Passing interception behavior detection method and system based on line-of-sight focus

Abstract

The application discloses a dribbling tackle behavior detection method and system based on a line-of-sight focus, and belongs to the technical field of motion behavior detection. The application takes a preprocessed competition video image sequence as input, pre-judges a competition area where a wonderful tackle may occur, and detects a tackle event in the competition area after determining that the line-of-sight of multiple persons is focused on the competition area. The application can greatly reduce the workload of an event detection algorithm, save computing power resources, improve time efficiency, thus reduce the high performance requirement of a terminal device, and is more easily popularized to general use scenarios. The application solves the problems of low dribbling tackle behavior detection accuracy and poor recognition effect in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motion behavior detection, in particular to a dribble steal behavior detection method and system based on a line of sight focal point. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] Sports video highlight detection is a technology that detects and filters out highlights in a competition video, which can help users browse the highlights of a normal competition in just a few minutes.

[0004] Existing sports video highlight detection is mainly based on image stream and audio stream information in a competition video, analyzes the image stream information to realize lens boundary detection, lens classification, player behavior recognition, and content recognition of text information such as video text titles; for the audio stream, effective information is extracted through voice recognition of predefined keywords, whistle detection, etc. to assist in the detection of sports events; at the same time, a predefined event model is established using sports field knowledge to realize event detection of highlights.

[0005] However, competition videos usually have a high frame rate, and directly detecting and analyzing sports events from the image stream requires high time and computing costs, so the scope of application is relatively narrow and it is difficult to promote; and because the competition site is usually too noisy, audio recognition is usually difficult, low in accuracy, and poor in recognition effect, and the contribution value to sports event detection is actually not high. SUMMARY

[0006] In order to solve the problems of the prior art, for the visual scene of basketball, the present application provides a dribble steal behavior detection method and system based on a line of sight focal point, which combines the poses of multiple people and the line of sight focal point, estimates the line of sight direction of the players on the court by analyzing the head posture and facial orientation of the players on the court, calculates the court area where the lines of sight of multiple people are focused at the same time, and determines whether a steal behavior has occurred by combining the real-time motion posture of the players in the focused area.

[0007] In a first aspect, the present application provides a dribble steal behavior detection method based on a line of sight focal point.

[0008] The dribble steal behavior detection method based on a line of sight focal point comprises:

[0009] Obtain a competition video and preprocess it to obtain a sequence of competition video images;

[0010] The game video image sequence is input into a preset multi-target tracking detector for processing to obtain a multi-person visual focus area and a multi-person visual focus time;

[0011] According to the multi-person visual focus area and the multi-person visual focus time, a frame cutting and a segment intercepting are performed on the game video to obtain a visual focus area video segment;

[0012] The visual focus area video segment is input into a trained group motion event classification model for processing to determine whether a dribble tackle behavior occurs and output a determination result.

[0013] Further, the game video image sequence is input into a preset multi-target tracking detector for processing to obtain a multi-person visual focus area and a multi-person visual focus time, and the method comprises the following steps:

[0014] The game video image sequence is input into a multi-target detector to obtain head images of each player and corresponding position coordinate information;

[0015] According to the head images and the corresponding position coordinate information of the head images, head posture estimation is performed on the players to obtain real-time gaze areas of each player;

[0016] According to the real-time gaze areas of each player, a multi-person visual focus area and a multi-person visual focus time are obtained.

[0017] Preferably, the game video image sequence is input into a multi-target detector to obtain head images of each player and corresponding position coordinate information, and the method comprises the following steps:

[0018] The game video image sequence is input into a body region detector to obtain body detection boxes of each player, and the body detection boxes are associated with identity IDs of the players;

[0019] The body detection boxes with the identity IDs are input into a multi-target tracker to obtain body region images with the identity IDs and real-time motion coordinates of each player with the identity IDs;

[0020] The body region images with the identity IDs are input into a head detector to obtain head images with the identity IDs;

[0021] According to the real-time motion coordinates of each player with the identity IDs, position coordinate information of the head images relative to the bodies is obtained.

