A method for detecting the number of times of juggling in football training and a computer-readable storage medium

Abstract

A method for detecting the number of juggling times in football training and a computer-readable storage medium, the method comprising: step 1, obtaining a juggling training video; step 2, intercepting the juggling training video into multiple frames of images, and sequentially inputting them into a trained football juggling key point detection network to obtain human skeleton point information and football point information in each frame of the image; step 3, based on the human skeleton point information and football point information in the image, judging whether the juggling condition is met according to the correlation between the football point and the human skeleton point during the juggling process, and if the juggling condition is met, it is recorded as one juggling; step 4, sequentially traversing the images to obtain the number of juggling times in football training. The present invention collects the required football training video through a user device with a camera function, and can accurately detect the number of football training juggling times of a football trainer through subsequent processing, which changes the existing method of observing and counting by a coach or a fellow trainee in the football training process.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for detecting the number of dribbling times in football training and a computer-readable storage medium. Background Art

[0002] Deep learning is a powerful machine learning technology that recognizes patterns and data features by simulating the neural network structure of the human brain. In the human key point detection technology, deep learning algorithms can identify and locate key parts of the human body, such as joints, head, etc., and are often used in fields such as action recognition, virtual reality, and augmented reality. This technology trains a large amount of image data to allow the model to automatically detect and understand the structure and posture of the human body, thereby achieving accurate recognition and analysis of complex human movements. At present, with the emphasis on youth sports, the proportion of sports training in schools is increasing. Visual detection is currently mainly used in push-ups, pull-ups and other detections to assist coaches in supervising the physical training of relevant students.

[0003] Currently, there is almost no detection of training contents such as juggling in football training through human key point detection technology. Summary of the invention

[0004] In view of the technical defects and technical drawbacks in the prior art, the embodiments of the present invention provide a method for detecting the number of dribbling times in football training and a computer-readable storage medium that overcomes the above problems or at least partially solves the above problems. The specific scheme is as follows:

[0005] As a first aspect of the present invention, a method for detecting the number of dribbling times in football training is provided, the method comprising:

[0006] Step 1, obtain the juggling training video;

[0007] Step 2, intercepting the juggling training video into multiple frames of images, and sequentially inputting them into the trained football juggling key point detection network, and obtaining the human skeleton point information and football point information in each frame of the image through the football juggling key point detection network;

[0008] Step 3, based on the human skeleton point information and the football point information in each frame of the image obtained in step 2, judging whether the juggling condition is met according to the correlation between the football point and the human skeleton point during the juggling process, and if the juggling condition is met, it is recorded as one juggling;

[0009] Step 4, based on the method of step 3, traverse the image in sequence to obtain the number of juggling times for football training.

[0010] Furthermore, the method also includes: obtaining a football juggling picture, marking football point information and human skeleton point information in the football juggling image to obtain a football juggling point data set; inputting the football juggling point data set into a football juggling key point detection network for training, and obtaining a trained football juggling key point detection network.

[0011] Furthermore, the method also includes: acquiring street view pictures, marking human skeleton point information in the street view pictures, and using the marked street view pictures and the football juggling pictures together as a football juggling point data set.

[0012] Further, based on the human skeleton point information and the football point information in each frame of the image obtained in step 2, judging whether the juggling condition is met according to the correlation between the football point and the human skeleton point during the juggling process, and if the juggling condition is met, it is recorded as one juggling, which specifically includes:

[0013] When the subject dribbles the ball with both legs:

[0014] If one leg of the subject in the image is in a bent state and the other leg is in a straight state, and at this time, the distance between the measured football point and all the measured human skeleton points is closest to the knee point of the bent leg, the state in the corresponding image is marked as state one; if in the case that state one has occurred, if both legs of the subject in the image are in a straight state and the football point is higher than the double hip points of the subject, the state in the corresponding image is marked as state two; if in the case that state two has occurred, if one leg of the subject in the image is in a straight state and the other leg is in a bent state, and at this time, the distance between the measured football point and all the measured human skeleton points is closest to the knee point of the bent leg, the state in the corresponding image is marked as state three; starting from the first frame of the image and traversing backwards in sequence, the subject completes a transition from state one to state two and then to state three, which is recorded as one double-leg and thigh juggling;

[0015] When the subject dribbles the ball with both feet:

[0016] If the distance between the football point in the image and all the human skeleton points of the subject is the closest to the ankle point on any side of the subject, the state in the corresponding image is marked as state one; if the football point in the image is higher than any knee point of the subject when state one has occurred, it is marked, and the state in the corresponding image is marked as state two; if the distance between the football point in the image and all the human skeleton points of the subject is the closest to the ankle point on the other side of the subject when state two has occurred, the state in the corresponding image is marked as state three; starting from the first frame of the image, the test subject completes a transition from state one to state two and then to state three, which is recorded as one double-footed dribbling;

[0017] When the subject juggles the ball with the thigh and foot alternately:

[0018] If one leg of the subject in the image is in a bent state, and the distance between the football point and all the human bone points of the subject is the closest to the corresponding hip point or knee point of the subject, the state in the corresponding image is marked as state one. When state one has occurred, if the distance between the football point in the image and all the human bone points of the subject is the closest to the corresponding ankle point of the subject, the state in the corresponding image is marked as state two; when state two has occurred, if the football point in the image is higher than any knee point of the subject, the state in the corresponding image is marked as state three; starting from the first frame of the image and traversing backwards in sequence, after one leg of the subject completes a transition from state one to state two and then to state three, if the other leg of the subject also completes a transition from state one to state two and then to state three, the subject has completed a whole thigh-foot alternating juggling, which is recorded as one thigh-foot alternating juggling.

[0019] Further, the human skeleton point information includes a left hip point, a right hip point, a left knee point, a right knee point, a left ankle point, and a right ankle point;

[0020] The left hip point, left knee point and left ankle point in the image are connected into a broken line as the left leg broken line. When the angle of the left leg broken line is less than 120 degrees, it is judged that the left leg of the subject in the corresponding image is in a bent state; when the angle is greater than 160 degrees, it is judged that the left leg of the subject in the corresponding image is in an extended state. Similarly, the right hip point, right knee point and right ankle point in the image are connected into a broken line as the right leg broken line. When the angle of the right leg broken line is less than 120 degrees, it is judged that the right leg of the subject in the corresponding image is in a bent state; when the angle is greater than 160 degrees, it is judged that the right leg of the subject in the corresponding image is in an extended state.

[0021] Furthermore, the human skeleton point information and the football point information in the image are detected by the football juggling key point detection network, which specifically includes the following steps:

[0022] Step 201: the input image of the football juggling key point detection network first passes through two convolutional layer structures in sequence to obtain an output, wherein the convolutional layer structure consists of a convolutional layer, a batch normalization layer and an ELU activation function;

[0023] Step 202: extract features from the output of step 201 through a shallow feature extraction layer;

[0024] Step 203: passing the output of the shallow feature extraction layer in step 202 through two convolutional layer structures to form two output branches, namely, convolutional layer output branch 1 and convolutional layer output branch 2, wherein the convolutional kernel stride and number of the convolutional layer structure corresponding to the convolutional layer output branch 2 are twice those of the convolutional layer structure corresponding to the convolutional layer output branch 1;

[0025] Step 204: inputting the output of step 203 into the first feature extraction layer, the first feature extraction layer includes a first multi-scale feature extraction module, the first multi-scale feature extraction module includes two corresponding input branches and two output branches, the convolution layer output branch 1 and the convolution layer output branch 2 are respectively input into the two input branches of the first multi-scale feature extraction module, and two output branches of the first multi-scale feature extraction module are obtained, which are the output branch 1 and the output branch 2 of the first multi-scale feature extraction module;

[0026] Step 205: using output branch 1 of the first multi-scale feature extraction module as input branch 1 of the second feature extraction layer, and using output branch 2 of the first multi-scale feature extraction module as input branch 2 of the second feature extraction layer; passing output branch 2 of the first multi-scale feature extraction module through a convolutional layer structure a to obtain input branch 3 of the second feature extraction layer; passing output branch 2 of the first multi-scale feature extraction module through a series of convolutional layer structures b and c to obtain input branch 4 of the second feature extraction layer;

[0027] Step 206: the second feature extraction layer includes a second multi-scale feature extraction module, the second multi-scale feature extraction module includes four input branches and four output branches, the four input branches of the second feature extraction layer are respectively input into the four input branches of the second multi-scale feature extraction module, and four output branches of the second multi-scale feature extraction module are obtained;

[0028] Step 207, the four output branches of the second multi-scale feature extraction module are respectively passed through four human feature extractors to obtain four new output branches, namely, new output branch 1, new output branch 2, new output branch 3 and new output branch 4. The new output branches 2, 3 and 4 are respectively passed through an upsampling structure and added to the new output branch 1. The added result is then passed through an ELU activation function to obtain an output;

[0029] Step 208, pass the output of the ELU activation function of step 207 through a convolution layer with 18 convolution kernels to obtain the final output result of the football juggling key point detection network, calculate the heat map of the 18 output channels of the football juggling key point detection network, and obtain the predicted key points of the human body and the football point.

