Advertisement operation system capable of avoiding same-type advertisement content

Abstract

The invention relates to the technical field of advertisement operation, in particular to an advertisement operation system for avoiding advertisement contents of the same type, which comprises an action collecting and processing unit, an action analyzing and avoiding unit and a self-learning adaptive unit. The motion analysis avoiding unit uses a human body posture estimation algorithm and a biomechanics principle to accurately calculate kinetic energy and potential energy of body parts of athletes, constructs a kinetic energy flow trajectory map, sets a Hash index, and judges whether to trigger an intelligent avoiding mechanism or not by comparing a new map with a historical map. And the self-learning and self-adaption unit optimizes an algorithm model according to feedback data, updates historical data, and improves the accurate recognition capability of the system on the actions of the athletes of the same type of sports event video advertisements.

Description

Technical Field

[0001] The present invention relates to the technical field of advertisement operation, and in particular to an advertisement operation system for avoiding the same type of advertisement contents. Background Art

[0002] Advertising operation is an important technology. In today's fiercely competitive advertising market, sports event video advertising has become a key area of ​​competition for many advertisers due to its huge audience and efficient communication effect. However, with the continuous increase in advertising volume, the problem of avoiding the same type of advertising content has become increasingly prominent, especially in terms of avoiding the comparison of athletes' movement characteristics, where traditional technical means are stretched to the limit.

[0003] Traditional systems mostly focus on simple feature analysis of a single dimension such as advertising images and texts. For example, by only comparing the appearance and posture of athletes or keywords in advertising copy, it is difficult to accurately capture the deep dynamic characteristics and internal energy changes of movements. In sports event advertising scenes, subtle differences in athletes' movements can lead to very different advertising connotations from audience perceptions, and existing technologies tend to ignore key factors such as muscle force sequence, joint coordinated movement rhythm, and movement energy transfer mode. In basketball games, the shooting actions in advertisements of different brands of sports shoes have seemingly similar limb trajectories, but the efficiency of force transmission from the legs to the arms and the energy distribution for maintaining body balance are unique. Traditional methods cannot distinguish them, resulting in the prevalence of similar advertisements, weakening brand recognition and advertising effectiveness. In order to solve this technical problem, we provide an advertising operation system that avoids similar advertising content. Summary of the invention

[0004] The purpose of the present invention is to provide an advertising operation system that avoids the same type of advertising content, so as to solve the problems raised in the above-mentioned background technology.

[0005] The present invention comprises an action acquisition and processing unit that collects and pre-processes sports event video advertisement samples; an action analysis and avoidance unit that constructs a map through human body posture and biomechanics calculations, stores and compares the data to decide whether to avoid; and a self-learning and adaptive unit that optimizes algorithms and data based on feedback to achieve accurate avoidance of the same type of advertisements and improve the quality of advertising operations.

[0006] To achieve the above-mentioned purpose, an advertising operation system for avoiding the same type of advertising content is provided, including a motion collection and processing unit, a motion analysis and avoidance unit, and a self-learning and adaptive unit;

[0007] The action collection and processing unit is used to collect sports event video advertisement samples containing athlete actions from the advertisement material library, and pre-process the samples to obtain an advertisement sample set;

[0008] The motion analysis and avoidance unit calculates the kinetic energy and potential energy of each part of the athlete's body during the movement process based on the mass parameters, joint coordinates and limb angle information of each part of the athlete's body through the human posture estimation algorithm and biomechanics principles, and constructs a motion energy flow trajectory map based on this, stores the motion energy flow trajectory map and related advertising metadata in the advertising material library, and establishes an index structure, and then uses a method based on the combination of dynamic time warping algorithm and energy distribution similarity calculation to compare the motion energy flow trajectory map with the historical advertising map in the advertising sample set, and determines whether to start the intelligent avoidance mechanism based on the comparison result;

[0009] The self-learning adaptive unit dynamically optimizes the motion energy flow trajectory extraction algorithm and the similarity calculation model based on the feedback data of the comparison results in the actual advertising delivery effect, and updates the historical advertising data in the database according to the optimization results.

[0010] As a further improvement of the technical solution, the action acquisition processing unit uses an image enhancement algorithm to enhance the video advertisement sample image during the preprocessing process, as follows:

[0011] Convert the video ad sample image from RGB color space to grayscale image, count the number of pixels at each grayscale level in the grayscale image, create an array of length 256, calculate the cumulative distribution function based on the array, and then calculate the mapping table based on the cumulative distribution function;

[0012] The grayscale image is traversed again, and the original grayscale value of each pixel is replaced by the value in the mapping table as a new grayscale value, and the new grayscale value is used to enhance the image.

