectrum analysis, and feature selection.

[0188] The original audio data undergoes noise reduction processing to eliminate background noise and irrelevant audio interference. The waveform of the original audio data is shown below:

[0189] Length: 10 seconds.

[0190] Sampling frequency: 44.1kHz.

[0191] The noise-reduced audio signal is divided into multiple short frequency bands to obtain short frequency sound signals.

[0192] The noise-reduced audio signal is divided into 5 short frequency bands, and the data for each short frequency band is as follows:

[0193] Short band number Audio sample segment (sec) 1 [0.1,0.2,0.3,0.4] 2 [0.5,0.2,0.1,0.3] 3 [0.3,0.4,0.5,0.6] 4 [0.6,0.7,0.4,0.3] 5 [0.7,0.8,0.6,0.5]

[0194] Fast Fourier Transform was used to perform spectral analysis on the audio data of each short frequency band, and volume features, frequency features, and prosodic features were extracted.

[0195] The analysis results for the short frequency band are as follows:

[0196] Short band number Volume feature (dB) Main frequency feature (Hz) Rhythm feature (Hz) 1 75 440 1.5 2 80 450 1.7 3 78 460 1.6 4 85 430 1.8 5 82 410 1.5

[0197] The recursive feature elimination method was used to select the sound features most relevant to the attention score.

[0198] The final output of relevant sound characteristics is as follows:

[0199] Relevant sound features = [volume features, main frequency features] = [75, 440].

[0200] The raw limb data is calibrated and filtered to eliminate noise and unnecessary interference.

[0201] Raw limb data:

[0202] Sampling frequency: 100Hz.

[0203] Sampling time: 10 seconds.

[0204] After calibration and filtering, smoothed limb data is obtained, as follows:

[0205] Time (sec) X-axis acceleration (g) Y-axis acceleration (g) Z-axis acceleration (g) Angular velocity (rad / s) 0 0.02 0.05 -0.02 0.1 0.01 0.03 0.06 -0.01 0.12 0.02 0.01 0.07 0 0.10 … … … … … 9.99 0.02 0.03 -0.01 0.1

[0206] Continuous limb movements are segmented into independent action units, and the start and end of the action are detected using specific thresholds for the movement.

[0207] Two independent action units were identified:

[0208] Clapping gesture:

[0209] Start time: 2 seconds.

[0210] End time: 3 seconds.

[0211] Waving gesture:

[0212] Start time: 5 seconds.

[0213] End time: 7 seconds.

[0214] Calculate the maximum range of motion of the limb. Calculate the main direction of the movement. Calculate the displacement from the starting position to the ending position of the movement. Calculate the duration of the movement. Calculate the frequency of the movement.

[0215] The feature extraction results for clapping and waving actions are as follows:

[0216] Clapping motion (2 to 3 seconds)

[0217] Range: 0.3g.

[0218] Direction: Y-axis.

[0219] Displacement distance: 0.4m.

[0220] Duration: 1 second.

[0221] Frequency: 1 time / second.

[0222] Waving gesture (5 to 7 seconds).

[0223] Range: 0.5g.

[0224] Direction: X-axis.

[0225] Displacement distance: 0.6m.

[0226] Duration: 2 seconds.

[0227] Frequency: 1 time / 2 seconds.

[0228] The Z-score standardization method is used to calculate the mean and standard deviation to standardize the features.

[0229] Feature Clapping (normalized feature) Waving (normalized feature) Amplitude (g) -1.0 1.0 Direction (X / Y) 0 1 Displacement distance (m) -0.5 0.5 Duration (sec) -1.0 1.0 Frequency (times / sec) 1.0 -0.5

[0230] After standardization, the final output of relevant limb features is summarized as follows:

[0231] Action type Amplitude (normalized) Direction (normalized) Displacement distance (normalized) Duration (normalized) Frequency (normalized) Clapping -1.0 0 -0.5 -1.0 1.0 Waving 1.0 1 0.5 1.0 -0.5

[0232] The relevant sound features are quantified and the sound features are divided according to specific sound ranges.

[0233] The sound features are divided into the following intervals:

[0234] Range 1: 0-50dB, indicator score 2.

[0235] Interval 2: 51-75dB, indicator score 4.

[0236] Range 3: 76-90dB, indicator score 6.

[0237] Range 4: 91-110dB, indicator score 8.

[0238] Set the motion characteristics to the following range:

[0239] Range 1 (0-0.2m), index score 2.

[0240] Range 2 (0.21-0.5m), index score 4.

