Pickleball event information generation method and system based on ai vision

Abstract

The application relates to the technical field of computer vision, and discloses a pickball match information generation method and system based on AI vision, which comprises the following steps: acquiring video continuous frames of a pickball match; performing physical displacement feature extraction processing according to the video continuous frames to obtain a motion speed distribution; performing time sequence clustering and screening processing according to the motion speed distribution to obtain a hitting time sequence and a video frame sequence; extracting a player position feature sequence and a dynamic action feature according to the video frame sequence, and calculating a field confrontation density; obtaining a rhythm fluctuation curve according to the hitting time sequence, and intercepting a round subsequence; inputting the hitting action type distribution of the round subsequence and the field confrontation density into a pre-trained evaluation model to obtain a matching degree score; and performing boundary optimization processing on the round subsequence according to the matching degree score and the rhythm fluctuation curve to obtain a key match segment. The method can realize precise automatic interception of a high-intensity confrontation round of a match.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and system for generating pickleball event information based on AI vision. Background Technology

[0002] Currently, with the increasing demand for digital analysis of sports, the automated extraction of event information from video stream data has become a core technology for improving competitive level analysis and spectator interaction. Especially in fast-paced competitive ball sports, the dynamic characteristics contained in video data, such as the frequency of ball strikes, player distribution, and the pace of the competition, are key to analyzing the depth of the event.

[0003] In a current technology, a common approach is to use general image recognition algorithms to detect targets and track motion trajectories in match videos, describing the match process by recording the coordinate changes of players or the ball within the frame. This type of approach primarily focuses on extracting physical spatial positions. However, when dealing with fast-paced scenarios like pickleball, which involve numerous close-quarters exchanges at the net, a single coordinate sequence is insufficient to effectively depict the intensity of the match and the shifts in tactical rhythm. Due to a lack of deep correlation between the temporal dimension of shot frequency and the spatial dimension of player distribution density, the system often fails to identify truly high-intensity, competitive segments from the massive amounts of raw video streams, resulting in fragmented and shallow match information.

[0004] Existing technologies present a technical problem: it is difficult to accurately locate high-intensity combat rounds from high-speed dynamic videos. Summary of the Invention

[0005] This invention provides a method and system for generating pickleball tournament information based on AI vision, in order to solve the technical problem in the prior art that it is difficult to accurately locate high-intensity competitive rounds from high-speed dynamic videos.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for generating pickleball event information based on AI vision, comprising: Acquire consecutive video frames of a pickleball match, and perform physical displacement feature extraction processing on the consecutive video frames to obtain the motion velocity distribution; The motion speed distribution is subjected to temporal clustering filtering to obtain the ball hitting time sequence, and the video frame sequence of the ball hitting time sequence within a preset time window is obtained; The video frame sequence is processed by a pre-trained spatiotemporal extraction model to obtain player position feature sequence and dynamic action feature. The spatial aggregation degree of the player position feature sequence and the dynamic action feature is calculated to obtain the field confrontation density. The frequency sliding statistics processing of the hitting time sequence is performed to obtain the rhythm fluctuation curve. The extreme value boundary positioning processing of the rhythm fluctuation curve is performed to obtain the high-intensity round boundary. Based on the high-intensity round boundary, the continuous frames of the video are truncated to obtain the round subsequence. The ball-hitting action type distribution is obtained by performing action recognition processing on the subsequence of the rounds. The ball-hitting action type distribution is then fused with the field confrontation density input into a pre-trained evaluation model for evaluation processing to obtain a matching score. Extract the subsequences of rounds whose matching scores are greater than a preset matching threshold, and perform boundary optimization processing on the subsequences of rounds according to the rhythm fluctuation curve to obtain key event segments.

[0007] Secondly, the present invention provides a pickleball tournament information generation system based on AI vision, comprising: The displacement extraction module is used to acquire continuous video frames of a pickleball match, and to perform physical displacement feature extraction processing on the continuous video frames to obtain the motion velocity distribution. The timing filtering module is used to perform timing clustering filtering on the motion speed distribution to obtain the ball hitting time sequence, and to obtain the video frame sequence of the ball hitting time sequence within a preset time window; The feature extraction and aggregation module is used to perform feature extraction processing on the video frame sequence using a pre-trained spatiotemporal extraction model to obtain player position feature sequences and dynamic action features, and to perform spatial aggregation degree calculation processing on the player position feature sequences and the dynamic action features to obtain the field confrontation density. The fluctuation boundary positioning module is used to perform frequency sliding statistical processing on the hitting time sequence to obtain a rhythm fluctuation curve, perform extreme value boundary positioning processing on the rhythm fluctuation curve to obtain a high-intensity round boundary, and perform a cropping operation on the continuous frames of the video according to the high-intensity round boundary to obtain a round subsequence. The fusion evaluation module is used to perform action recognition processing on the round subsequence to obtain the distribution of hitting action types, and to perform fusion evaluation processing on the distribution of hitting action types and the field confrontation density input into the pre-trained evaluation model to obtain a matching score. The segment generation module is used to extract the round subsequences whose matching score is greater than a preset matching threshold, and to perform boundary optimization processing on the round subsequences according to the rhythm fluctuation curve to obtain key event segments.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention extracts the motion speed distribution by acquiring continuous frames of video from a peak ball match, performs temporal clustering and frequency sliding statistics on the extracted physical displacement features, and then obtains the shot time series and rhythm fluctuation curve. This process abandons the traditional single analysis mode that relies solely on coordinate trajectory extraction, but instead delves into the dramatic changes in the underlying physical motion speed from the continuous video stream, transforming the fleeting high-speed shot into quantifiable temporal pulses and dynamic frequency fluctuations, effectively capturing the inherent rhythm of the match in the time dimension. This significantly improves the sensitivity of shot rhythm capture in high-frequency peak ball match scenarios, laying a solid temporal foundation for accurate positioning in subsequent high-intensity rallies.

[0009] (2) This invention utilizes a pre-trained spatiotemporal extraction model to extract features from video frame sequences, obtaining player position feature sequences and dynamic action features. Based on these features, spatial aggregation is calculated to obtain the field confrontation density. This mechanism introduces a joint deconstruction of the players' gathering posture and dynamic action evolution in key areas such as the net area, breaking the limitation of existing single-target detection algorithms that lack the ability to quantify spatial pressure. It can accurately map the activity of isolated pixels on a two-dimensional plane into field spatial features that reflect the intensity of the match. Thus, it achieves an objective and in-depth quantitative evaluation of the intensity of multi-target spatial confrontation, effectively filling the technical gap in existing technologies that cannot deeply characterize high-intensity net confrontation postures.