[0022] Preferably, according to the head images and the corresponding position coordinate information of the head images, head posture estimation is performed on the players to obtain real-time gaze areas of each player, and the method comprises the following steps:

[0023] The head images are input into a head angle classifier to obtain head orientation angle information of each player;

[0024] According to the head orientation angle information of each athlete and the position information of the head image, three-dimensional position coordinates of the head center point of each athlete are obtained;

[0025] According to the head orientation angle information of each athlete and the three-dimensional position coordinates of the head center point, a real-time gaze area of each athlete is obtained.

[0026] Further preferably, the three-dimensional position coordinates of the head center point of each athlete are obtained according to the head orientation angle information of each athlete and the position information of the head image, comprising:

[0027] The head image is thresholded to remove background pixels, obtain a head binary image, and perform connected component analysis on the head binary image;

[0028] The center of the largest connected component is calculated and taken as the head center point of the athlete;

[0029] According to the position information of the head image, the head center point coordinates are restored to the corresponding competition video image sequence, and are mapped to the three-dimensional space of the court through coordinate transformation to obtain the three-dimensional position coordinates of the head center point.

[0030] Further, the gaze focus area video segment is input into the trained group sports event classification model for processing, comprising:

[0031] The gaze focus area video segment and the body detection frame are input into the trained group sports event classification model to extract the body skeleton of each athlete;

[0032] According to the body skeleton of each athlete, it is judged whether a dribbling tackle behavior occurs.

[0033] Further, before the competition video image sequence is input into the preset multi-target tracking detector for processing, further comprising:

[0034] According to the actual coordinate information three-dimensional geometric position of the court marker line and the basket, and the mutual relationship of the corresponding key points of the court marker line and the basket in the competition video image sequence, camera calibration and court three-dimensional space reconstruction are performed.

[0035] In a second aspect, the present application provides a dribbling tackle behavior detection system based on gaze focus;

[0036] The dribbling tackle behavior detection system based on gaze focus comprises:

[0037] A first acquisition module is configured to acquire competition video and perform preprocessing to obtain a competition video image sequence;

[0038] The second acquisition module is configured to input the competition video image sequence into a preset multi-target tracking detector for processing, acquire a multi-person visual focus area and a multi-person visual focus time, and perform frame cutting and segment interception on the competition video according to the multi-person visual focus area and the multi-person visual focus time to acquire a visual focus area video segment.

[0039] The dribble steal behavior detection module is configured to input the visual focus area video segment into the trained group motion event classification model for processing to determine whether a dribble steal behavior occurs and output a determination result.

[0040] In a third aspect, the present application provides an electronic device;

[0041] An electronic device includes a memory and a processor, and computer instructions stored in the memory and running on the processor, when the computer instructions are run by the processor, the steps of the dribble steal behavior detection method based on a visual focus point are completed.

[0042] In a fourth aspect, the present application provides a computer readable storage medium;

[0043] A computer readable storage medium is used to store computer instructions, when the computer instructions are executed by a processor, the steps of the dribble steal behavior detection method based on a visual focus point are completed.

[0044] Compared with the prior art, the present application has the following beneficial effects:

[0045] The technical solution provided by the present application pre-judges the possible wonderful steal area of the competition field, and only detects the steal event in the focused competition field area, which can greatly reduce the workload of the event detection algorithm, save the computing power resources, improve the time efficiency, thus reduce the high performance requirement of the terminal device, and is more easy to popularize to the general use scene.

[0046] The dribble steal behavior detection combined with the multi-person pose and the visual focus point is an important and practical task, and is a method for identifying the wonderful steal segment of the player in the competition video of the visual scene, and has a very broad development prospect. BRIEF DESCRIPTION OF DRAWINGS

[0047] The drawings accompanying the specification of the present application serve to provide a further understanding of the present application, and the illustrative embodiments of the present application and their descriptions serve to explain the present application, and do not constitute an improper limitation on the present application.

[0048] Figure 1 The dribble steal behavior detection method based on a visual focus point provided by the embodiments of the present application is shown in the flowchart;

[0049] Figure 2A flowchart of a line-of-sight estimation method provided by an embodiment of the present application is shown in FIG. 1.