[0030] Further, the shallow feature extraction layer includes a human feature extractor;

[0031] The human feature extractor comprises two convolutional layer structures, a convolutional layer structure without an ELU activation function, and a spectrum fusion module; the input of the human feature extractor comprises two branches, the first branch firstly obtains the input amplitude spectrum and phase spectrum through Fourier transform, and the second branch sequentially obtains the output through various convolutional layer structures of the human feature extractor, and the corresponding output is also Fourier transformed to obtain its corresponding amplitude spectrum and phase spectrum;

[0032] The spectrum fusion module takes the phase spectrum obtained by the first branch and the amplitude spectrum obtained by the second branch as input, and obtains the final output of the human feature extractor through inverse Fourier transform.

[0033] Further, in step 204, the step of inputting the convolutional layer output branch 1 and the convolutional layer output branch 2 into two input branches of the first multi-scale feature extraction module respectively to obtain two output branches of the first multi-scale feature extraction module specifically includes:

[0034] The output branch 1 of the convolution layer is passed through four human feature extractors to obtain the output of branch 1; the output branch 2 of the convolution layer is passed through four human feature extractors to obtain the output of branch 2; the output of the branch 1 is added to the output of the branch 2 through a downsampling structure, and then an ELU activation function is used to obtain the final output of branch 2; the output of the branch 2 is added to the output of the branch 1 through an upsampling structure, and then an ELU activation function is used to obtain the final output of branch 1; the final output of branch 1 and the final output of branch 2 are the output branch 1 and output branch 2 of the first multi-scale feature extraction module;

[0035] The downsampling structure is a convolution kernel with a size of 3 3. Convolutional layer structure with a stride of 2; the upsampling structure has a convolution kernel size of 1 1. It consists of a convolutional layer with a stride of 1, a batch normalization layer, and an upsampling layer.

[0036] Furthermore, in step 205, the convolution kernel strides of convolution layer a, convolution layer b, and convolution layer c are all 2, and the number of convolution kernels is twice the number of input channels of the corresponding convolution layer structure.

[0037] As a second aspect of the present invention, a computer-readable storage medium is further provided, wherein a computer program is stored in the computer-readable storage medium. When the computer program is executed by a computer, the computer executes the method for detecting the number of dribbling times in football training as described in any one of the above items.

[0038] The present invention has the following beneficial effects:

[0039] 1. A method for detecting the number of juggling times in football training provided by the present invention only requires shooting a juggling training video, intercepting the shot video into multiple frames of images, and sequentially inputting the images into a football juggling key point detection network, obtaining the football point position and human skeleton point position information in the output result of the football juggling key point detection network, and judging whether the juggling condition is met according to the correlation between the football point position and the human skeleton point position during the juggling process, thereby detecting the juggling action performed by the subject and obtaining the corresponding number of juggling times;

[0040] 2. Based on the technical solution of the present invention, the user can collect the required football training video through the camera in the user device with a camera function, such as a mobile phone, a tablet computer and other devices, and through subsequent processing, the number of football training juggling times of the football trainer can be accurately detected, which changes the existing method of observing and counting by coaches or fellow trainees during football training. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A schematic diagram of a flow chart of a method for detecting the number of dribbling times in soccer training provided by an embodiment of the present invention;

[0042] Figure 2 A schematic diagram of the measured human skeleton points and football points provided by an embodiment of the present invention;

[0043] Figure 3 It is a schematic diagram of the points of a double-leg and thigh ball juggling provided by an embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram of the positions of a double-footed ball juggling technique provided by an embodiment of the present invention;

[0045] Figure 5 This is a schematic diagram of the positions of a thigh-foot alternating juggling technique provided by an embodiment of the present invention;

[0046] Figure 6 1 is a schematic diagram of a soccer juggling key point detection network provided by an embodiment of the present invention;