[0013] As a further improvement of the present technical solution, the motion analysis and avoidance unit includes an energy flow trajectory extraction module. The energy flow trajectory extraction module and the human body posture estimation algorithm adopt a convolutional neural network model based on deep learning. The specific steps are as follows:

[0014] Extract sports event video advertisement samples from the advertisement sample set, and perform frame-by-frame segmentation to obtain a video frame sequence;

[0015] Each frame of image is input into a pre-trained convolutional neural network, which includes a convolution layer, a pooling layer and a fully connected layer. The convolution layer uses a convolution kernel to slide on the image to perform a convolution operation, and obtains feature maps of different levels by weighted summation of the convolution kernel and image pixels. The pooling layer uses a maximum pooling operation to slide a window of a fixed size on the feature map, and selects the maximum value in the window as the output after pooling. The fully connected layer performs a fully connected operation on the feature vector output by the pooling layer, converts the feature vector into a predicted value of the joint point coordinates, and finally obtains an accurate athlete's body posture estimation result.

[0016] As a further improvement of the technical solution, the energy flow trajectory extraction module calculates the kinetic energy and potential energy of each part of the athlete's body based on the principles of biomechanics, as follows:

[0017] For kinetic energy calculation, according to the formula ;in are the mass parameters of various parts of the body, The instantaneous speed of the part is calculated according to the change of the joint point coordinates. The speed value is obtained by differential calculation of the joint point coordinates in consecutive frames, and then the kinetic energy value of each part at different times is calculated according to the speed value;

[0018] For potential energy calculation, according to the formula ;in is the height of the body parts relative to the ground plane. The height change of each part is determined by the coordinate information of the joint points, and the potential energy value is calculated in combination with the mass parameter;

[0019] The calculated kinetic energy and potential energy values ​​are then integrated according to the time series to construct an energy flow data sequence that describes the energy changes during the athlete's movement, which serves as the basic data for constructing the sports energy flow trajectory map.

[0020] As a further improvement of the technical solution, the steps of constructing a motion energy flow trajectory map in the energy flow trajectory extraction module are as follows:

[0021] The kinetic energy and potential energy values ​​of each part of the body are used as the two dimensional data of the map, and time is used as the horizontal axis to construct a two-dimensional energy flow trajectory map;

[0022] The kinetic energy and potential energy values ​​of each body part at different time points are marked and connected to form the energy flow trajectory curve of the part; the curves of different body parts together constitute a complete motion energy flow trajectory map, and metadata information related to the advertisement is also annotated in the map;

[0023] The metadata information related to the advertisement includes the brand to which the advertisement belongs, the theme of the advertisement, and the time of delivery.

[0024] As a further improvement of the present technical solution, the action analysis avoidance unit includes a data construction storage module, and the method for establishing an index structure in the data construction storage module is as follows:

[0025] The indexing method based on hash algorithm is used to perform hash calculation on the feature vector of the motion energy flow trajectory map to obtain the corresponding hash value;

[0026] A hash index table is constructed by using the hash value as the index key and the storage address of the advertising metadata and the graph in the advertising material library as the index value.

[0027] As a further improvement of the present technical solution, the action analysis avoidance unit includes a delivery comparison avoidance module, and the delivery comparison avoidance module uses a method based on a dynamic time warping algorithm combined with energy distribution similarity calculation to perform comparison steps, which are specifically as follows:

[0028] The dynamic time warping algorithm is used to warp the time axis of the motion energy flow trajectory map and the historical advertising map in the advertising sample set to find the optimal matching path between the two maps in the time dimension.

[0029] Along the optimal matching path, the energy distribution similarity at the corresponding time is calculated. For the energy distribution similarity calculation, the similarity calculation formula based on Euclidean distance is used to calculate the kinetic energy distribution similarity and the potential energy distribution similarity respectively, and the total energy distribution similarity is obtained by combining the two.

[0030] A similarity threshold is set, and the calculated total energy distribution similarity is compared with the threshold. If the similarity is greater than the threshold, it is determined to be similar advertising content and the intelligent avoidance mechanism is activated. If the similarity is less than or equal to the threshold, it is determined to be non-similar advertising content.