[0050] Figure 3 A flowchart of a method for obtaining a multi-person line-of-sight focus area provided by an embodiment of the present application is shown in FIG. 2. DETAILED DESCRIPTION

[0051] It should be noted that the following detailed description is merely exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0052] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting, as the scope of the application will be limited only by the appended claims. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. Unless otherwise required by context, singular terms shall include pluralities and vice versa. It will be further understood that the terms "comprises" and "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or groups thereof.

[0053] The embodiments in the present application and the features in the embodiments can be combined with each other in the case of no conflict.

[0054] Embodiment One

[0055] In the prior art, the time cost and calculation cost required for motion behavior detection are high, the use range is narrow, and the recognition effect is poor. Therefore, the present application provides a dribble steal behavior detection method based on line-of-sight focus. The method takes a pre-processed game video image sequence as input, pre-judges the game area where a wonderful steal may occur, determines the multi-person line-of-sight focus game area, and then detects the steal event inside the area. First, the pre-processed game video is input into a multi-target detection tracker to obtain a body part detection box associated with the identity ID of each player on the court, and a head detection box is obtained therefrom. The physical coordinates of the head center point are obtained by combining the three-dimensional reconstruction result of the court, and the real-time head direction of each player is obtained by using the head image of each player. Then, the physical coordinates and direction information of the head are associated by using the motion trajectory of the player to obtain a head three-dimensional pose estimation result, which approximately replaces the line-of-sight direction and is mapped into the court physical space, so as to obtain a multi-person line-of-sight focus area, and the image sequence before and after the line-of-sight focus stage in this area is segmented for three-dimensional human body pose estimation. The pose estimation result is classified to determine whether a steal event actually occurred in the area.

[0056] Next, the method will be described in combination with Figures 1-3The method for detecting the dribble steal behavior based on the gaze focus point is disclosed in the embodiment. The method comprises the following steps:

[0057] S0, obtaining a competition video and performing preprocessing, and obtaining a competition video image sequence.

[0058] If the original competition video is directly input into the detector, a huge amount of calculation will be generated. Therefore, the competition video needs to be frame-processed.

[0059] In the subsequent target detection process, there may be a problem of reduced detection effect due to the fast movement speed of the moving target. Therefore, a filter is used for noise reduction preprocessing to obtain a smooth frame sequence and improve the accuracy of the subsequent posture estimation step.

[0060] S1, camera calibration and three-dimensional space reconstruction of the court are performed according to the three-dimensional geometric positions of the court mark line and the basket, and the mutual relationship of the corresponding key points of the court mark line and the basket in the competition video image sequence.

[0061] In an actual competition scene, only ready-made video images can be obtained, and the traditional camera calibration method cannot be used. In the competition video, the court mark line and the basketball stand always exist in the camera shot, and the shape is relatively regular, so they can be used as key information to assist the camera self-calibration.

[0062] Specifically, first, the competition video image sequence is subjected to grayscale processing, the court mark line is segmented from the original image using an edge detection operator, a binary image with obvious court mark line is obtained through thresholding, and the straight line coordinate parameters are extracted using Hough transformation. A plurality of pairs of key point coordinates are selected to complete the calibration of the camera, the internal and external parameters of the camera and the projection matrix are calculated, and the court is three-dimensionally reconstructed on this basis.

[0063] For example, according to the regulations of the International Basketball Federation, the size of a standard basketball court is 28 meters long and 15 meters wide, and the height of a standard basketball stand is 3.05 meters (the size of a non-standard basketball court can also be obtained by measurement). By comparing the relationship between the key point coordinates of the court mark line in the two-dimensional image and the actual three-dimensional space coordinates corresponding thereto, the perspective transformation coefficient for converting the two-dimensional image space into the real three-dimensional space is solved, and the three-dimensional reconstruction of the court is completed.

[0064] S2, inputting the competition video image sequence into a preset multi-target tracking detector for processing to obtain a plurality of gaze focus areas and a plurality of gaze focus times; the specific steps comprise:

[0065] S201, input the competition video image sequence into the multi-target detector to obtain the head image of each player and the corresponding position coordinate information. The specific steps include:

[0066] S2011, input the competition video image sequence into the body region detector to obtain the body detection frame of each player and associate it with the identity ID of each player; the body region detector detects the body part of the player in a single frame of picture and generates a body detection frame. The body region detector can be a body target detection model based on yolo.