[0241] Range 3 (0.51-0.8m), index score 6.

[0242] Range 4 (0.81-1.0m), index score 8.

[0243] The weights assigned to vocal features and body features are as follows:

[0244] Voice feature weight: 0.6.

[0245] Body feature weight: 0.4.

[0246] Attention score = (voice feature score × voice feature weight) + (body feature score × body feature weight).

[0247] Attention score = (6 × 0.6) + (4 × 0.4) = 3.6 + 1.6 = 5.2.

[0248] The following thresholds are set: explosiveness score threshold: 0.8; focus score threshold: 0.7; focus score threshold: 5.0. Athlete A's current explosiveness score is 0.95, focus score is 0.80, and focus score is 5.2, therefore it is judged as a highlight moment.

[0249] In summary, this embodiment provides a method for real-time capture of exciting moments based on artificial intelligence. By collecting athletes' physiological and kinematic data, including height, weight, heart rate, speed, and range of motion, a comprehensive perception of the athlete's physical and mental state is constructed. This data not only reflects the athlete's basic physical characteristics but also describes their motor abilities and physical condition, laying the foundation for subsequent explosive power assessment. Comprehensive perception data facilitates a more accurate assessment of the athlete's overall condition, providing a reliable basis for identifying exciting moments.

[0250] By utilizing recurrent neural network models and combining motion and physiological data, an intelligent assessment of athletes' explosive power can be achieved. This comprehensive analysis based on kinematic and physiological characteristics can accurately capture explosive power, a key indicator, providing an important basis for identifying highlight moments. Compared to relying solely on subjective judgment, this data-driven explosive power assessment is more objective and accurate, helping to eliminate the influence of human factors.

[0251] By collecting athletes' facial expression data in real time and extracting facial features using computer vision technology, focus scores are assessed. Focus is a crucial psychological factor affecting athlete performance, and this focus assessment based on facial micro-expressions more closely reflects the athletes' actual psychological state. Compared to simple human observation, this focus assessment method is more objective and accurate, providing data support for the rational allocation of athletes.

[0252] Real-time acquisition of auditory and physical data from the audience, coupled with feature extraction and pattern recognition, allows for the assessment of athlete attention scores. The audience's emotional and behavioral responses directly reflect their level of engagement with the athletes' performances; this multimodal data-driven attention analysis provides a more comprehensive and objective picture. This not only facilitates the identification of key moments but also provides higher-quality content for broadcasting events, enhancing the viewing experience.

[0253] Three thresholds are set for explosive power, focus, and attention scores. A moment is only recognized when all three scores meet the threshold simultaneously. This strict logical relationship ensures more reliable moment identification and avoids false positives or false negatives. At the same time, this threshold setting is flexible and can be optimized and adjusted according to the characteristics of different sports, continuously improving identification accuracy.

[0254] Utilizing technologies such as phase detection autofocus and gimbal auto-adjustment, the system tracks and aligns with athletes in real time, ensuring clarity and continuity of exciting footage. This intelligent camera technology, based on multi-sensor fusion, can handle complex scenarios such as fast-moving athletes, guaranteeing high-quality images. Furthermore, by combining previously identified key moments, the system can automatically trigger continuous shooting to capture the best possible instants.

[0255] This AI-based real-time capture solution for exciting moments achieves intelligent and automated processes throughout the entire process, from sensing athlete status, assessing key indicators, identifying key moments, to finally capturing the exciting footage. It not only significantly improves the accuracy and reliability of identifying key moments but also presents viewers with a superior viewing experience. This innovative practice, which deeply integrates cutting-edge technology with the sports industry, will undoubtedly become a major driving force for the development of smart sports, injecting new momentum into the sport.