[0031] As a further improvement of the technical solution, the self-learning adaptive unit adopts an optimization algorithm based on gradient descent when dynamically optimizing the motion energy flow trajectory extraction algorithm, as follows:

[0032] According to the proportion of the advertisements that are not correctly identified as the same type in the feedback data of the advertising effect, a loss function is constructed, and the gradient of the loss function with respect to the parameters in the motion energy flow trajectory extraction algorithm is calculated;

[0033] According to the gradient, the parameters are updated according to the update formula of the gradient descent algorithm, and the above steps of defining the loss function, calculating the gradient and updating the parameters are repeated until the stopping condition is met. The stopping condition is that the change in the loss function between two adjacent iterations is less than a set threshold.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] In an advertising operation system for avoiding the same type of advertising content, the motion analysis avoidance unit constructs a motion energy flow trajectory map based on the human posture estimation algorithm and biomechanical principles, deeply captures the motion details and energy change characteristics, accurately distinguishes the same type of advertisements, avoids misjudgment and omission, and improves the accuracy of advertising delivery. The intelligent and dynamic avoidance decision-making ability is based on a rigorous comparison process and dynamic threshold setting, and the avoidance strategy is timely initiated according to the similarity of the advertisements, thereby enhancing the scientificity and flexibility of advertising delivery planning and optimizing the audience's advertising experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is an overall block diagram of the present invention.

[0037] The meaning of each number in the figure is:

[0038] 1. Motion acquisition and processing unit; 2. Motion analysis and avoidance unit; 21. Energy flow trajectory extraction module; 22. Data construction and storage module; 23. Delivery comparison and avoidance module; 3. Self-learning and adaptive unit. DETAILED DESCRIPTION

[0039] 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 embodiments of the present invention, not all of the 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.

[0040] The present invention provides an advertising operation system for avoiding the same type of advertising content. Figure 1 As shown, it includes a motion collection and processing unit 1, a motion analysis and avoidance unit 2, and a self-learning and adaptive unit 3;

[0041] The action collection and processing unit 1 is used to collect sports event video advertisement samples containing athlete actions from the advertisement material library, and pre-process them to obtain an advertisement sample set;

[0042] During the preprocessing process, the motion acquisition processing unit 1 uses an image enhancement algorithm to enhance the video advertisement sample image, as follows:

[0043] For the video advertisement sample image, it is first converted from the RGB color space to a grayscale image, and the grayscale value of each pixel is calculated by the formula to obtain the grayscale image. In the subsequent histogram equalization processing, the grayscale image is easier to calculate and process the grayscale distribution of the image, which simplifies the complexity of data processing, reduces the amount of calculation, and can quickly convert the color image into a single-channel grayscale image.

[0044] Count the number of pixels at each gray level in the grayscale image. For an 8-bit image, the gray level range is 0-255. Create an array of length 256. ,in Represented as grayscale, Indicates grayscale The number of pixels, traverse the entire grayscale image, for each pixel gray value ,Will The value of is added by 1. The histogram can intuitively reflect the distribution of gray levels in the image. It is the basic data of the histogram equalization algorithm. It can accurately obtain the gray feature information of the image, facilitate the subsequent adjustment of the gray distribution, and provide a basis for calculating the cumulative distribution function.

[0045] Create an array of length 256 , used to store the value of the cumulative distribution function, first initialize , and then through the recursive formula ;calculate ,in After the calculation is completed, The value of is normalized, that is, ;in is the total number of pixels in the image, so Indicates that the gray level is less than or equal to The cumulative distribution function is the ratio of pixels in the image. It can describe the cumulative distribution of gray levels in the image and is used to determine the mapping relationship of gray levels. The mapping target of each gray level can be adjusted according to the overall gray distribution of the image to achieve contrast enhancement, providing key data for calculating the mapping table.

[0046] Create a mapping table with a length of 256 , used to map the original grayscale to the new grayscale, through the formula Calculated, where It is expressed as a rounding-down operation. For an 8-bit image, the maximum grayscale level is 255. The mapping table can determine the new grayscale level corresponding to each original grayscale level according to the cumulative distribution function, thereby achieving image enhancement. Through a reasonable mapping relationship, the originally concentrated grayscale distribution in the image can be stretched to enhance the contrast. A mapping rule for converting the original grayscale level to the new grayscale level is obtained, which provides a specific conversion method for image enhancement.

[0047] Traverse the grayscale image again, for the original grayscale value of each pixel , replace it with As the new grayscale value, this completes the image enhancement processing based on histogram equalization, so that the contrast of the image is improved, and the details of the darker and brighter areas in the image are clearer, which is more conducive to subsequent operations such as human posture estimation.

[0048] The motion analysis and avoidance unit 2 includes an energy flow trajectory extraction module 21 , a data construction and storage module 22 , and a delivery comparison and avoidance module 23 .

[0049] The energy flow trajectory extraction module 21 calculates the kinetic energy and potential energy of each part of the athlete's body during the movement process based on the human body posture estimation algorithm and biomechanical principles, according to the mass parameters, joint point coordinates and limb angle information of each part of the athlete's body, and constructs a sports energy flow trajectory map accordingly.