[0010] (3) This invention integrates the distribution of hitting action types in the round subsequence with the field confrontation density input evaluation model for scoring, and performs extreme value boundary positioning and boundary optimization processing on the high-scoring round subsequence according to the rhythm fluctuation curve, finally generating key match segments. This integrated evaluation architecture organically combines the action types in the time dimension with the confrontation density in the spatial dimension, replacing the traditional single-dimensional coarse judgment with multi-dimensional composite feature weighting; at the same time, it strictly constrains the start and end boundaries of video slices by using the gradient descent slope of the rhythm fluctuation curve, so that the segment extraction has a rigorous digital basis and logical closed loop. Thus, it realizes the automated and high-precision extraction of core highlight moments, which not only greatly reduces the computing power and time loss of manual video editing or retrieval, but also ensures that the final output match information segments have extremely high integrity and focus in tactical review and viewing experience. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the process for generating pickleball tournament information based on AI vision provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the AI ​​vision-based peak ball event information generation system provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a method for generating pickleball event information based on AI vision. In this invention, the events include, but are not limited to, all pickleball entertainment activities, such as formal competitions, general competitions, or regular entertainment activities. The method of this embodiment includes the following steps: S1, acquire consecutive video frames of the pickleball match, and perform physical displacement feature extraction processing on the consecutive video frames to obtain the motion speed distribution; S2, perform temporal clustering filtering on the motion speed distribution to obtain the ball-hitting time sequence, and obtain the video frame sequence of the ball-hitting time sequence within a preset time window; S3, use a pre-trained spatiotemporal extraction model to perform feature extraction processing on the video frame sequence to obtain player position feature sequence and dynamic action feature, and perform spatial aggregation degree calculation processing on the player position feature sequence and the dynamic action feature to obtain field confrontation density. S4, perform frequency sliding statistical processing on the hitting time sequence to obtain a rhythm fluctuation curve, perform extreme value boundary positioning processing on the rhythm fluctuation curve to obtain a high-intensity round boundary, and perform a truncation operation on the continuous frames of the video according to the high-intensity round boundary to obtain a round subsequence. S5, perform action recognition processing on the round subsequence to obtain the ball hitting action type distribution, and perform fusion evaluation processing on the ball hitting action type distribution and the field confrontation density input pre-trained evaluation model to obtain the matching degree score; S6, extract the round subsequences whose matching score is greater than the preset matching threshold, and perform boundary optimization processing on the round subsequences according to the rhythm fluctuation curve to obtain key event segments.

[0014] In step S1, continuous video frames of the pickleball match are acquired, and physical displacement feature extraction processing is performed on the continuous video frames to obtain the motion speed distribution.

[0015] Specifically, physical displacement feature extraction is performed on the consecutive frames of the video to obtain the motion velocity distribution, including: The pixel displacement vector is calculated by using the Farneback optical flow algorithm on the consecutive frames of the video. Based on preset spatial ratio parameters and preset inter-frame time interval parameters, the pixel displacement vector is subjected to a dimension transformation operation to obtain the motion velocity distribution.

[0016] In one implementation, continuous frames of a pickleball match video are acquired. This embodiment extracts an image sequence from a 1080p resolution raw video stream at a rate of 30 frames per second using a video decoding component. Each frame in the image sequence is then converted to grayscale to obtain a grayscale image sequence. A Gaussian filtering algorithm is used to perform spatial domain convolution smoothing on the grayscale image sequence to remove high-frequency noise components. It should be noted that the Gaussian kernel size used in the Gaussian filtering algorithm is set to 5×5. The method for determining the Gaussian kernel size is as follows: a set of static background frames without moving targets is extracted from the historical video dataset, and different odd-numbered Gaussian kernels of different dimensions are applied for filtering. The peak signal-to-noise ratio (PSNR) of the smoothed image under each Gaussian kernel dimension is calculated, and the smallest odd-numbered kernel dimension that makes the PSNR value exceed 35 dB for the first time is extracted and directly designated as the Gaussian kernel size. The processed images are combined in chronological order to output the continuous video frames.

[0017] In one implementation, the pixel displacement vector is calculated using the Farneback optical flow algorithm on consecutive video frames. A first image and a second image, sequentially adjacent in time, are selected from the consecutive video frames. A local quadratic polynomial expansion algorithm is used to fit polynomial coefficients to the local neighborhoods of corresponding pixels in the two images, generating a first polynomial coefficient matrix and a second polynomial coefficient matrix. Based on the assumption of constant global displacement between the first and second polynomial coefficient matrices, the horizontal and vertical pixel displacement offset values ​​of the corresponding pixels in the image coordinate system are calculated and combined to form the pixel displacement vector. It should be noted that the pyramid scaling factor involved in the Farneback optical flow algorithm is set to 0.5, the number of pyramid layers is set to 3, and the window size is set to 15. It is worth noting that the method for determining the above-mentioned key algorithm parameters is as follows: a sample video sequence containing standard pickle hitting motion is selected as a calibration set, and the motion trajectory of the artificial bounding box is marked on the calibration set; an exhaustive search algorithm is executed in the preset parameter search grid space to calculate the arithmetic mean of the endpoint Euclidean distance error between the output optical flow motion trajectory and the artificially marked trajectory under different parameter combinations; the parameter combination corresponding to the global minimum value of the arithmetic mean is extracted as the calibration value of the above-mentioned key algorithm parameters.

[0018] In one implementation, the pixel displacement vector is subjected to a dimension transformation operation based on a preset spatial ratio parameter and a preset inter-frame time interval parameter to obtain the motion velocity distribution. For each pixel displacement vector in the two-dimensional image plane, its horizontal and vertical displacement components are multiplied by the preset spatial ratio parameter to convert them into absolute physical displacement components. All absolute physical displacement components are divided by the preset inter-frame time interval parameter to obtain the physical velocity components at the corresponding spatial coordinate positions. The physical velocity components at all spatial coordinate positions are encapsulated into a spatial matrix according to the original two-dimensional plane topology to generate a feature tensor characterizing the overall macroscopic motion intensity, thus obtaining the motion velocity distribution.

[0019] It should be noted that the method for determining the preset spatial ratio parameter is as follows: obtaining the physical width constant of the official standard playing field; using an edge detection algorithm to identify the total horizontal pixel span of the playing field corresponding to the physical width constant in the calibration reference frame of the continuous video frames; dividing the physical width constant by the total horizontal pixel span, and determining the calculated quotient as the preset spatial ratio parameter. The method for determining the preset inter-frame time interval parameter is as follows: obtaining the hardware sampling frame rate value of the video acquisition device; calculating the reciprocal relationship between a constant and the hardware sampling frame rate value, and calibrating the calculated reciprocal value as the preset inter-frame time interval parameter.

[0020] For example, this embodiment decodes and extracts a sports video with a frame rate of 30 frames per second. The system uses an edge detection algorithm to identify that a standard field with a physical width of 6.10 meters occupies a total horizontal pixel span of 305 pixels in the image. Dividing 6.10 meters by 305 pixels yields a preset spatial ratio parameter of 0.02 meters per pixel. Simultaneously, dividing a constant by 30 yields a preset inter-frame time interval parameter of approximately 0.0333 seconds. Subsequently, the Farneback optical flow algorithm calculates that the displacement vector of a certain local area between adjacent frames is 10.2 pixels horizontally and -5.8 pixels vertically. In this embodiment, multiplying 10.2 pixels by 0.02 meters per pixel yields an absolute physical displacement component of 0.204 meters, and dividing 0.204 meters by 0.0333 seconds yields a physical velocity component of approximately 6.12 meters per second; similarly, performing an algebraic conversion on the -5.8 pixels yields a physical velocity component of approximately -3.48 meters per second. Finally, the system encapsulates the physical velocity components corresponding to all pixel coordinates into feature tensors and outputs a complete motion velocity distribution; this velocity distribution reflects the intensity of motion in a local area, not the absolute ball speed.

[0021] In step S2, the motion speed distribution is subjected to temporal clustering filtering to obtain the ball-hitting time sequence, and the video frame sequence of the ball-hitting time sequence within a preset time window is obtained.