[0067] S2012, input the body detection frame with the identity ID into the multi-target tracker to obtain the body region image with the identity ID and the real-time motion coordinates of each player with the identity ID; the multi-target tracker obtains the position of each player in each frame of image and associates the front and back frames to obtain the motion trajectory of each person.

[0068] S2013, input the body region image with the identity ID into the head detector to obtain the head image with the identity ID; the head detector extracts the head detection frame from the body detection frame. The head detector can be a head detection model based on yolo.

[0069] S2014, obtain the position coordinate information of the head image relative to the body according to the real-time motion coordinates of each player with the identity ID.

[0070] S202, estimate the head posture of the player according to the head image and the position coordinate information corresponding to the head image to obtain the real-time gaze area of each player. The steps include:

[0071] S2021, input the head image into the head angle classifier to obtain the head orientation angle information of each player; the deep residual network is used as the head angle classifier, which is first trained using head image data with labels, and the head image is input into the model to obtain the head angle classification result.

[0072] Specifically, first, the head orientation angle dataset is prepared: 10 players are selected to simulate the competition scene, the real-time head turning angle is obtained by wearing a head-mounted inertial measurement unit, and the head image data is collected at the same time, the head orientation angle of each player is matched and associated with the corresponding head image. The angle obtained by the head-mounted inertial measurement unit is continuous, and the probability of the front and back of the player's head appearing in the competition is comparable, so 360° can be equally divided into several labels for the classifier to learn.

[0073] Based on the comprehensive consideration of real-time and accuracy, a deep residual network is used as the head angle classifier. The head orientation angle dataset is used for training, and small batch stochastic gradient descent method can be used for parameter optimization. After the model converges, the parameters are saved. In actual use, the head image is input into the trained model, and the head angle classification result is output.

[0074] S2022, according to the head orientation angle information and the position information of the head image of each athlete, the three-dimensional position coordinates of the head center point of each athlete are obtained;

[0075] Specifically, first, the head image is thresholded to remove background pixels, obtain the head binary image, and perform connected component analysis on the head binary image. Then, the center of the largest connected domain is calculated as the head center point of the athlete. Then, according to the position information of the head image, the head center point coordinates are restored to the corresponding competition video image sequence, and are mapped to the three-dimensional space of the court through coordinate transformation to obtain the three-dimensional position coordinates of the head center point.

[0076] S2023, according to the head orientation angle information and the three-dimensional position coordinates of the head center point of each athlete, the real-time gaze area of each athlete is obtained.

[0077] S203, according to the real-time gaze area of each athlete, the multi-person visual focus area and the multi-person visual focus time are obtained.

[0078] Since the subsequent operation needs to use the head direction of the athlete to approximate the visual direction of the athlete, the head detection accuracy is required to be high, in order to avoid false detection and other interference. In this step, a two-step detection strategy is adopted. For example, in the first step, the detection frame of each athlete's body part is obtained, and the identity ID (used to distinguish different athletes) is associated respectively. In the second step, the head detection frame is extracted from the body detection frame, that is, the reliable output result of the body area detector is used as the input of the head detector. Since the head and the body have a relatively fixed positional relationship, the method of extracting the head detection frame from the body detection frame can improve the accuracy of head recognition.

[0079] The detector first normalizes the pre-processed video frame image to a uniform size. Due to the different distances of the athletes from the camera, the sizes in the picture are greatly different. Multi-scale detection can be used to improve the accuracy. Here, 32 times, 16 times and 8 times downsampling rates are used to divide the image into a plurality of grids. A certain number of preset detection boxes are set in each grid, and whether it is an effective athlete body region is judged by setting a confidence probability threshold for the detection box. Then the effective body region image is segmented from the original image and input into the head detector for detection, and the position coordinate information of the body part image relative to the complete image is recorded. Finally, the head detector extracts the effective head image for subsequent gaze estimation operation, and the position coordinate information of the head image relative to the original image is calculated and restored with the help of the previously saved body position coordinate information.

[0080] The multi-target tracker can use the DEEPSORT algorithm as a framework. On the basis of extracting the body detection box of each athlete, the Kalman filter is used to predict the state of the front and back frames, and the motion features and appearance features are combined to associate the trajectories of the athletes.