[0047] Figure 7 Schematic diagram of the structure of a convolutional layer in a soccer juggling key point detection network provided by an embodiment of the present invention;

[0048] Figure 8 It is a schematic diagram of the structure of a human feature extractor in a football juggling key point detection network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] like Figure 1 As shown, an embodiment of the present invention provides a method for detecting the number of dribbling times in football training, the method comprising:

[0051] Step 1: Place the camera in front of or beside the subject to shoot a juggling training video;

[0052] Step 2, intercepting the juggling training video into multiple frames of images, and sequentially inputting them into the trained football juggling key point detection network, and obtaining the human skeleton point information and football point information in each frame of the image through the football juggling key point detection network;

[0053] Step 3, based on the human skeleton point information and the football point information in each frame of the image obtained in step 2, judging whether the juggling condition is met according to the correlation between the football point and the human skeleton point during the juggling process, and if the juggling condition is met, it is recorded as one juggling;

[0054] Step 4, based on the method of step 3, traverse the image in sequence to obtain the number of juggling times for football training.

[0055] The present invention provides a method for detecting the number of juggling times in football training. It only requires shooting a juggling training video, and obtaining football point position and human skeleton point position information in the output result of the football juggling key point detection network through a football juggling key point detection network. It is judged whether the juggling conditions are met according to the correlation between the football point position and the human skeleton point position during the juggling process, so as to detect the juggling action performed by the subject and obtain the corresponding number of juggling times.

[0056] Based on the technical solution of the present invention, the user can collect the required football training video through the camera in the user device with camera function, such as mobile phone, tablet computer and other devices, and the number of football training dribbling times of the football trainer can be accurately detected through subsequent processing, which changes the existing method of observing and counting by coaches or fellow trainees in the football training process.

[0057] In some embodiments, the method further includes training a football juggling key point detection network as follows:

[0058] A football juggling picture is obtained, and the football point information and human skeleton point information are marked in the football juggling image to obtain a football juggling point data set; the football juggling point data set is input into a football juggling key point detection network for training, and a trained football juggling key point detection network is obtained.

[0059] Optionally, in order to increase the data set, the present invention further obtains street view pictures, marks human skeleton point information in the street view pictures, and uses the marked street view pictures and football juggling pictures together as the football juggling point data set.

[0060] In the above embodiment, the soccer juggling pictures obtained should include all pictures of complete juggling actions of each juggling method, so that the data set contains image samples of all required juggling actions.

[0061] like Figure 2 As shown, in some embodiments, the acquired human skeleton point information includes nose 0, left eye 1, right eye 2, left ear 3, right ear 4, left shoulder point 5, right shoulder point 6, left elbow point 7, right elbow point 8, left wrist point 9, right wrist point 10, left hip point 11, right hip point 12, left knee point 13, right knee point 14, left ankle point 15, right ankle point 16, a total of 17 points, and the football point information includes football point 17; in the embodiment of the present invention, the left hip point 11, right hip point 12, left knee point 13, right knee point 14, left ankle point 15, right ankle point 16 are mainly used as human skeleton points, and these points are used to judge whether the legs of the subject are in a bent state. It is in a straight state. Specifically, the left hip point 11, the left knee point 13, and the left ankle point 15 in the image are connected into a broken line as the left leg broken line. When the angle of the left leg broken line is less than 120 degrees, it is judged that the left leg of the subject in the corresponding image is in a bent state; when the angle is greater than 160 degrees, it is judged that the left leg of the subject in the corresponding image is in a straight state; similarly, the right hip point 12, the right knee point 14, and the right ankle point 16 in the image are connected into a broken line as the right leg broken line. When the angle of the right leg broken line is less than 120 degrees, it is judged that the right leg of the subject in the corresponding image is in a bent state; when the angle is greater than 160 degrees, it is judged that the right leg of the subject in the corresponding image is in a straight state.