[0022] The motion speed distribution is subjected to temporal clustering filtering to obtain the ball-hitting time series, including: The motion speed distribution is filtered according to a preset speed threshold to obtain candidate event frames for hitting the ball; The mean filtering algorithm is used to smooth the consecutive candidate hit event frames to obtain a smoothed velocity sequence; The smoothed velocity sequence is grouped using the K-means clustering algorithm to obtain time distribution clusters, and the peak velocity time points in the time distribution clusters are extracted to obtain the ball hitting time sequence.

[0023] In one implementation, the motion speed distribution is filtered according to a preset speed threshold to obtain candidate event frames for hitting the ball. In this embodiment, all physical speed components in the motion speed distribution are traversed, and time frames corresponding to physical speed components whose absolute values ​​are greater than the preset speed threshold are extracted. These time frames are then marked and extracted as candidate event frames for hitting the ball.

[0024] It should be noted that the method for determining the preset speed threshold is as follows: A historical calibration video dataset containing professional peak ball match hitting actions is collected, such as publicly available PPA Tour recordings. 500 standard hitting clips are randomly selected, and the peak speed values ​​at all actual hitting moments are calculated using an optical flow algorithm. The arithmetic mean and overall standard deviation of all peak speed values ​​are calculated. The arithmetic mean is subtracted by twice the overall standard deviation, and the difference is determined as the preset speed threshold. As a preferred embodiment, in the absence of massive historical data for dynamic calibration, the preset speed threshold can be directly set to the empirically recommended value of 5.0 m / s.

[0025] It is worth noting that the mean filtering algorithm is used to smooth the continuous candidate hit event frames to obtain a smoothed velocity sequence. In this embodiment, the candidate hit event frames are arranged in chronological order to construct a time observation window with a length equal to the preset sliding window length. The time observation window is slid sequentially on the time axis according to the single frame step, and the arithmetic mean of the physical velocity components corresponding to all candidate event frames within the time observation window is calculated. The original velocity value of the center frame of the time observation window is replaced with the arithmetic mean, and a smoothed and denoised continuous velocity value sequence is output, which is determined as the smoothed velocity sequence.

[0026] It should be noted that the method for determining the preset sliding window length is as follows: the average physical duration corresponding to the swing and force phase of a regular hitting action in historical match videos is statistically analyzed; this average physical duration is multiplied by the sampling frame rate of the video stream, and the product is rounded up to obtain the preset sliding window length.

[0027] In one implementation, the smoothed velocity sequence is grouped using the K-means clustering algorithm to obtain time distribution clusters, and the peak velocity time points in the time distribution clusters are extracted to obtain the shot time sequence. This embodiment uses time points as a one-dimensional coordinate axis, and data points with physical velocity components in the smoothed velocity sequence are used as one-dimensional clustering samples. The silhouette coefficient method is used to iteratively calculate the average silhouette coefficient corresponding to different candidate cluster numbers within a preset candidate cluster number interval. The candidate cluster number corresponding to the global maximum value of the average silhouette coefficient is extracted as the optimal cluster center number parameter. Based on the optimal cluster center number parameter, the one-dimensional clustering samples are divided into a corresponding number of time interval sets using the K-means clustering algorithm to obtain the time distribution clusters. This embodiment extracts the sample point where the velocity amplitude reaches a local maximum value within each time distribution cluster, records the timestamp corresponding to the sample point, and combines all extracted timestamps to generate the shot time sequence. Simultaneously, with each timestamp in the shot time sequence as the center, an equal number of video frames are symmetrically extracted in both the past and future time directions to obtain the video frame sequence within the preset time window.

[0028] It should be noted that the preset candidate cluster number range is determined based on the statistical upper and lower limits of the number of consecutive hits in a typical round in a historical single game. In this embodiment, it is set to 2 to 15. The preset time window length is set to 21 to 35 frames. This value is chosen because the average physical duration of a single complete hit action in a pickleball, including the backswing, hitting, and follow-through, is approximately 0.7 to 1.1 seconds. At a sampling rate of 30 frames per second, this window can completely cover the action features of a closed loop.

[0029] For example, in this embodiment, a preset speed threshold of 5.0 m / s is calculated based on a historical dataset. The system detects a local speed of 6.8 m / s at a certain time point in the motion speed distribution. Since 6.8 m / s is greater than 5.0 m / s, the system extracts this frame as a candidate event frame for hitting the ball. Subsequently, the system calculates a preset sliding window length of 5 frames based on the average swing duration and performs mean filtering on the speeds of the 5 consecutive candidate event frames to remove local noise fluctuations. Next, the system uses the silhouette coefficient method to traverse and search within the candidate cluster number range of 2 to 15, calculating that the average silhouette coefficient of the current detected segment reaches its maximum value when the number of candidate clusters is 3, thus determining the optimal cluster center number parameter to be 3. The system uses the K-means clustering algorithm to divide the sample points on the one-dimensional time axis into 3 time distribution clusters, and extracts the 152nd, 210th, and 278th frames corresponding to the speed peaks within these 3 time distribution clusters as the hitting time series. Finally, the system reads the preset time window as 21 frames; taking the 152nd frame in the ball-hitting time sequence as the center, the system extracts 10 video frames from the past and 10 frames from the future, and finally obtains a video frame sequence containing 21 consecutive video frames.

[0030] In step S3, the video frame sequence is processed by a pre-trained spatiotemporal extraction model to obtain player position feature sequence and dynamic action feature. The spatial aggregation degree of the player position feature sequence and the dynamic action feature is calculated to obtain the field confrontation density.

[0031] The video frame sequence is processed using a pre-trained spatiotemporal extraction model to obtain player position feature sequences and dynamic action features, including: The video frame sequence is analyzed using the pre-trained spatiotemporal extraction model to extract the key point coordinates of the video frame sequence in a two-dimensional image, thereby obtaining the player position feature sequence. The player position feature sequence is used to perform cross-frame motion state inference calculation using the pre-trained spatiotemporal extraction model to obtain dynamic action features.

[0032] In one implementation, the pre-trained spatiotemporal extraction model is used to perform feature extraction processing on the video frame sequence to obtain player position feature sequences and dynamic action features. The pre-trained spatiotemporal extraction model employs a deep neural network based on a two-stream architecture. Its spatial extraction branch uses a ResNet-50 network with fully connected layers removed as the backbone to extract multi-scale visual features, and cascades a Feature Pyramid Network (FPN). The output feature map of the Feature Pyramid Network is uniformly reduced to 256 dimensions. Its temporal extrapolation branch includes a Long Short-Term Memory (LSTM) network layer that receives the 256-dimensional feature vector input. The hidden layer dimension of the LSM network layer is set to 512, and the number of layers is set to 2 to match the high dynamic temporal feature dimensions of the pickleball motion.

[0033] It should be noted that the training process of the pre-trained spatiotemporal extraction model is as follows: A sample dataset containing a large number of pickleball match video frames and corresponding player physical positions and action type labels is constructed; the sample dataset is input into the spatiotemporal extraction model, using the mean squared error loss between predicted coordinates and true coordinates, and the cross-entropy loss between predicted action categories and true categories as joint optimization objectives; gradient descent and backpropagation operations are performed using the adaptive moment estimation algorithm. The initial learning rate of the model is set to 0.0001, the batch size is set to 32, and the time step is consistent with the number of frames in the preset time window; the network weight parameters are fixed when the loss value of the joint optimization objective on the validation set no longer decreases for 10 consecutive iterations, thus obtaining the pre-trained spatiotemporal extraction model.