[0050] Among them, the human body posture estimation algorithm in the energy flow trajectory extraction module 21 adopts a convolutional neural network model based on deep learning, and the specific steps are as follows:

[0051] Sports event video advertisement samples are extracted from the advertisement sample set and segmented frame by frame to obtain video frame sequences, thereby improving data processing efficiency and making data flow more orderly.

[0052] For each frame of the image, it is input into the convolution layer, and the convolution kernel slides on the image according to the set step size. For each pixel in the image , calculated according to the convolution calculation formula:

[0053] ;

[0054] in is the convolution kernel function, For the input image pixel value, the weighted sum of the convolution kernel and the image pixels is used to obtain feature maps of different levels, extract the key features of the image, and accurately locate the contours of the athlete's body parts, providing a reliable basis for subsequent posture estimation.

[0055] The feature map output by the convolution layer enters the pooling layer. The pooling layer uses a fixed-size window to slide on the feature map to reduce the data dimension, reduce the amount of calculation, and speed up the processing process. For the maximum pooling operation, according to the formula:

[0056] ;

[0057] in The pixel value of the feature map input to the pooling layer is selected, and the maximum value in the window is selected as the output after pooling. The pooling operation can reduce the resolution of the feature map and the amount of data while retaining the main feature information, improving the computational efficiency of the network and reducing the risk of overfitting.

[0058] The feature vector output by the pooling layer enters the fully connected layer. Each neuron in the fully connected layer performs a fully connected operation with all the feature vector elements output by the pooling layer to integrate feature information and comprehensively integrate image features to predict joint point coordinates. The input of a neuron is ;in To connect The input neuron The weight of a neuron, For the The weights and biases of neurons are learned through a large amount of labeled data during the model pre-training process. By learning the mapping relationship between image features and joint coordinates in a large amount of labeled data, the feature vectors are converted into predicted values ​​of joint coordinates, and finally accurate athlete posture estimation results are obtained, including the coordinate position information of each joint in the image.

[0059] This information will serve as the basic data for subsequent biomechanical calculations, and will be used to further construct motion energy flow trajectory map operations, using the learned mapping relationship to accurately predict joint point coordinates, providing an accurate posture data foundation for subsequent complex analysis.

[0060] The energy flow trajectory extraction module 21 calculates the kinetic energy and potential energy of each part of the athlete's body based on the biomechanical principles, as follows:

[0061] First, the coordinates of the joint points of the same body part in the continuous video frames are obtained. Taking the elbow joint of the athlete's arm as an example, the coordinates of the joint points in the first Frame coordinates and Coordinates , it reflects the movement status through the position changes of joint points in adjacent frames, can accurately capture the details of limb movement, and provide an accurate data source for calculating instantaneous speed.

[0062] Calculate the displacement vector of the part between two frames based on the coordinate change ; This step can intuitively reflect the movement direction and distance of the joint point in three-dimensional space, which helps to improve the accuracy of subsequent speed calculations and reduce error accumulation.

[0063] Calculate the time interval between two frames , calculate the instantaneous velocity vector based on the displacement vector and time interval In this way, the movement speed of body parts at a specific moment can be accurately obtained, the movement in three-dimensional space can be taken into account, the movement state of limbs can be fully reflected, and the core parameters can be provided for kinetic energy calculation.

[0064] The mass parameters of each body part are known According to the kinetic energy formula Calculating the kinetic energy value of the part provides basic data for subsequent analysis of the athlete's sports energy distribution, and by calculating the kinetic energy of each part, a comprehensive understanding of the overall sports energy situation can be obtained.

[0065] Take the ground as the reference plane and determine the height of the body parts relative to the reference plane through the coordinate information of the joint points , taking the athlete's ankle joint as an example, its coordinates are , then its height ,A clear and unified reference plane can ensure the consistency and accuracy of potential energy calculations, make the potential energy comparison of different body parts meaningful, and facilitate the subsequent energy flow map construction.

[0066] For the same body part in different frames, the above height determination process is repeated to obtain its height change. This can track the vertical position change of the body part during movement, accurately reflect the increase or decrease of potential energy, and provide a basis for accurately calculating the potential energy value.

[0067] According to the potential energy formula , combined with known quality parameters and the calculated height and the acceleration due to gravity , calculate the potential energy value of each part. This formula is based on the basic principle of gravitational potential energy and can accurately quantify the potential energy changes caused by changes in the relative height of body parts. It plays an important role in analyzing the energy conversion and overall energy distribution of athletes during exercise.