[0062] In some embodiments, based on the human skeleton point information and the football point information in each frame of the image obtained in step 2, judging whether the juggling condition is met according to the correlation between the football point and the human skeleton point during the juggling process, and if the juggling condition is met, then it is recorded as one juggling, which specifically includes:

[0063] like Figure 3 As shown, when the subject dribbles the ball with both legs:

[0064] If one leg of the subject in the image is in a bent state and the other leg is in a straight state, and at this time, the distance between the measured football point and all the measured human skeleton points is closest to the knee point of the bent leg, the state in the corresponding image is marked as state one; if in the case that state one has occurred, if both legs of the subject in the image are in a straight state and the football point is higher than the double hip points of the subject, the state in the corresponding image is marked as state two; if in the case that state two has occurred, if one leg of the subject in the image is in a straight state and the other leg is in a bent state, and at this time, the distance between the measured football point and all the measured human skeleton points is closest to the knee point of the bent leg, the state in the corresponding image is marked as state three; starting from the first frame of the image and traversing backwards in sequence, the subject completes a transition from state one to state two and then to state three, which is recorded as one double-leg and thigh juggling;

[0065] like Figure 4 As shown, when the subject dribbles the ball with both feet:

[0066] If the distance between the football point in the image and all the human skeleton points of the subject is the closest to the ankle point on any side of the subject, the state in the corresponding image is marked as state one; if the football point in the image is higher than any knee point of the subject when state one has occurred, it is marked, and the state in the corresponding image is marked as state two; if the distance between the football point in the image and all the human skeleton points of the subject is the closest to the ankle point on the other side of the subject when state two has occurred, the state in the corresponding image is marked as state three; starting from the first frame of the image, the test subject completes a transition from state one to state two and then to state three, which is recorded as one double-footed dribbling;

[0067] like Figure 5 As shown, when the subject dribbles the ball with the thigh and foot alternately:

[0068] If one leg of the subject in the image is in a bent state, and the distance between the football point and all the human bone points of the subject is the closest to the corresponding hip point or knee point of the subject, the state in the corresponding image is marked as state one. When state one has occurred, if the distance between the football point in the image and all the human bone points of the subject is the closest to the corresponding ankle point of the subject, the state in the corresponding image is marked as state two; when state two has occurred, if the football point in the image is higher than any knee point of the subject, the state in the corresponding image is marked as state three; starting from the first frame of the image and traversing backwards in sequence, after one leg of the subject completes a transition from state one to state two and then to state three, if the other leg of the subject also completes a transition from state one to state two and then to state three, the subject has completed a whole thigh-foot alternating juggling, which is recorded as one thigh-foot alternating juggling.

[0069] In the embodiment of the present invention, juggling the ball once for each leg is counted as one leg-thigh juggling, juggling the ball once for each foot is counted as one foot juggling, and juggling the ball once for each thigh and foot is counted as one thigh-foot alternating juggling. It should be noted that, in actual applications, juggling the ball once for a leg or foot can also be counted as one juggling.

[0070] Figure 6 A schematic diagram of a soccer juggling key point detection network provided by an embodiment of the present invention, such as Figure 6 As shown, in some embodiments, the detection of human skeleton point information and football point information in the image is realized through a football juggling key point detection network, which specifically includes the following steps:

[0071] Step 201: The input image of the football juggling key point detection network first passes through two convolutional layer structures in sequence to obtain an output, such as Figure 7 As shown in the figure, the convolutional layer structure consists of a convolutional layer, a batch normalization layer and an ELU activation function. The expression formula for the output of the input image through the convolutional layer structure is as follows:

[0072]

[0073] Among them, M represents the feature map, Mi is the input feature map, and the output Mo of this convolutional layer structure is obtained through a convolutional layer, a batch normalization layer and an ELU activation function;

[0074] Step 202: extract features from the output of step 201 through a shallow feature extraction layer;

[0075] Wherein, the shallow feature extraction layer is composed of a human feature extractor;

[0076] like Figure 8 As shown, the human feature extractor includes two convolutional layer structures, a convolutional layer structure that does not include an ELU activation function, and a spectrum fusion module; the input of the human feature extractor includes two branches, the first branch first obtains the input amplitude spectrum and phase spectrum through Fourier transform, and the second branch sequentially obtains the output through various convolutional layer structures of the human feature extractor, and the corresponding output is also Fourier transformed to obtain its corresponding amplitude spectrum and phase spectrum;

[0077] The spectrum fusion module takes the phase spectrum obtained by the first branch and the amplitude spectrum obtained by the second branch as input, and obtains the final output of the human feature extractor through inverse Fourier transform.