[0081] Judging the real-time gaze area requires head pose estimation of the athletes. The head pose can be represented by a six-dimensional vector, i.e. (φ x , φ y , φ z , t x , t y , t z ). The real-time head orientation information is represented by (φ x , φ y , φ z ), and the corresponding real-time head center physical coordinates are represented by (t x , t y , t z ).

[0082] Generally, the image quality of the competition video is low, the size of the head image is small, and the situations such as occlusion and facing away from the camera are common in sports. It is difficult to extract facial key points such as eyes, nose tip and mouth corner, and it is impossible to support the regression task of accurately estimating the head orientation angle. Therefore, the head orientation estimation here is defined as a classification task. There is a constant attack and defense conversion in the competition, so the probability of the front and back of the head is not much different. Therefore, 360° can be equally divided into 10 labels to represent the head orientation angle. The head image obtained is input into the head angle classifier to obtain the head orientation angle information of each athlete, and the midpoint of each label range is recorded as (φ x , φ y , φ z ).

[0083] Further analysis and processing of the obtained images of the athlete's head, and associated with the position coordinate information can obtain the three-dimensional spatial coordinates of the head center point. In order to obtain higher positioning accuracy, first, the head image is thresholded to remove the background pixels, obtain the head binary image, and perform connected component analysis to remove other small area pixel interference, and then calculate the center of gravity of the largest connected component, which is determined as the head center point of the athlete in the two-dimensional image. Then, according to the position information of the head detection frame in the original image, the two-dimensional head center point is restored to the original image, and further mapped to the three-dimensional space of the court by coordinate transformation, to obtain the three-dimensional position coordinates of the athlete's head center point (t x , t y , t z ).

[0084] Obtain the real-time gaze area of each athlete. The real-time head orientation angle information (φ x , φ y , φ z ) of each athlete is associated with the three-dimensional spatial coordinates of the head center point (t x , t y , t z ). From the three-dimensional coordinates of the head center point, the head orientation angle is mapped and projected in three-dimensional space to obtain the gaze area of each athlete in the three-dimensional space of the court.

[0085] For the obtained real-time gaze area of each athlete, it is regarded as a point set in three-dimensional space, and clustering operation is performed. The clustering center of the largest cluster is regarded as the center of the multi-person gaze focusing area, and the coordinates are marked as (x, y, z). The three-dimensional coordinates of the center point are restored to the two-dimensional coordinates in the original image. A cropping frame is generated with the coordinates as the center to crop the original image. At the same time, the time range of the video segment is the time period when the gaze focusing center point exists and 5 seconds before and after it. Among them, in order to exclude the interference of the athlete's full court scanning gaze during the game (when multiple players appear scanning gaze, the wonderful steal event usually does not occur), if the time length of the multi-person gaze focusing area center point exceeds a certain threshold (such as 3 seconds), it is regarded as an effective gaze focusing area center and the video is cropped; if the following two situations occur, the segment is directly regarded as a video segment of no interest and excluded: (1) the multi-person gaze focusing time is too short; (2) when the gazes of each athlete are relatively dispersed, the clusters in the clustering result are relatively dispersed and have comparable volume, and there is no multi-person gaze focusing area.

[0086] S3, input the gaze focusing area video segment into the trained group motion event classification model for processing to determine whether a dribble steal behavior occurs, and output the determination result. The specific steps include:

[0087] S301, input the line-of-sight focus area video segment and the body detection box into the trained group motion event classification model to extract the body skeleton of each athlete; wherein the group motion event classification model can be a classification model based on a C3D convolutional neural network.

[0088] S302, determine whether a dribble steal behavior occurs according to the body skeleton of each athlete.

[0089] The present embodiment judges whether a steal event occurs in the multi-person line-of-sight focus area. Basketball games belong to multi-person motion events, and when a steal event occurs, the images of the body parts of the dribbler and the stealer usually have serious occlusion problems. If the single-person motion posture estimation result is used as the basis for classification of motion events, a large amount of information about group motion behavior will be lost, resulting in low event classification accuracy. Therefore, the line-of-sight focus area video segment obtained in step S2 is directly subjected to group motion event analysis.