[0068] The calculated potential energy values ​​of each body part at different times are integrated with the previously calculated kinetic energy values ​​in a time series, with time as the horizontal axis and the kinetic energy and potential energy values ​​as different vertical axis dimensions. An energy flow data sequence describing the energy changes during the athlete's movement is constructed. This can intuitively display the dynamic changes of the athlete's energy over time during the entire movement process, clearly present the conversion relationship between kinetic energy and potential energy, and provide a comprehensive, accurate and orderly data basis for constructing a sports energy flow trajectory map.

[0069] The steps of constructing the motion energy flow trajectory map in the energy flow trajectory extraction module 21 are as follows:

[0070] Time is used as the horizontal axis because the athlete's movement process progresses over time. The time dimension can clearly show the sequence and duration of the movement. By accurately recording the time corresponding to each data point, it can accurately reflect the timing of energy changes during the movement, providing a basic framework for subsequent analysis of movement rhythm, the time point of energy peak, etc., making the energy changes of the entire movement process visual and quantifiable for analysis.

[0071] The kinetic energy and potential energy values ​​of each part of the body are used as the two vertical coordinate dimensions of the graph. The kinetic energy value reflects the energy state related to the intensity and speed of the movement of the body parts, and the potential energy value reflects the energy of the body parts due to their relative height position. Setting kinetic energy and potential energy as the vertical coordinate dimensions respectively can comprehensively display the two main forms of energy of body parts in motion and their mutual conversion relationship, which is helpful for in-depth analysis of the mechanical properties of motion and the laws of energy distribution.

[0072] For each body part, at different time points , get the corresponding kinetic energy value and potential energy value , mark these data points in a two-dimensional coordinate system, that is and ,Precisely marking these data points can accurately reflect the energy state of body parts at a specific moment, ,provide a basis for drawing accurate trajectory curves for subsequent connections, ,so that the curves can truly reflect the trend of energy changes over time, ,avoiding curve distortion caused by data marking errors, ,which affects the analysis and judgment of motion energy changes.

[0073] Arrange the kinetic energy data points of the same body part in chronological order. Connect them with line segments in sequence to form the kinetic energy flow trajectory curve of the body part. This curve can intuitively show the changing trend of the kinetic energy of the body part during the movement, which is helpful for analyzing the changing law of the movement speed and the force of the body part.

[0074] Similarly, the potential energy data points of the same body part They are connected in sequence to form a potential energy flow trajectory curve, which reflects the potential energy changes caused by the relative height changes of body parts during movement. For example, the potential energy first increases and then decreases in a jumping action. This curve is of great significance for understanding the vertical movement characteristics and energy conversion timing in exercise.

[0075] The kinetic energy flow trajectory curves and potential energy flow trajectory curves drawn for different body parts are drawn together in the same two-dimensional coordinate system. In this way, the entire graph can fully display the energy changes of various body parts of the athlete during exercise and their mutual relationships.

[0076] The interweaving and comparison of curves of different parts can analyze which parts are the main energy providers or consumers in a certain action, as well as the mode of energy transfer and collaborative work between the parts. This provides rich information for in-depth research on the mechanical principles and characteristics of sports movements, and also provides a comprehensive and detailed data presentation for comparison with historical advertising maps to determine the similarity of advertising content.

[0077] The data construction storage module 22 stores the motion energy flow trajectory map and related advertisement metadata in the advertisement material library and establishes an index structure.

[0078] The method for establishing the index structure in the data construction storage module 22 is as follows:

[0079] For the motion energy flow trajectory map, we first need to determine its eigenvector. The map can be discretized in the time dimension, that is, according to a fixed time interval Divide the time period, and count the mean, variance, maximum, minimum of the kinetic energy and potential energy of each body part in each time period, as well as the correlation coefficient of energy between different body parts as characteristic elements. These characteristic elements are combined to form the characteristic vector of the motion energy flow trajectory map. ; This way of extracting feature vectors can reduce the dimensionality of complex graph data, extract key information, reduce the complexity of data storage and calculation, and retain the main features of the graph, providing a suitable data form for subsequent efficient hash calculation and index construction, so that similar graphs have similar representations in the feature vector space, which facilitates rapid similarity screening through hash algorithms.

[0080] Select MurmurHash function as the hash function, which is used for the input feature vector , a fixed-length hash value is calculated through a series of complex bit operations and mixed operations ,The purpose of the hash function is to map different feature vectors to the hash value space as evenly as possible and reduce hash conflicts. Its formula is:

[0081] ;

[0082] in is a large prime number, The actual MurmurHash function is much more complex than this and contains more bit operations and optimization strategies. Through this hash calculation, the high-dimensional feature vector is converted into a low-dimensional hash value, which makes it possible to quickly locate and compare possible similar motion energy flow trajectory maps in a large-scale advertising material library, greatly improving the efficiency of data retrieval and avoiding the high computational cost of comparing the entire map one by one.