[0034] Specifically, the input video frame sequence of the spatiotemporal extraction model is uniformly cropped and scaled to an image size of 224×224 pixels. The player center point coordinates in the player position feature sequence are pixel coordinates (x, y) in a two-dimensional image plane. The probability distribution matrix of the specific action category, which includes action category labels, is predefined as a set containing the following six discrete categories: forehand drive, backhand drive, net volley, overhead smash, serve, and regular transition block. The sample dataset is constructed by annotators using open-source video annotation tools, such as CVAT, to annotate the two-dimensional center point coordinates of the players frame by frame, and assigning action labels of one of the above six discrete categories to continuous action slices.

[0035] It is worth noting that the video frame sequence is analyzed using the pre-trained spatiotemporal extraction model to extract the keypoint coordinates of the video frame sequence in a two-dimensional image, thus obtaining the player position feature sequence. The video frame sequence is then input into the spatial extraction branch of the pre-trained spatiotemporal extraction model to extract a high-dimensional image feature map. A non-maximum suppression algorithm is used to extract multiple local extremum pixel coordinates with response values ​​greater than a preset human confidence threshold from the high-dimensional image feature map. At most four keypoint coordinates corresponding to players participating on the field are extracted within a single frame, and these coordinates are projected back into the original two-dimensional image to obtain a set of center point coordinates for multiple players corresponding to each video frame. All player center point coordinates are arranged according to the chronological order of the video frame sequence, and the player position feature sequence is output.

[0036] In one implementation, the player position feature sequence is subjected to cross-frame motion state deduction calculation using the pre-trained spatiotemporal extraction model to obtain dynamic action features. The first-order displacement difference and second-order velocity difference between adjacent frame coordinates in the player position feature sequence are calculated to construct a motion state feature vector sequence. This motion state feature vector sequence is input into the temporal deduction branch of the pre-trained spatiotemporal extraction model, outputting a probability distribution matrix of specific action categories occurring within the current time window. Simultaneously, an algebraic averaging operation is performed using the first-order displacement difference and the inter-frame time interval to calculate the corresponding average movement speed value. The probability distribution matrix and the average movement speed value are concatenated and encapsulated to determine the dynamic action features. It should be noted that the probability distribution matrix contains vectors of probabilities for various hitting behaviors, thus obtaining the distribution of hitting action types.

[0037] Specifically, the spatial aggregation degree calculation is performed on the player position feature sequence and the dynamic action feature to obtain the field confrontation density, including: Using the DBSCAN clustering algorithm, the core point determination calculation is performed on the player position feature sequence according to the preset clustering radius parameter and the preset minimum sample number parameter to obtain the high-density confrontation area; The field adversarial density is obtained by performing a correlation mapping operation between the dynamic action characteristics and the high-density adversarial region.

[0038] In one implementation, the DBSCAN clustering algorithm is used to calculate the core points of the player position feature sequence based on a preset clustering radius parameter and a preset minimum sample number parameter, thereby obtaining a high-density confrontation region. The coordinates of the two-dimensional center points of multiple consecutive frames in the player position feature sequence are projected and superimposed along the time axis onto the same two-dimensional spatial plane to generate a static point cloud distribution map. A spatial neighborhood is drawn with each key point coordinate in the static point cloud distribution map as the center and the preset clustering radius parameter as the physical radius. The number of other key point coordinates contained within this spatial neighborhood is counted. It is determined whether the number is greater than or equal to the preset minimum sample number parameter; if so, the corresponding key point coordinates are marked as core points. A connectivity operation is performed on all core points whose geometric distance is less than the preset clustering radius parameter to form a spatial cluster. The two-dimensional geometric bounding polygon corresponding to the spatial cluster containing the most core points is extracted, and its boundary range is determined as the high-density confrontation region.

[0039] It should be noted that the method for determining the preset clustering radius parameter is as follows: Obtain the set of actual physical position coordinates of players during close-range net play in historical pickleball match videos; calculate and plot a k-distance graph for the set of physical position coordinates; find the physical spatial distance value corresponding to the significant abrupt inflection point of the distance curve; and directly use this physical spatial distance value as the preset clustering radius parameter. The method for determining the preset minimum sample size parameter is as follows: Combine the maximum number of participants on one side of the court in a pickleball doubles match (e.g., 2 to 4 players) with the number of frames superimposed on a single time window; use the theoretical coordinate point cloud base number generated after projecting multiple frames as a reference; and set the preset minimum sample size parameter to a constant of 5.

[0040] It is worth noting that the field confrontation density is obtained by performing a correlation mapping operation between the dynamic action features and the high-density confrontation area. The geometric area of ​​the high-density confrontation area is calculated using the polygon area calculation formula; the average movement speed value from the dynamic action features and the maximum probability value from the probability distribution matrix are extracted; the maximum probability value is multiplied by the average movement speed value to obtain a dynamic activity index; the dynamic activity index is divided by the sum of the geometric area value and a preset minimum anti-overflow constant, such as 0.0001, to obtain a quantitative evaluation value characterizing the intensity of player movement and the sense of spatial pressure per unit area, which is then determined as the field confrontation density.

[0041] In step S4, frequency sliding statistics are performed on the hitting time sequence to obtain a rhythm fluctuation curve. Extreme value boundary positioning processing is performed on the rhythm fluctuation curve to obtain a high-intensity round boundary. Based on the high-intensity round boundary, the continuous frames of the video are truncated to obtain a round subsequence.

[0042] The frequency sliding statistical processing of the ball-hitting time series yields a rhythm fluctuation curve, including: Calculate the time interval values ​​between adjacent time points in the ball-hitting time sequence, and extract the time segments corresponding to the time interval values ​​that are less than a preset duration threshold; According to the preset window size parameters and preset sliding step parameters, a sliding window frequency statistics operation is performed on the time segment to obtain a discrete frequency sequence; The discrete frequency sequence is smoothed using a Gaussian filtering algorithm to obtain the rhythm fluctuation curve.

[0043] In one implementation, the time interval between adjacent time points in the ball-hitting time sequence is calculated, and time segments corresponding to time intervals less than a preset duration threshold are extracted. In this embodiment, the ball-hitting time sequence is traversed chronologically, and a subtraction operation is performed on two adjacent ball-hitting timestamps to obtain a frame interval. The frame interval is multiplied by a preset inter-frame time interval parameter to obtain the absolute physical time span, which is then determined as the time interval value. The time interval value is compared with the preset duration threshold, and time intervals with values ​​less than the preset duration threshold are extracted and determined as the time segments.

[0044] It should be noted that the method for determining the preset duration threshold is as follows: statistically analyze the consecutive shot intervals marked as high-intensity confrontations in historical professional peakball matches, and draw a frequency distribution histogram of the time intervals; calculate the time span value corresponding to when the cumulative frequency in the frequency distribution histogram reaches 85%, and mark it as the preset duration threshold.