[0078] Step 203: passing the output of the shallow feature extraction layer in step 202 through two convolutional layer structures to form two output branches, namely, convolutional layer output branch 1 and convolutional layer output branch 2, wherein the convolutional kernel stride and number of the convolutional layer structure corresponding to the convolutional layer output branch 2 are twice those of the convolutional layer structure corresponding to the convolutional layer output branch 1;

[0079] Step 204: inputting the output of step 203 into the first feature extraction layer, the first feature extraction layer includes a first multi-scale feature extraction module, the first multi-scale feature extraction module includes two corresponding input branches and two output branches, the convolution layer output branch 1 and the convolution layer output branch 2 are respectively input into the two input branches of the first multi-scale feature extraction module, and two output branches of the first multi-scale feature extraction module are obtained, which are the output branch 1 and the output branch 2 of the first multi-scale feature extraction module;

[0080] The first convolution layer output branch and the second convolution layer output branch are respectively input into two input branches of the first multi-scale feature extraction module, and the two output branches of the first multi-scale feature extraction module are obtained, which specifically includes:

[0081] The output branch 1 of the convolution layer is passed through four human feature extractors to obtain the output of branch 1; the output branch 2 of the convolution layer is passed through four human feature extractors to obtain the output of branch 2; the output of the branch 1 is added to the output of the branch 2 through a downsampling structure, and then an ELU activation function is used to obtain the final output of branch 2; the output of the branch 2 is added to the output of the branch 1 through an upsampling structure, and then an ELU activation function is used to obtain the final output of branch 1; the final output of branch 1 and the final output of branch 2 are the output branch 1 and output branch 2 of the first multi-scale feature extraction module;

[0082] The downsampling structure is a convolution kernel with a size of 3 3. Convolutional layer structure with a stride of 2; the upsampling structure has a convolution kernel size of 1 1. It consists of a convolutional layer with a stride of 1, a batch normalization layer, and an upsampling layer.

[0083] Step 205: using output branch 1 of the first multi-scale feature extraction module as input branch 1 of the second feature extraction layer, and using output branch 2 of the first multi-scale feature extraction module as input branch 2 of the second feature extraction layer; passing output branch 2 of the first multi-scale feature extraction module through a convolutional layer structure a to obtain input branch 3 of the second feature extraction layer; passing output branch 2 of the first multi-scale feature extraction module through a series of convolutional layer structures b and c to obtain input branch 4 of the second feature extraction layer;

[0084] Step 206: the second feature extraction layer includes a second multi-scale feature extraction module, the second multi-scale feature extraction module includes four input branches and four output branches, the four input branches of the second feature extraction layer are respectively input into the four input branches of the second multi-scale feature extraction module, and four output branches of the second multi-scale feature extraction module are obtained;

[0085] Among them, the second multi-scale feature extraction module includes four human feature extractors 4. The four input branches of the second feature extraction layer are respectively passed through four human feature extractors 4 obtains the corresponding four output branches, each output branch is added with the result obtained by an upsampling structure of the other three output branches, and the final output branch corresponding to the corresponding input branch is obtained, thereby obtaining four final output branches corresponding to the four input branches, i.e., the four output branches of the second multi-scale feature extraction module.

[0086] Step 207, the four output branches of the second multi-scale feature extraction module are respectively passed through four human feature extractors to obtain four new output branches, namely, new output branch 1, new output branch 2, new output branch 3 and new output branch 4. The new output branches 2, 3 and 4 are respectively passed through an upsampling structure and added to the new output branch 1. The added result is then passed through an ELU activation function to obtain an output;

[0087] Step 208, pass the output of the ELU activation function of step 207 through a convolution layer with 18 convolution kernels to obtain the final output result of the football juggling key point detection network, calculate the heat map of the 18 output channels of the football juggling key point detection network, and obtain the predicted key points of the human body and the football point.

[0088] An embodiment of the present invention further provides a computer-readable storage medium, wherein a computer program is stored in the computer-readable storage medium. When the computer program is executed by a computer, the computer executes the method for detecting the number of dribbling times in football training as described in any one of the above items.