[0083] The calculated hash value As the index key, this is because the hash value can quickly locate the approximate range of graphs with similar features, and use the advertising metadata and the storage address of the graph in the advertising material library as the index value. When you need to query a historical advertising graph that is similar to a certain motion energy flow trajectory graph, you only need to calculate the hash value of the graph, and then quickly find the corresponding index key in the hash index table to obtain the relevant advertising metadata and storage address, and directly obtain the possibly similar advertising graph for further precise comparison.

[0084] The delivery comparison and avoidance module 23 then uses a method based on a combination of a dynamic time warping algorithm and energy distribution similarity calculation to compare the motion energy flow trajectory map with the historical advertising map in the advertising sample set, and determines whether to activate the intelligent avoidance mechanism based on the comparison result.

[0085] The steps of performing comparison in the delivery comparison avoidance module 23 using a method based on a combination of a dynamic time warping algorithm and energy distribution similarity calculation are as follows:

[0086] Assume that the time series of the motion energy flow trajectory map is , the time series of the historical advertising graph is The goal of the dynamic time warping algorithm is to find a mapping relationship of time series so that the difference between the two graphs in the time dimension is minimized by constructing a cost matrix , where the elements Indicates that No. Time point and No. The cost of aligning time points, ;in and are the energy value vectors of the body parts corresponding to the two time points, and then fill the cost matrix through dynamic programming to find the energy vector from the starting point To the end The minimum cost path is the optimal matching path. The algorithm can handle the situation where the length of the time series is inconsistent, and effectively find the best way to align the two spectra in time, which provides the correct correspondence in the time dimension for the subsequent accurate calculation of the similarity of energy distribution and avoids misjudgment caused by differences in the time axis.

[0087] Along the optimal matching path, for each corresponding time point pair, the kinetic energy distribution similarity is calculated. The kinetic energy value vector of the motion energy flow trajectory at a corresponding time point is assumed to be , is the number of body parts, and the kinetic energy value vector of the historical advertising graph is , using the similarity calculation formula based on Euclidean distance:

[0088] ;

[0089] in, The Euclidean distance calculated in this way can quantify the degree of difference between two kinetic energy value vectors in space. The distance is converted into similarity through the above formula, which is simple and intuitive and can effectively reflect the similarity of kinetic energy distribution, and a kinetic energy distribution similarity value between 0 and 1 is obtained. The closer to 1, the more similar the kinetic energy distribution is.

[0090] Similarly, for the calculation of potential energy distribution similarity, let the potential energy value vector of the motion energy flow trajectory at the corresponding time point be , is the number of body parts, and the potential energy value vector of the historical advertising graph is , according to the above kinetic energy similarity calculation formula, the potential energy similarity is obtained This calculation can accurately measure the similarity of potential energy distributions, complement the kinetic energy distribution similarity calculation, and comprehensively evaluate the similarity of energy distributions of the two maps.

[0091] Comprehensive kinetic energy distribution similarity Similarity to potential energy distribution The total energy distribution similarity is obtained. The calculation formula of energy distribution similarity is: ,in is the weight of kinetic energy similarity. This calculation can comprehensively consider the role of kinetic energy and potential energy in judging the similarity of advertisements, and evaluate the energy similarity of two graphs more comprehensively and accurately.

[0092] Setting thresholds based on historical experience , the calculated total energy distribution similarity With threshold For comparison, if > , it is determined to be similar advertising content and the intelligent avoidance mechanism is activated. This means that the advertisement has a high similarity with the historical advertisement in terms of motion energy flow characteristics, and there may be problems of content duplication or high similarity, which requires avoidance processing. , it is determined to be non-similar advertising content and the intelligent avoidance mechanism is not activated, indicating that the advertisement has significant differences from historical advertisements in terms of motion energy flow characteristics. It can be delivered normally or further advertising operations can be carried out. In this way, the advertising content can be effectively screened and managed, and the quality and diversity of advertising delivery can be improved.

[0093] The self-learning adaptive unit 3 dynamically optimizes the motion energy flow trajectory extraction algorithm and the similarity calculation model based on the feedback data of the comparison results in the actual advertising delivery effect, and updates the historical advertising data in the database according to the optimization results.