[0045] It is worth noting that, according to preset window size parameters and preset sliding step parameters, a sliding window frequency statistics operation is performed on the time segment to obtain a discrete frequency sequence. In this embodiment, a time observation window with a length equal to the preset window size parameter is constructed, and the time observation window is slid along the time axis using the preset sliding step parameter as the advancing unit. After each slide, the total number of occurrences of the ball-hitting events contained within the current time observation window is counted. This total number of occurrences is divided by the absolute time span corresponding to the preset window size parameter to calculate the ball-hitting frequency scalar for the current window interval. The ball-hitting frequency scalars calculated from all sliding positions are then concatenated according to the chronological order of the sliding time to output the discrete frequency sequence.

[0046] It should be noted that the preset window size parameter is determined by obtaining the number of video frames corresponding to the average duration of a typical single attack and defense round in historical statistics, and calibrating it as the preset window size parameter. The preset sliding step size parameter is determined by dividing the preset window size parameter by a constant six and performing a floor operation to obtain the preset sliding step size parameter.

[0047] In one implementation, a Gaussian filtering algorithm is used to smooth the discrete frequency sequence to obtain a rhythm fluctuation curve. This embodiment constructs a one-dimensional Gaussian kernel function, and performs discrete convolution calculations with the discrete frequency sequence using the one-dimensional Gaussian kernel function to eliminate local high-frequency components in the sequence, outputting a smoothed rhythm fluctuation curve.

[0048] It should be noted that the standard deviation parameter of the one-dimensional Gaussian kernel function is determined by using the Fast Fourier Transform algorithm to perform frequency domain analysis on the global rhythm frequency sequence in the historical calibration dataset, extracting the upper limit value of the main frequency band with an energy accumulation ratio of 90%, and calibrating it as the preset cutoff frequency; then, calculating the variance value corresponding to the preset cutoff frequency in the time domain, and determining the square root of the variance value as the standard deviation parameter.

[0049] For example, in this embodiment, the time points of three consecutive ball-hitting events are extracted as frames 155, 160, and 166, and the frame intervals between adjacent ball-hitting events are calculated to be 5 and 6 frames. Combined with a preset inter-frame time interval parameter, time interval values ​​of 0.17 seconds and 0.20 seconds are obtained. The system reads a calibrated preset duration threshold of 0.30 seconds. Since both of the above two time interval values ​​are less than this threshold, the continuous interval is merged and extracted into a time segment. Subsequently, the system performs sliding statistics on the video sequence with a preset window size parameter of 30 frames (corresponding to an absolute time span of approximately 1.0 second) and a preset sliding step parameter of 5 frames. Within a certain sliding window, the system counts a total of 2 ball-hitting events. Dividing 2 by 1.0 seconds, the scalar ball-hitting frequency of this window is calculated to be 2.0 times / second. A discrete frequency sequence is generated by sliding the records of all scalars over time. Finally, the system applies a Gaussian filter algorithm with a standard deviation parameter of 2.0 to perform convolutional smoothing, outputting a continuous rhythm fluctuation curve.

[0050] Among them, the high-intensity round boundary is obtained by performing extreme value boundary localization processing on the rhythm fluctuation curve, including: The first and second derivatives of the rhythm fluctuation curve are calculated, and the coordinates of the curves where the first derivative is zero and the second derivative meets the preset extreme curvature conditions are extracted to obtain the local maximum points. Extract the local maxima points that are greater than a preset frequency peak threshold as core peak points; The frequency change slope of the curves on both sides is calculated with the core peak point as the center, and the curves on both sides are truncated according to the preset slope change threshold to obtain the high-intensity round boundary.

[0051] In one implementation, the first and second derivatives of the rhythm fluctuation curve are calculated, and curve coordinates points where the first derivative is zero and the second derivative meets a preset extreme curvature condition are extracted to obtain local maxima. This embodiment uses the central difference method to calculate the first and second time partial derivatives of each data point on the rhythm fluctuation curve. All data points are traversed, and data points where the first time partial derivative is equal to a constant zero or where a positive or negative sign crosses are selected as candidate points. The second time partial derivative values ​​corresponding to the candidate points are extracted and compared with the preset extreme curvature condition. The time coordinates corresponding to data points that meet the preset extreme curvature condition are determined as the local maxima.

[0052] It should be noted that the preset extreme curvature condition is that the value of the second-order time partial derivative must be less than a constant zero and its absolute value must be greater than a preset significant curvature constant. The preset significant curvature constant is determined by collecting manually marked real rhythm peak points, calculating the amplitude of the second derivative of all marked peak points, and extracting the lower quartile of the amplitude distribution as the preset significant curvature constant.

[0053] It is worth noting that the local maxima points with frequencies greater than a preset peak value threshold are extracted as core peak points. In this embodiment, the frequency amplitudes corresponding to all extracted local maxima points are compared with the preset peak value threshold. Local maxima points with frequency amplitudes less than or equal to the preset peak value threshold are removed, and points with frequency amplitudes greater than the preset peak value threshold are retained and marked as core peak points.

[0054] It should be noted that the method for determining the preset frequency peak threshold is as follows: calculate the overall mean and overall standard deviation of the discrete frequency sequence of the entire game; add 1.5 times the overall standard deviation to the overall mean, and set the calculated sum as the preset frequency peak threshold.

[0055] In one implementation, the frequency change slope of the curves on both sides is calculated with the core peak point as the center, and the curves on both sides are truncated according to a preset slope threshold to obtain the high-intensity round boundary. In this embodiment, the time coordinate corresponding to the core peak point is used as the reference point, and the rhythm fluctuation curve is traversed frame by frame along the time axis in both the past and future directions. The difference between the frequency value of the current frame and the frequency value of the previous frame is calculated, and then this difference is divided by the inter-frame time interval to obtain the instantaneous frequency change slope. When traversing in the past direction, the time node where the instantaneous frequency change slope is greater than the preset slope threshold is marked as the starting boundary; when traversing in the future direction, the time node where the instantaneous frequency change slope is less than the negative value of the preset slope threshold is marked as the ending boundary. The starting boundary and the ending boundary are combined and extracted to output the high-intensity round boundary, and the continuous video frames are truncated and read according to the high-intensity round boundary to generate the round subsequence.

[0056] It should be noted that the method for determining the preset slope threshold is as follows: statistically analyze the instantaneous frequency change slope sample set of the start and end boundary region of historical high-intensity segments, extract the absolute value of the slope corresponding to the 5 percentile of the cumulative probability distribution of the sample set, and set it as the preset slope threshold.

[0057] For example, in this embodiment, a central difference calculation is performed on the smoothed rhythm fluctuation curve, and the first derivative at frame 230 is detected to be zero, and its second derivative value is -0.03 times / second. 3 The absolute value of the point, 0.03, is greater than the preset significant curvature constant of 0.01, satisfying the preset extreme curvature condition, and is therefore determined to be a local maximum. The system reads the frequency amplitude of this point as 2.1 times / second, and finds that it is greater than the preset frequency peak threshold of 1.5 times / second, thus extracting it as the core peak point. Subsequently, the system calculates the slope by expanding the time window to both sides with frame 230 as the center. When calculating in the past direction, it finds that the instantaneous frequency change slope at frame 218 reaches 0.08 times / second², which is greater than the preset change slope threshold of 0.05 times / second², and marks it as the starting boundary; when calculating in the future direction, it finds that the instantaneous frequency change slope at frame 242 is less than -0.05 times / second², and marks it as the ending boundary. The system finally extracts the time interval from frame 218 to frame 242 to generate a high-intensity round boundary, and extracts the corresponding round subsequence based on this.