[0089] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting the number of times a football is dribbled during training, characterized in that: The method comprises: Step 1, obtain the juggling training video; Step 2, intercepting the juggling training video into multiple frames of images, and sequentially inputting them into the trained football juggling key point detection network, and obtaining the human skeleton point information and football point information in each frame of the image through the football juggling key point detection network; Step 3, based on the human skeleton point information and the football point information in each frame of the image obtained in step 2, judging whether the juggling condition is met according to the correlation between the football point and the human skeleton point during the juggling process, and if the juggling condition is met, it is recorded as one juggling; Step 4, based on the method in step 3, traverse the image in sequence to obtain the number of juggling times for football training; Wherein, step 3 specifically includes: When the subject dribbles the ball with both legs: If one leg of the subject in the image is in a bent state and the other leg is in a straight state, and at this time, the distance between the measured football point and all the measured human skeleton points is closest to the knee point of the bent leg, the state in the corresponding image is marked as state one; if in the case that state one has occurred, if both legs of the subject in the image are in a straight state and the football point is higher than the double hip points of the subject, the state in the corresponding image is marked as state two; if in the case that state two has occurred, if one leg of the subject in the image is in a straight state and the other leg is in a bent state, and at this time, the distance between the measured football point and all the measured human skeleton points is closest to the knee point of the bent leg, the state in the corresponding image is marked as state three; starting from the first frame of the image and traversing backwards in sequence, the subject completes a transition from state one to state two and then to state three, which is recorded as one double-leg and thigh juggling; When the subject dribbles the ball with both feet: If the distance between the football point in the image and all the human skeleton points of the subject is the closest to the ankle point on any side of the subject, the state in the corresponding image is marked as state one; if the football point in the image is higher than any knee point of the subject when state one has occurred, it is marked, and the state in the corresponding image is marked as state two; if the distance between the football point in the image and all the human skeleton points of the subject is the closest to the ankle point on the other side of the subject when state two has occurred, the state in the corresponding image is marked as state three; starting from the first frame of the image, the test subject completes a transition from state one to state two and then to state three, which is recorded as one double-footed dribbling; When the subject juggles the ball with the thigh and foot alternately: If one leg of the subject in the image is in a bent state, and the distance between the football point and all the human bone points of the subject is the closest to the corresponding hip point or knee point of the subject, the state in the corresponding image is marked as state one. When state one has occurred, if the distance between the football point in the image and all the human bone points of the subject is the closest to the corresponding ankle point of the subject, the state in the corresponding image is marked as state two; when state two has occurred, if the football point in the image is higher than any knee point of the subject, the state in the corresponding image is marked as state three; starting from the first frame of the image and traversing backwards in sequence, after one leg of the subject completes a transition from state one to state two and then to state three, if the other leg of the subject also completes a transition from state one to state two and then to state three, the subject has completed a whole thigh-foot alternating juggling, which is recorded as one thigh-foot alternating juggling.

2. The method for detecting the number of times of dribbling in football training according to claim 1, characterized in that: The method also includes: obtaining a football juggling picture, marking football point information and human skeleton point information in the football juggling image to obtain a football juggling point data set; inputting the football juggling point data set into a football juggling key point detection network for training, and obtaining a trained football juggling key point detection network.

3. The method for detecting the number of dribbling times in football training according to claim 2, characterized in that: The method further includes: acquiring a street view picture, marking human skeleton point information in the street view picture, and using the marked street view picture and the football juggling picture together as a football juggling point data set.

4. The method for detecting the number of times of dribbling in football training according to claim 1, characterized in that: The human skeleton point information includes a left hip point, a right hip point, a left knee point, a right knee point, a left ankle point, and a right ankle point; The left hip point, left knee point and left ankle point in the image are connected into a broken line as the left leg broken line. When the angle of the left leg broken line is less than 120 degrees, it is judged that the left leg of the subject in the corresponding image is in a bent state; when the angle is greater than 160 degrees, it is judged that the left leg of the subject in the corresponding image is in an extended state. Similarly, the right hip point, right knee point and right ankle point in the image are connected into a broken line as the right leg broken line. When the angle of the right leg broken line is less than 120 degrees, it is judged that the right leg of the subject in the corresponding image is in a bent state; when the angle is greater than 160 degrees, it is judged that the right leg of the subject in the corresponding image is in an extended state.