[0094] The self-learning adaptive unit 3 uses a gradient descent-based optimization algorithm when dynamically optimizing the motion energy flow trajectory extraction algorithm, as follows:

[0095] Construct a loss function based on the proportion of ads that are not correctly identified as the same type in the advertising effect feedback data. ,set up It is the actual ad type judgment result, 1 for correct and 0 for wrong. is the prediction result obtained by the current motion energy flow trajectory extraction algorithm. The loss function can be expressed using the cross entropy loss formula:

[0096] ;

[0097] The loss function can quantify the degree of difference between the current algorithm's prediction results and the actual results, and provides a clear goal for the optimization algorithm, which is to minimize this difference. This provides a measurable standard for the subsequent optimization process to guide the algorithm to adjust in a more accurate direction.

[0098] Parameters in the motion energy flow trajectory extraction algorithm , calculate the loss function About these parameters Gradient This requires the use of the chain rule to gradually derive the specific mathematical expression of the motion energy flow trajectory extraction algorithm. If the calculation of a certain step in the motion energy flow trajectory extraction algorithm is , is the input data, and the loss function ,So The gradient represents the rate of change of the loss function in the parameter space, and points out the direction in which the parameters should be adjusted. The parameters can be effectively updated based on this direction information to reduce the loss function value. The gradient value corresponding to each parameter is obtained, which provides key adjustment direction information for parameter updating.

[0099] According to the calculated gradient , according to the update formula of the gradient descent algorithm Update the parameters, where is the learning rate, which is used to control the step size of parameter update. The above steps of defining loss function, calculating gradient and updating parameters are repeated until the change of loss function in two adjacent iterations is less than a set threshold. Multiple iterations can make the algorithm continuously approach the optimal parameter setting. With the progress of iteration, the algorithm performance is gradually improved, and the extraction of motion energy flow trajectory is more accurate. Finally, an optimized motion energy flow trajectory extraction algorithm is obtained, which can better adapt to the changes in advertising data and improve the accuracy of avoiding the same type of advertising content.

[0100] In the present invention, the action acquisition and processing unit 1 extracts advertising samples from the advertising material library and performs preprocessing. The action analysis and avoidance unit 2 uses the human posture estimation algorithm and biomechanical principles to accurately calculate the kinetic energy and potential energy of the athlete's body parts, and constructs a motion energy flow trajectory map, and sets a hash index. By comparing the new map with the historical map, it is determined whether to trigger the intelligent avoidance mechanism. The self-learning and adaptive unit 3 optimizes the algorithm model and updates the historical data based on the feedback data, so as to improve the system's ability to accurately identify the athlete's actions in the video advertisements of the same type of sports events.

[0101] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An advertising operation system for avoiding the same type of advertising content, characterized in that: It includes a motion collection and processing unit (1), a motion analysis and avoidance unit (2), and a self-learning and adaptive unit (3); The action collection and processing unit (1) is used to collect sports event video advertisement samples containing athlete actions from an advertisement material library, and pre-process the samples to obtain an advertisement sample set; The motion analysis and avoidance unit (2) calculates the kinetic energy and potential energy of each part of the athlete's body during the motion process based on the mass parameters, joint coordinates and limb angle information of each part of the athlete's body through a human posture estimation algorithm and biomechanical principles, and constructs a motion energy flow trajectory map based on this, stores the motion energy flow trajectory map and related advertising metadata in an advertising material library, and establishes an index structure, and then uses a method based on a combination of a dynamic time warping algorithm and energy distribution similarity calculation to compare the motion energy flow trajectory map with historical advertising maps in an advertising sample set, and determines whether to start an intelligent avoidance mechanism based on the comparison result; The self-learning adaptive unit (3) dynamically optimizes the motion energy flow trajectory extraction algorithm and the similarity calculation model based on the feedback data of the comparison results in the actual advertising delivery effect, and updates the historical advertising data in the database according to the optimization results.

2. The advertising operation system for avoiding the same type of advertising content according to claim 1, characterized in that: The action acquisition processing unit (1) uses an image enhancement algorithm to enhance the video advertisement sample image during the preprocessing process, as follows: Convert the video ad sample image from RGB color space to grayscale image, count the number of pixels at each grayscale level in the grayscale image, create an array of length 256, calculate the cumulative distribution function based on the array, and then calculate the mapping table based on the cumulative distribution function; The grayscale image is traversed again, and the original grayscale value of each pixel is replaced by the value in the mapping table as a new grayscale value, and the new grayscale value is used to enhance the image.