[0058] In step S5, the action recognition processing of the round subsequence is performed to obtain the distribution of hitting action types. The distribution of hitting action types is then fused with the field confrontation density input to the pre-trained evaluation model to obtain the matching score.

[0059] The process involves fusing the distribution of the ball-striking motion types with the pre-trained evaluation model inputting the field confrontation density to obtain a matching score, including: The distribution of the ball-hitting action types is normalized to obtain action index scores, and the field confrontation density is normalized to obtain density index scores. Using a pre-trained linear regression model, the action index score and the density index score are weighted and summed according to a preset feature weight distribution to obtain the matching score.

[0060] In one implementation, the round subsequence is processed for action recognition to obtain a distribution of hitting action types. In this embodiment, the round subsequence is input into a pre-constructed action classification network, which outputs the classification confidence score of each frame within the round subsequence belonging to a specific hitting action category. This embodiment extracts the action category corresponding to the global maximum value of the classification confidence score as a single-frame action label; the total frequency of each action category in the round subsequence is counted in chronological order, and the total frequency is divided by the total number of video frames in the round subsequence to generate a frequency distribution sequence of each action category occurring within the corresponding time window. This outputs a vector containing the probabilities of multiple hitting behaviors, thus obtaining the hitting action type distribution.

[0061] It should be noted that the pre-constructed action classification network is a temporal behavior recognition network containing three-dimensional convolutional layers. Its training process is as follows: a large number of video clips containing human action labels are acquired to construct a training sample set; the training sample set is input into the network, with the objective function being to minimize the cross-entropy loss between the predicted classification category and the real human action label; the network weight parameters are iteratively updated using an adaptive moment estimation algorithm, where the initial learning rate is set to 0.001 and the batch size is set to 16; training stops when the classification accuracy on the independent validation set is stable at over 90% for 5 consecutive iterations.

[0062] It is worth noting that the distribution of the hitting action types is normalized to obtain the action index score, and the field confrontation density is normalized to obtain the density index score. Since a time span corresponding to a round subsequence usually contains multiple local time windows corresponding to the hitting time sequences, this embodiment extracts the field confrontation density corresponding to all local time windows contained in the round subsequence, calculates the maximum value of all field confrontation densities, and determines the maximum value as the field confrontation density used to evaluate the global state of that round.

[0063] Subsequently, this embodiment extracts the occurrence frequency of high-intensity confrontation categories from the distribution of hitting action types, including net volleys and overhead smashes. The occurrence frequencies of these high-intensity confrontation categories are summed to obtain feature count values ​​characterizing the intensity of the technical actions. A preset historical maximum and minimum action frequency are obtained. Using a maximum-minimum normalization algorithm, the feature count value is subtracted from the historical minimum action frequency to obtain a first difference; the historical maximum action frequency is subtracted from the historical minimum action frequency to obtain a second difference; the first difference is divided by the second difference to obtain a dimensionless value mapped to the interval between constant zero and constant one, which is determined as the action index score. Similarly, the determined field confrontation density is extracted, and a preset historical maximum and minimum density are obtained. The same maximum-minimum normalization algorithm is used to perform algebraic operations, and the calculated result is determined as the density index score.

[0064] It should be noted that the aforementioned preset historical maximum and minimum values ​​of action frequency and preset historical maximum and minimum values ​​of density are determined based on the following: extracting a high-intensity segment dataset of 100 historical matches of the same scale, calculating the action frequency and confrontation density set of all historical segments; using box plot statistics to calculate the first and third quartiles of the above dataset, and obtaining the interquartile range; subtracting 1.5 times the interquartile range from the first quartile as the lower limit, and adding 1.5 times the interquartile range to the third quartile as the upper limit, removing outliers exceeding these upper and lower limits; extracting the maximum and minimum absolute values ​​from the remaining valid data samples, and labeling them as the aforementioned preset historical maximum and preset historical minimum values, respectively.

[0065] In one implementation, a pre-trained linear regression model is used to calculate the matching score by weighting and summing the action index score and the density index score according to a preset feature weight distribution. This embodiment extracts the first and second network regression coefficients fixed in the pre-trained linear regression model. The first network regression coefficient is determined as the action index weight, and the second network regression coefficient is determined as the density index weight. These weights together constitute the preset feature weight distribution. This embodiment performs a multiplication operation on the action index score and the action index weight to generate an action-weighted value; it also performs a multiplication operation on the density index score and the density index weight to generate a density-weighted value; and it performs an addition calculation on the action-weighted value and the density-weighted value to output a continuous scalar value, which is determined as the matching score.

[0066] It should be noted that the construction and training process of the pre-trained linear regression model is as follows: Action index scores and density index scores corresponding to historical round sub-sequences are extracted as a two-dimensional independent variable sample set. This two-dimensional independent variable sample set consists of 500 historical match videos. The corresponding one-dimensional dependent variable labels are obtained by professional referees who score the intensity of the competition on a percentage scale and then normalize the scores. The objective function is to minimize the mean square error between the predicted value of the linear combination of the two-dimensional independent variables and the one-dimensional dependent variable labels. Ordinary least squares is used to solve the parameters algebraically. After completing the numerical solution and fitting, the regression coefficient parameters are solidified and extracted, thus obtaining the pre-trained linear regression model.

[0067] For example, this embodiment detects and extracts a round subsequence containing 24 frames. The action classification network identifies a total of 5 hitting actions, of which 3 are classified as high-intensity confrontation actions such as net volleys or overhead smashes. The feature count value representing the intensity of the technical action is calculated to be 0.125 times / frame, which is 3 times divided by 24 frames. Since this round subsequence contains two local time windows corresponding to the hitting time points, the system extracts the field confrontation density calculated in the previous steps for these two windows, which are 0.58 and 0.62 respectively, and extracts the maximum value of 0.62 as the field confrontation density of this round subsequence. The system reads the historical maximum value of the action frequency after denoising by box plot statistics, which is 0.15 times / frame, and the minimum value is 0.02 times / frame. The system subtracts 0.02 from 0.125 to get the first difference of 0.105, subtracts 0.02 from 0.15 to get the second difference of 0.13, and divides 0.105 by 0.13 to calculate a motion index score of approximately 0.80.

[0068] In step S6, the round subsequences with matching scores greater than a preset matching threshold are extracted, and the round subsequences are subjected to boundary optimization processing according to the rhythm fluctuation curve to obtain key event segments.

[0069] Specifically, the key event segments are obtained by performing boundary optimization processing on the round subsequence based on the rhythm fluctuation curve, including: The round subsequence is divided into intervals according to a preset frame ratio parameter to obtain the tail time interval; The frequency drop slope of the rhythm fluctuation curve corresponding to the tail time interval is obtained, and the video frames in the round subsequence are removed according to the frequency drop slope to obtain the optimized subsequence. The optimized subsequence is then subjected to video encoding and data encapsulation operations to obtain key event segments.

[0070] In one implementation, the subsequences of rounds with matching scores greater than a preset matching threshold are extracted. In this embodiment, the matching scores of all subsequences to be filtered are compared with the preset matching threshold, and sequences with matching scores less than or equal to the preset matching threshold are removed, while sequences with matching scores greater than the preset matching threshold are retained.