5. The method for detecting the number of times of dribbling in football training according to claim 1, characterized in that: The football juggling key point detection network realizes the detection of human skeleton point information and football point information in the image, and specifically comprises the following steps: Step 201: the input image of the football juggling key point detection network first passes through two convolutional layer structures in sequence to obtain an output, wherein the convolutional layer structure consists of a convolutional layer, a batch normalization layer and an ELU activation function; Step 202: extract features from the output of step 201 through a shallow feature extraction layer; Step 203: passing the output of the shallow feature extraction layer in step 202 through two convolutional layer structures to form two output branches, namely, convolutional layer output branch 1 and convolutional layer output branch 2, wherein the convolutional kernel stride and number of the convolutional layer structure corresponding to the convolutional layer output branch 2 are twice those of the convolutional layer structure corresponding to the convolutional layer output branch 1; Step 204: inputting the output of step 203 into the first feature extraction layer, the first feature extraction layer includes a first multi-scale feature extraction module, the first multi-scale feature extraction module includes two corresponding input branches and two output branches, the convolution layer output branch 1 and the convolution layer output branch 2 are respectively input into the two input branches of the first multi-scale feature extraction module, and two output branches of the first multi-scale feature extraction module are obtained, which are the output branch 1 and the output branch 2 of the first multi-scale feature extraction module; Step 205: using output branch 1 of the first multi-scale feature extraction module as input branch 1 of the second feature extraction layer, and using output branch 2 of the first multi-scale feature extraction module as input branch 2 of the second feature extraction layer; passing output branch 2 of the first multi-scale feature extraction module through a convolutional layer structure a to obtain input branch 3 of the second feature extraction layer; passing output branch 2 of the first multi-scale feature extraction module through a series of convolutional layer structures b and c to obtain input branch 4 of the second feature extraction layer; Step 206: the second feature extraction layer includes a second multi-scale feature extraction module, the second multi-scale feature extraction module includes four input branches and four output branches, the four input branches of the second feature extraction layer are respectively input into the four input branches of the second multi-scale feature extraction module, and four output branches of the second multi-scale feature extraction module are obtained; Step 207, the four output branches of the second multi-scale feature extraction module are respectively passed through four human feature extractors to obtain four new output branches, namely, new output branch 1, new output branch 2, new output branch 3 and new output branch 4. The new output branches 2, 3 and 4 are respectively passed through an upsampling structure and added to the new output branch 1. The added result is then passed through an ELU activation function to obtain an output; Step 208, pass the output of the ELU activation function of step 207 through a convolution layer with 18 convolution kernels to obtain the final output result of the football juggling key point detection network, calculate the heat map of the 18 output channels of the football juggling key point detection network, and obtain the predicted key points of the human body and the football point.

6. The method for detecting the number of dribbling times in football training according to claim 5, characterized in that: The shallow feature extraction layer includes a human feature extractor; The human feature extractor comprises two convolutional layer structures, a convolutional layer structure without an ELU activation function, and a spectrum fusion module; the input of the human feature extractor comprises two branches, the first branch firstly obtains the input amplitude spectrum and phase spectrum through Fourier transform, and the second branch sequentially obtains the output through various convolutional layer structures of the human feature extractor, and the corresponding output is also Fourier transformed to obtain its corresponding amplitude spectrum and phase spectrum; The spectrum fusion module takes the phase spectrum obtained by the first branch and the amplitude spectrum obtained by the second branch as input, and obtains the final output of the human feature extractor through inverse Fourier transform.

7. The method for detecting the number of times of dribbling in football training according to claim 5, characterized in that: In step 204, the step of inputting the convolutional layer output branch 1 and the convolutional layer output branch 2 into two input branches of the first multi-scale feature extraction module respectively to obtain two output branches of the first multi-scale feature extraction module specifically includes: The output branch 1 of the convolution layer is passed through four human feature extractors to obtain the output of branch 1; the output branch 2 of the convolution layer is passed through four human feature extractors to obtain the output of branch 2; the output of the branch 1 is added to the output of the branch 2 through a downsampling structure, and then an ELU activation function is used to obtain the final output of branch 2; the output of the branch 2 is added to the output of the branch 1 through an upsampling structure, and then an ELU activation function is used to obtain the final output of branch 1; the final output of branch 1 and the final output of branch 2 are the output branch 1 and output branch 2 of the first multi-scale feature extraction module; The downsampling structure is a convolution kernel with a size of 3 3. Convolutional layer structure with a stride of 2; the upsampling structure has a convolution kernel size of 1 1. It consists of a convolutional layer with a stride of 1, a batch normalization layer, and an upsampling layer.

8. The method for detecting the number of times of dribbling in football training according to claim 5, characterized in that ,In step 205, the convolution kernel stride of convolution layer a, convolution layer b and convolution layer c is 2, and the number of convolution kernels is twice the number of input channels of the corresponding convolution layer structure.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, the computer executes the method for detecting the number of dribbling times in football training as claimed in any one of claims 1 to 8.

Images