3. The advertising operation system for avoiding the same type of advertising content according to claim 2, characterized in that: The motion analysis and avoidance unit (2) comprises an energy flow trajectory extraction module (21), wherein the human body posture estimation algorithm in the energy flow trajectory extraction module (21) adopts a convolutional neural network model based on deep learning, and the specific steps are as follows: Extract sports event video advertisement samples from the advertisement sample set, and perform frame-by-frame segmentation to obtain a video frame sequence; Each frame of image is input into a pre-trained convolutional neural network, which includes a convolution layer, a pooling layer and a fully connected layer. The convolution layer uses a convolution kernel to slide on the image to perform a convolution operation, and obtains feature maps of different levels by weighted summation of the convolution kernel and image pixels. The pooling layer uses a maximum pooling operation to slide a window of a fixed size on the feature map, and selects the maximum value in the window as the output after pooling. The fully connected layer performs a fully connected operation on the feature vector output by the pooling layer, converts the feature vector into a predicted value of the joint point coordinates, and finally obtains an accurate athlete's body posture estimation result.

4. The advertising operation system for avoiding the same type of advertising content according to claim 3, characterized in that: The energy flow trajectory extraction module (21) calculates the kinetic energy and potential energy of each part of the athlete's body based on the principles of biomechanics, specifically as follows: For kinetic energy calculation, according to the formula ;in are the mass parameters of various parts of the body, The instantaneous speed of the part is calculated according to the change of the joint point coordinates. The speed value is obtained by differential calculation of the joint point coordinates in consecutive frames, and then the kinetic energy value of each part at different times is calculated according to the speed value; For potential energy calculation, according to the formula ;in is the height of the body parts relative to the ground plane. The height change of each part is determined by the coordinate information of the joint points, and the potential energy value is calculated in combination with the mass parameter; The calculated kinetic energy and potential energy values ​​are then integrated according to the time series to construct an energy flow data sequence that describes the energy changes during the athlete's movement, which serves as the basic data for constructing the sports energy flow trajectory map.

5. The advertising operation system for avoiding the same type of advertising content according to claim 4, characterized in that: The steps of constructing a motion energy flow trajectory map in the energy flow trajectory extraction module (21) are specifically as follows: The kinetic energy and potential energy values ​​of each part of the body are used as the two dimensional data of the map, and time is used as the horizontal axis to construct a two-dimensional energy flow trajectory map; The kinetic energy and potential energy values ​​of each body part at different time points are marked and connected to form the energy flow trajectory curve of the part; the curves of different body parts together constitute a complete motion energy flow trajectory map, and metadata information related to the advertisement is also annotated in the map; The metadata information related to the advertisement includes the brand to which the advertisement belongs, the theme of the advertisement, and the time of delivery.

6. The advertising operation system for avoiding the same type of advertising content according to claim 5, characterized in that: The action analysis avoidance unit (2) comprises a data construction storage module (22), and a method for establishing an index structure in the data construction storage module (22) is specifically as follows: The indexing method based on hash algorithm is used to perform hash calculation on the feature vector of the motion energy flow trajectory map to obtain the corresponding hash value; A hash index table is constructed by using the hash value as the index key and the storage address of the advertising metadata and the graph in the advertising material library as the index value.

7. The advertising operation system for avoiding the same type of advertising content according to claim 5, characterized in that: The motion analysis avoidance unit (2) comprises a delivery comparison avoidance module (23), wherein the delivery comparison avoidance module (23) performs a comparison step using a method based on a combination of a dynamic time warping algorithm and energy distribution similarity calculation, which is specifically as follows: The dynamic time warping algorithm is used to warp the time axis of the motion energy flow trajectory map and the historical advertising map in the advertising sample set to find the optimal matching path between the two maps in the time dimension. Along the optimal matching path, the energy distribution similarity at the corresponding time is calculated. For the energy distribution similarity calculation, the similarity calculation formula based on Euclidean distance is used to calculate the kinetic energy distribution similarity and the potential energy distribution similarity respectively, and the total energy distribution similarity is obtained by combining the two. A similarity threshold is set, and the calculated total energy distribution similarity is compared with the threshold. If the similarity is greater than the threshold, it is determined to be similar advertising content and the intelligent avoidance mechanism is activated. If the similarity is less than or equal to the threshold, it is determined to be non-similar advertising content.

8. The advertising operation system for avoiding the same type of advertising content according to claim 1, characterized in that: The self-learning adaptive unit (3) adopts an optimization algorithm based on gradient descent when dynamically optimizing the motion energy flow trajectory extraction algorithm, which is specifically as follows: According to the proportion of the advertisements that are not correctly identified as the same type in the feedback data of the advertising effect, a loss function is constructed, and the gradient of the loss function with respect to the parameters in the motion energy flow trajectory extraction algorithm is calculated; According to the gradient, the parameters are updated according to the update formula of the gradient descent algorithm, and the above steps of constructing the loss function are repeated. For the parameters in the motion energy flow trajectory extraction algorithm, the gradient of the loss function with respect to the parameters is calculated until the stopping condition is met. The stopping condition is that the change in the loss function between two adjacent iterations is less than a set threshold.

Images