[0071] It should be noted that the method for determining the preset matching degree threshold is as follows: collect real high-intensity round segments and regular transition segments marked by manual in historical pickleball events, calculate the matching degree scores of all historical segments as a sample set; use the receiver operating characteristic curve analysis algorithm to statistically analyze the sample set, calculate the Youden index corresponding to the curve; extract the critical matching degree score corresponding to the maximum value of the Youden index, and mark it as the preset matching degree threshold.

[0072] It is worth noting that the round subsequence is divided into intervals according to a preset frame rate ratio parameter to obtain the tail time interval. In this embodiment, the total number of video frames contained in the round subsequence is obtained, and the total number of video frames is multiplied by the preset frame rate ratio parameter and rounded down to obtain a tail frame constant. Starting from the last frame of the round subsequence, a continuous time range equal to the tail frame constant is extracted in reverse order along the time axis, and this range is used to construct the tail time interval.

[0073] It should be noted that the preset frame rate ratio parameter is determined by statistically analyzing the average number of buffered frames from the completion of the last valid shot in historical high-intensity match segments to the player fully returning to a static ready position; the average number of buffered frames is divided by the average total number of video frames in historical high-intensity match segments, and the calculated quotient is determined as the preset frame rate ratio parameter.

[0074] In one implementation, the frequency decline slope of the rhythm fluctuation curve corresponding to the tail time interval is obtained, and video frames in the round subsequence are removed according to the frequency decline slope to obtain an optimized subsequence. This embodiment extracts a one-dimensional discrete frequency data point set contained in the rhythm fluctuation curve within the tail time interval; the one-dimensional discrete frequency data point set is linearly fitted using the least squares method, and the negative correlation slope value of the fitted line is extracted as the frequency decline slope. This embodiment determines whether the absolute value of the frequency decline slope is greater than a preset attenuation threshold; if so, it is determined that there is obvious motion attenuation at the tail of the current time segment. At this time, a reverse traversal is performed starting from the last frame of the round subsequence, and video frames with instantaneous frequency values ​​lower than a preset silent frequency constant on the corresponding rhythm fluctuation curve are removed until a video frame with an instantaneous frequency value greater than or equal to the preset silent frequency constant is detected. The remaining sequence is output as the optimized subsequence; if not, all video frames are retained, and the original round subsequence is determined as the optimized subsequence.

[0075] It should be noted that the preset attenuation judgment threshold is determined by extracting a slope sample set of historical normal continuous motion attenuation phases, calculating the arithmetic mean of the slope sample set, and calibrating it as the preset attenuation judgment threshold. The preset silent frequency constant is determined by statistically analyzing the upper limit of the 95% confidence interval of the frequency background noise generated by random disturbances in non-match states, such as receiving the ball or changing sides.

[0076] It is worth noting that video encoding and data encapsulation operations are performed on the optimized subsequence to obtain key event segments. In this embodiment, the continuous image matrix stream contained in the optimized subsequence is input into a video encoder, and inter-frame compression and image spatial encoding operations are performed using a preset video compression encoding standard to generate a visual media stream. Simultaneously, the start and end timestamps, the distribution of hitting action types, and the corresponding matching scores of the optimized subsequence in the original video stream are extracted and serialized into a structured metadata stream. In this embodiment, the visual media stream and the structured metadata stream are interleaved and multiplexed, and then uniformly encapsulated and written into a standard media container file to generate complete digital assets of the event segment, i.e., the key event segment.

[0077] For example, in this embodiment, a matching score of 0.78 is calculated for a subsequence of 24 frames. The preset matching threshold calculated using historical subject operating characteristic curve calibration is 0.70. The system compares this and finds that 0.78 is greater than 0.70, thus extracting this subsequence. Subsequently, the system reads the preset frame ratio parameter of 0.15, multiplies 24 frames by 0.15 and rounds down to obtain a 3-frame tail constant. The system then extracts 3 frames from the tail as the tail time interval. The system calculates the absolute value of the frequency decrease slope obtained from the rhythm fluctuation curve fitting within this tail time interval, finding it to be 0.05 times / second², greater than the preset attenuation threshold of 0.02 times / second², and the instantaneous frequency values ​​corresponding to the last two frames have dropped below the preset silent frequency constant. The system then removes these last two video frames, retaining the preceding 22 frames as the optimized subsequence. Finally, the system compresses and encodes the 22-frame sequence using the H.264 standard, and then merges and packages it with metadata such as start time and score to generate key event segment files.

[0078] In summary, this invention accurately extracts the hitting time sequence reflecting the intensity of underlying motion by combining optical flow physical displacement analysis of continuous video frames with temporal clustering screening. It utilizes a pre-trained spatiotemporal extraction model to analyze spatial dimension coordinates and deduce cross-frame action features, and combines time axis projection and DBSCAN clustering algorithms to transform discrete micro-motion states into quantified field confrontation density. It constructs a rhythm fluctuation curve based on frequency sliding statistics, and uses the extreme values ​​of the first and second derivatives and curvature conditions to strictly locate the boundaries of high-intensity rallies. Finally, it inputs the distribution of hitting action types and field confrontation density into a linear regression model for multi-dimensional feature weighted fusion evaluation, and performs frame-level elimination optimization on high-scoring rallies based on the frequency decrease slope of the tail time interval. This invention abandons the crude and shallow analysis mode of the existing technology that relies solely on two-dimensional isolated coordinate trajectory tracking, and successfully leaps to a deep match analysis architecture jointly driven by multiple modalities of hitting rhythm in the time dimension, pressure density in the spatial dimension, and action characteristics. It realizes objective quantitative scoring, precise time sequence delimitation, and fully automated interception of core high-intensity confrontation rounds in complex dynamic scenarios of rapid transitions and high-frequency clashes in pickleball, eliminating subjective bias caused by human experience intervention and the computational power consumption of massive blind retrieval.

[0079] Reference Figure 2 The second embodiment of the present invention provides a pickleball tournament information generation system based on AI vision, comprising: The displacement extraction module is used to acquire continuous video frames of a pickleball match, and to perform physical displacement feature extraction processing on the continuous video frames to obtain the motion velocity distribution. The timing filtering module is used to perform timing clustering filtering on the motion speed distribution to obtain the ball hitting time sequence, and to obtain the video frame sequence of the ball hitting time sequence within a preset time window; The feature extraction and aggregation module is used to perform feature extraction processing on the video frame sequence using a pre-trained spatiotemporal extraction model to obtain player position feature sequences and dynamic action features, and to perform spatial aggregation degree calculation processing on the player position feature sequences and the dynamic action features to obtain the field confrontation density. The fluctuation boundary positioning module is used to perform frequency sliding statistical processing on the hitting time sequence to obtain a rhythm fluctuation curve, perform extreme value boundary positioning processing on the rhythm fluctuation curve to obtain a high-intensity round boundary, and perform a cropping operation on the continuous frames of the video according to the high-intensity round boundary to obtain a round subsequence. The fusion evaluation module is used to perform action recognition processing on the round subsequence to obtain the distribution of hitting action types, and to perform fusion evaluation processing on the distribution of hitting action types and the field confrontation density input into the pre-trained evaluation model to obtain a matching score. The segment generation module is used to extract the round subsequences whose matching score is greater than a preset matching threshold, and to perform boundary optimization processing on the round subsequences according to the rhythm fluctuation curve to obtain key event segments.