[0256] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0257] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time capture of exciting scenes based on artificial intelligence, characterized in that, include: S1. Collect the athlete's physiological and kinematic data, including the athlete's height, weight, heart rate, and blood pressure; and the kinematic data including the athlete's speed, range of motion, frequency of motion, and joint angles. S2. Extract features from the motion data to obtain motion features, extract features from the physiological data to obtain physiological features, construct an explosive power evaluation model based on a recurrent neural network, input the motion features and the physiological features, and output the athlete's explosive power score; S3. Collect athletes' facial expression data in real time, extract expression features from the facial expression data, establish a focus assessment index, and obtain athletes' focus score based on the expression features. S4. Real-time collection of external attention data, including the sound data and body data of the audience on site, feature extraction of the sound data and body data, and calculation of the athlete's attention score through a pattern recognition algorithm; The process of feature extraction from the sound data includes: The sound data is subjected to noise reduction processing, and the continuous sound data is segmented to obtain short-frequency sound signals; Spectral analysis is performed on the short-frequency sound signal to obtain short-frequency sound features, which include volume features, frequency features, and prosodic features. The short-frequency sound features are filtered using a recursive feature elimination method to obtain relevant sound features; The process of feature extraction from the limb data includes: The limb data is calibrated and filtered, and continuous limb movements are divided into independent movement units, including clapping and waving movements. Extract spatial and temporal features from the action unit. The spatial features include the amplitude, direction, and displacement distance of the action, and the temporal features include the duration and frequency of the action. The spatial and temporal features are standardized to obtain relevant limb features; The process of calculating an athlete's attention score using a pattern recognition algorithm includes: quantifying the relevant voice features and the relevant body features, including the voice range of the relevant voice features and the movement range of the relevant body features; setting an index score for each voice range and each movement range; and weighted summing the voice range index score and the movement range index score to obtain the attention score. S5. Set the explosive power score threshold, the focus score threshold, and the attention score threshold. When the explosive power score exceeds the explosive power score threshold, the focus score exceeds the focus score threshold, and the attention score exceeds the attention score threshold, it is determined to be a highlight moment. S6. Focus on the athlete's face and automatically adjust the shooting angle to capture exciting moments.

2. The method for real-time capture of exciting scenes based on artificial intelligence according to claim 1, characterized in that, In step S2, the process of feature extraction from the motion data includes: S21. Using a deep learning-based target detection algorithm, locate the contours and coordinates of the athlete's key body parts; S22. Construct a human dynamics model based on the motion data to assess muscle strength; S23. Calculate the motion velocity and acceleration of the key parts based on the streamer method; S24. Evaluate the mass distribution of the athlete's limbs based on the motion data, and calculate the moment of inertia by combining the moment of inertia.

3. The method for real-time capture of exciting scenes based on artificial intelligence according to claim 1, characterized in that, In step S2, the process of constructing a burst power evaluation model based on a recurrent neural network includes: The motion data and the physiological data are cleaned and standardized to obtain standardized data. Construct a basic recurrent neural network model, and set the number of layers and the number of neurons in each layer; Historical data is collected, and the athlete's historical athletic and physiological characteristics are extracted from the historical data. The historical athletic and physiological characteristics are used as inputs, and the explosive power score is used as outputs to train the basic recurrent neural network model. The model parameters that meet the test accuracy are retained to obtain the explosive power evaluation model.

4. The method for real-time capture of exciting scenes based on artificial intelligence according to claim 1, characterized in that, In step S3, the process of extracting facial expression features from the facial expression data includes: The Viola-Jones face detection algorithm is used to detect face regions, and N key feature points are set in the face regions using an active shape model. Based on the location of the key feature points, the face region is divided into facial feature regions, which include the eyes, eyebrows, nose, lips, and face area. Calculate the geometric quantization index between the key feature points, wherein the geometric quantization index includes distance, angle and area; Texture analysis is performed on each of the facial feature regions to extract Gabor and LBP texture features; Track the motion trajectory of the key feature points in consecutive frames, and calculate the temporal features of each key feature point as the dynamic features of the face region; An expression vector is constructed using the geometric quantization index, the Gabor and LBP texture features, and the dynamic features, and the expression vector is used as the expression feature.

5. The method for real-time capture of exciting scenes based on artificial intelligence according to claim 1, characterized in that, In step S3, the process of establishing a focus assessment index includes: establishing a focus assessment model; collecting facial expression data of athletes under different levels of focus and labeling them with corresponding focus score tags; training the focus assessment model and retaining model parameters that meet the preset accuracy; and using the facial expression features as input to output a focus score.

6. The method for real-time capture of exciting scenes based on artificial intelligence according to claim 1, characterized in that, In step S6, the process of focusing on the athlete's face includes: using phase detection autofocus technology, determining the focal plane position by detecting high-frequency signals in the image, and combining face detection and eye detection technologies to achieve focus on the athlete's face.

7. The method for real-time capture of exciting scenes based on artificial intelligence according to claim 1, characterized in that, In step S6, the process of automatically adjusting the shooting angle includes: The motion state of the camera body is detected using a three-axis accelerometer and a gyroscope sensor. Based on the position and movement of the athletes in the footage, the shooting angle is automatically adjusted through a gimbal control algorithm. By combining image recognition technology, the system tracks and locks onto the position of athletes in the frame, capturing exciting moments.

Images