[0090] The group motion event classification model uses a deep neural network. The principle is to extract the body skeletons of multiple athletes and further estimate their postures, analyze the group motion events, and determine whether a steal event has occurred. Since the body part detection boxes of the players present in step S2 have been obtained, they are used as inputs to the event classification model together with the video, which can exclude the interference of other objects on the court on the extraction of body skeletons and posture estimation, thereby improving the accuracy.

[0091] First, a large basketball game dataset is used to train the group motion event classification model. In the dataset, 50 complete basketball game videos shot from different angles and containing athlete bounding boxes and human skeleton annotations are selected, and video segments of different motion events are extracted and manually labeled as steal events and non-steal events.

[0092] After the model is trained, the multi-person line-of-sight focus area video obtained in step S2 is first input, the body skeletons of each athlete are extracted and labeled based on the body detection box (such as using the openpose framework), which is input to the model. The model determines whether a steal event has occurred and outputs the determination result.

[0093] Embodiment Two

[0094] The present embodiment discloses a dribble steal behavior detection system based on a line-of-sight focus point, comprising:

[0095] A first acquisition module configured to acquire a game video and perform preprocessing to obtain a game video image sequence.

[0096] The second acquisition module is configured to input the competition video image sequence into a preset multi-target tracking detector for processing, acquire a multi-person visual focus area and a multi-person visual focus time, and perform frame cutting and segment interception on the competition video according to the multi-person visual focus area and the multi-person visual focus time to acquire a visual focus area video segment.

[0097] The dribble steal behavior detection module is configured to input the visual focus area video segment into a trained group motion event classification model for processing to determine whether a dribble steal behavior occurs and output a determination result.

[0098] It should be noted that the first acquisition module, the second acquisition module and the dribble steal behavior detection module correspond to the steps in Embodiment One, and the above modules have the same examples and application scenarios as the corresponding steps, but are not limited to the content disclosed in Embodiment One. It should be noted that the above modules can be executed in a computer system such as a set of computer executable instructions as part of the system.

[0099] Embodiment Three

[0100] Embodiment Three of the present application provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and running on the processor, when the computer instructions are executed by the processor, the steps of the above dribble steal behavior detection method based on visual focus are completed.

[0101] Embodiment Four

[0102] Embodiment Four of the present application provides a computer readable storage medium for storing computer instructions, when the computer instructions are executed by the processor, the steps of the above dribble steal behavior detection method based on visual focus are completed.

[0103] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of the flows and / or blocks in the flowchart and / or block diagram can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device realize a function specified in one or more flows and / or blocks in the flowchart and / or block diagram. Figure 1 The function specified in one or more flows or blocks Figure 1 The device for realizing the function specified in one or more flows or blocks.

[0104] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 of the flow or flows and / or blocks Figure 1 of the block or blocks specified in the flow.

[0105] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions that are executed on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 of the flow or flows and / or blocks Figure 1 of the block or blocks specified in the flow.

[0106] The above description of the various embodiments can have emphasized different features and variations. Details that are not described for a particular embodiment can be found in the description of the other embodiments.

[0107] The above description is only preferred embodiments of the present application, and is not intended to limit the present application. The present application can be variously changed and modified by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of the present application.

Claims

1. A method for detecting a ball interception behavior based on a line-of-sight focus, characterized in that, The method comprises the following steps: acquiring a competition video and preprocessing the competition video to obtain a competition video image sequence; inputting the competition video image sequence into a preset multi-target tracking detector for processing to obtain a multi-person visual focus area and a multi-person visual focus time, including: inputting the competition video image sequence into a multi-target detector to obtain head images of each athlete and corresponding position coordinate information; performing head posture estimation on the athletes according to the head images and the position coordinate information corresponding to the head images to obtain real-time gaze areas of each athlete, including: inputting the head images into a head angle classifier to obtain head orientation angle information of each athlete; obtaining three-dimensional position coordinates of head center points of each athlete according to the head orientation angle information of each athlete and the position information of the head images; and obtaining real-time gaze areas of each athlete according to the head orientation angle information of each athlete and the three-dimensional position coordinates of the head center points; obtaining the multi-person visual focus area and the multi-person visual focus time according to the real-time gaze areas of each athlete; performing frame cropping and segment cutting on the competition video according to the multi-person visual focus area and the multi-person visual focus time to obtain a visual focus area video segment; inputting the visual focus area video segment into a trained group sports event classification model for processing to determine whether a dribble tackle behavior occurs and output a determination result.