[0080] It should be noted that the AI ​​vision-based peak ball event information generation system provided in this embodiment of the invention is used to execute all the process steps of the AI ​​vision-based peak ball event information generation method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0081] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0082] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for generating pickleball tournament information based on AI vision, characterized in that, include: Acquire consecutive video frames of a pickleball match, and perform physical displacement feature extraction processing on the consecutive video frames to obtain the motion velocity distribution; The motion speed distribution is subjected to temporal clustering filtering to obtain the ball hitting time sequence, and the video frame sequence of the ball hitting time sequence within a preset time window is obtained; The video frame sequence is processed by a pre-trained spatiotemporal extraction model to obtain player position feature sequence and dynamic action feature. The spatial aggregation degree of the player position feature sequence and the dynamic action feature is calculated to obtain the field confrontation density. The frequency sliding statistics processing of the hitting time sequence is performed to obtain the rhythm fluctuation curve. The extreme value boundary positioning processing of the rhythm fluctuation curve is performed to obtain the high-intensity round boundary. Based on the high-intensity round boundary, the continuous frames of the video are truncated to obtain the round subsequence. The ball-hitting action type distribution is obtained by performing action recognition processing on the subsequence of the rounds. The ball-hitting action type distribution is then fused with the field confrontation density input into a pre-trained evaluation model for evaluation processing to obtain a matching score. Extract the subsequences of rounds whose matching scores are greater than a preset matching threshold, and perform boundary optimization processing on the subsequences of rounds according to the rhythm fluctuation curve to obtain key event segments.

2. The method for generating pickleball tournament information based on AI vision according to claim 1, characterized in that, The step of extracting physical displacement features from the consecutive video frames to obtain the motion velocity distribution includes: The pixel displacement vector is calculated by using the Farneback optical flow algorithm on the consecutive frames of the video. Based on preset spatial ratio parameters and preset inter-frame time interval parameters, the pixel displacement vector is subjected to a dimension transformation operation to obtain the motion velocity distribution.

3. The method for generating pickleball tournament information based on AI vision according to claim 1, characterized in that, The step of performing time-series clustering and filtering on the motion speed distribution to obtain the ball-hitting time series includes: The motion speed distribution is filtered according to a preset speed threshold to obtain candidate event frames for hitting the ball; The mean filtering algorithm is used to smooth the consecutive candidate hit event frames to obtain a smoothed velocity sequence; The smoothed velocity sequence is grouped using the K-means clustering algorithm to obtain time distribution clusters, and the peak velocity time points in the time distribution clusters are extracted to obtain the ball hitting time sequence.

4. The method for generating pickleball tournament information based on AI vision according to claim 1, characterized in that, The process of using a pre-trained spatiotemporal extraction model to extract features from the video frame sequence to obtain player position feature sequences and dynamic action features includes: The video frame sequence is analyzed using the pre-trained spatiotemporal extraction model to extract the key point coordinates of the video frame sequence in a two-dimensional image, thereby obtaining the player position feature sequence. The player position feature sequence is used to perform cross-frame motion state inference calculation using the pre-trained spatiotemporal extraction model to obtain dynamic action features.

5. The method for generating pickleball tournament information based on AI vision according to claim 1, characterized in that, The step of calculating the spatial aggregation degree of the player position feature sequence and the dynamic action feature to obtain the field confrontation density includes: Using the DBSCAN clustering algorithm, the core point determination calculation is performed on the player position feature sequence according to the preset clustering radius parameter and the preset minimum sample number parameter to obtain the high-density confrontation area; The field adversarial density is obtained by performing a correlation mapping operation between the dynamic action characteristics and the high-density adversarial region.

6. The method for generating pickleball tournament information based on AI vision according to claim 1, characterized in that, The frequency sliding statistical processing of the ball-hitting time series to obtain the rhythm fluctuation curve includes: Calculate the time interval values ​​between adjacent time points in the ball-hitting time sequence, and extract the time segments corresponding to the time interval values ​​that are less than a preset duration threshold; According to the preset window size parameters and preset sliding step parameters, a sliding window frequency statistics operation is performed on the time segment to obtain a discrete frequency sequence; The discrete frequency sequence is smoothed using a Gaussian filtering algorithm to obtain the rhythm fluctuation curve.

7. The method for generating pickleball tournament information based on AI vision according to claim 1, characterized in that, The process of locating the extreme value boundary of the rhythm fluctuation curve to obtain the high-intensity round boundary includes: The first and second derivatives of the rhythm fluctuation curve are calculated, and the coordinates of the curves where the first derivative is zero and the second derivative meets the preset extreme curvature conditions are extracted to obtain the local maximum points. Extract the local maxima points that are greater than a preset frequency peak threshold as core peak points; The frequency change slope of the curves on both sides is calculated with the core peak point as the center, and the curves on both sides are truncated according to the preset slope change threshold to obtain the high-intensity round boundary.

8. The method for generating pickleball tournament information based on AI vision according to claim 1, characterized in that, The process of fusing the distribution of the ball-striking action types with the field confrontation density input into a pre-trained evaluation model to obtain a matching score includes: The distribution of the ball-hitting action types is normalized to obtain action index scores, and the field confrontation density is normalized to obtain density index scores. Using a pre-trained linear regression model, the action index score and the density index score are weighted and summed according to a preset feature weight distribution to obtain the matching score.

9. The method for generating pickleball tournament information based on AI vision according to claim 1, characterized in that, The step of performing boundary optimization processing on the round subsequence based on the rhythm fluctuation curve to obtain key event segments includes: The round subsequence is divided into intervals according to a preset frame ratio parameter to obtain the tail time interval; The frequency drop slope of the rhythm fluctuation curve corresponding to the tail time interval is obtained, and the video frames in the round subsequence are removed according to the frequency drop slope to obtain the optimized subsequence. The optimized subsequence is then subjected to video encoding and data encapsulation operations to obtain key event segments.

10. A pickleball tournament information generation system based on AI vision, characterized in that, include: The displacement extraction module is used to acquire continuous video frames of a pickleball match, perform physical displacement feature extraction processing on the continuous video frames, and obtain the motion velocity distribution. The timing filtering module is used to perform timing clustering filtering on the motion speed distribution to obtain the ball hitting time sequence, and to obtain the video frame sequence of the ball hitting time sequence within a preset time window; The feature extraction and aggregation module is used to perform feature extraction processing on the video frame sequence using a pre-trained spatiotemporal extraction model to obtain player position feature sequences and dynamic action features, and to perform spatial aggregation degree calculation processing on the player position feature sequences and the dynamic action features to obtain the field confrontation density. The fluctuation boundary positioning module is used to perform frequency sliding statistical processing on the hitting time sequence to obtain a rhythm fluctuation curve, perform extreme value boundary positioning processing on the rhythm fluctuation curve to obtain a high-intensity round boundary, and perform a cropping operation on the continuous frames of the video according to the high-intensity round boundary to obtain a round subsequence. The fusion evaluation module is used to perform action recognition processing on the round subsequence to obtain the distribution of hitting action types, and to perform fusion evaluation processing on the distribution of hitting action types and the field confrontation density input into the pre-trained evaluation model to obtain a matching score. The segment generation module is used to extract the round subsequences whose matching score is greater than a preset matching threshold, and to perform boundary optimization processing on the round subsequences according to the rhythm fluctuation curve to obtain key event segments.

Images