2. The line-of-sight focus-based pick-and-roll play detection method of claim 1, wherein, The step of inputting the competition video image sequence into the multi-target detector to obtain the head images of each athlete and the corresponding position coordinate information comprises the following steps: inputting the competition video image sequence into a body region detector to obtain body detection boxes of each athlete and associate the body detection boxes with athlete IDs; inputting the body detection boxes with the athlete IDs into a multi-target tracker to obtain body region images with the athlete IDs and real-time motion coordinates of each athlete with the athlete IDs; inputting the body region images with the athlete IDs into a head detector to obtain head images with the athlete IDs; obtaining position coordinate information of the head images relative to the bodies according to the real-time motion coordinates of each athlete with the athlete IDs.

3. The line-of-sight focus-based pick-and-roll play detection method of claim 1, wherein, The step of obtaining the three-dimensional position coordinates of the head center points of each athlete according to the head orientation angle information of each athlete and the position information of the head images comprises the following steps: performing threshold processing on the head images to remove background pixels, obtaining head binary images, and performing connected domain analysis on the head binary images; calculating a center of a maximum connected domain as a head center point of the athlete; restoring the head center point coordinates to corresponding competition video image sequences according to the position information of the head images, and mapping the head center point coordinates to a three-dimensional space of a playing field through coordinate transformation to obtain three-dimensional position coordinates of the head center points.

4. The line-of-sight focus-based pick-and-roll play detection method of claim 1, wherein, The step of inputting the visual focus area video segment into the trained group sports event classification model for processing comprises the following steps: inputting the visual focus area video segment and the body detection boxes into the trained group sports event classification model to extract body skeletons of each athlete; determining whether a dribble tackle behavior occurs according to the body skeletons of each athlete.

5. The line-of-sight focus-based pick-and-roll play detection method of claim 1, wherein, Before the step of inputting the competition video image sequence into the preset multi-target tracking detector for processing, the method further comprises the following steps: According to the three-dimensional geometric positions of the court marking lines and the basket, and the mutual relationship of the court marking lines and the basket corresponding key points in the game video image sequence, camera calibration and three-dimensional space reconstruction of the court are performed.

6. A visual line focus-based ball handling steal detection system, characterized by, The method comprises the following steps: A first acquisition module is configured to acquire a game video and perform preprocessing, and acquire a game video image sequence; A second acquisition module is configured to input the game video image sequence into a preset multi-target tracking detector for processing, and acquire multi-person visual focus areas and multi-person visual focus times; According to the multi-person visual focus areas and the multi-person visual focus times, the game video image sequence is input into a multi-target detector to acquire head images of each player and corresponding position coordinate information; According to the head images and the position coordinate information corresponding to the head images, head posture estimation is performed on the players to acquire real-time gaze areas of each player, which comprises the following steps: the head images are input into a head angle classifier to acquire head orientation angle information of each player; according to the head orientation angle information of each player and the position information of the head images, three-dimensional position coordinates of head center points of each player are acquired; according to the head orientation angle information of each player and the three-dimensional position coordinates of the head center points, real-time gaze areas of each player are acquired; According to the real-time gaze areas of each player, multi-person visual focus areas and multi-person visual focus times are acquired, The game video is frame cropped and segment intercepted to acquire visual focus area video segments; A dribble steal behavior detection module is configured to input the visual focus area video segments into a trained group motion event classification model for processing to determine whether a dribble steal behavior occurs, and output a determination result.

7. An electronic device, comprising: A computer device comprising a memory and a processor, and computer instructions stored in the memory and running on the processor, when the computer instructions are run by the processor, the steps of any one of claims 1-5 are completed.

8. A computer-readable storage medium, characterized in that, A computer program product for storing computer instructions, when the computer instructions are executed by a processor, the steps of any one of claims 1-5 are completed.

Images