A high-definition video stream pixel tracking training method based on AI application

Abstract

The present invention relates to the field of pixel tracking technology, specifically a method for training pixel tracking of high-definition video streams based on AI applications, comprising the following steps: based on the set video input parameters, a DCNN network is used to analyze the texture, edge and color distribution of the video frame, and the convolution kernel size and shape are adjusted to match the changes in video features to obtain feature data mapping. In the present invention, high-definition video streams are processed by combining deep convolution and recursive neural networks to achieve fine tracking and efficient analysis. The accuracy and speed of pixel tracking are significantly improved by refining local and global information in high-definition video streams and adjusting dynamic parameters. Especially in rapidly changing video scenes, the color and texture changes of pixels can be accurately extracted and analyzed through frame-by-frame analysis, and the pixel tracking performance is optimized, so that when processing large-scale high-definition video data, real-time adjustments can be made to ensure the accuracy and timeliness of tracking results.

Description

Technical Field

[0001] The present invention relates to the field of pixel tracking technology, and in particular to a high-definition video stream pixel tracking training method based on AI application. Background Art

[0002] Pixel tracking is an image processing technology that is mainly used to analyze and resolve pixel-level changes in video content. It can identify and track the movement and changes of pixels between video frames and is widely used in motion detection, visual effects generation, and security monitoring. Pixel tracking technology can help improve the accuracy of video analysis. By tracking the movement trajectory of specific pixels or pixel groups, high-precision target tracking and behavior analysis can be achieved.

[0003] Among them, the AI ​​application high-definition video stream pixel tracking training method involves the use of artificial intelligence technology to optimize and execute the pixel tracking process. In this method, the AI ​​model is trained to automatically identify and track pixel changes in the video stream, with the aim of improving the speed and accuracy of pixel tracking. It has a wide range of uses, such as in the visual system of self-driving cars for real-time monitoring and response to changes in the surrounding environment, or in security monitoring systems for accurate identification of abnormal activities. AI-enabled pixel tracking technology can process large amounts of high-definition video data, providing faster processing speeds and higher analysis efficiency.

[0004] Existing technologies have problems with insufficient processing speed and tracking accuracy when processing high-definition video streams. In particular, when faced with high resolution and the increased amount of data it brings, traditional technologies find it difficult to accurately identify and track the movement of pixels in real time, especially when video scenes change rapidly. This technical limitation becomes a bottleneck in applications that require precise responses, such as real-time monitoring and dynamic interactions, where delays or tracking errors can lead to the loss of key information. In addition, existing methods also show shortcomings when adapting to complex lighting and dynamic scenes, and often affect the overall tracking effect due to failure to optimize, limiting their application in changing environments. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and propose a high-definition video stream pixel tracking training method based on AI application.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a high-definition video stream pixel tracking training method based on AI application, comprising the following steps:

[0007] S1: Based on the set video input parameters, the DCNN network is used to analyze the texture, edge and color distribution of the video frame, and the convolution kernel size and shape are adjusted to match the changes in video features to obtain feature data mapping;

[0008] S2: Based on the feature data mapping, integrate the features of the multi-resolution video stream, refine the local and global information in the video stream, and adjust the feature tracking parameters to match the changes of illumination and dynamic scenes to generate scene adaptation information;

[0009] S3: Based on the scene adaptation information, a recursive neural network is used to memorize the time series information of the video, perform frame-by-frame analysis on the video frames, extract fine-grained features of color changes and texture differences, and gradually refine the feature parsing process to generate a detail recognition configuration;

[0010] S4: Based on the detail recognition configuration, adjust the recursive neural network parameters to match the newly received video stream data, adjust the tracking parameters through a continuous recursive feature optimization process, refine the tracking responsiveness to future video frames, and generate a pixel tracking model.

[0011] The present invention is improved in that the steps of analyzing the texture, edge and color distribution of the video frame are specifically as follows:

[0012] S111: Based on the set video input parameters, use the DCNN network to receive the video frame, adjust and match the resolution and frame rate of the video stream, and obtain DCNN configuration information;

[0013] S112: Based on the DCNN configuration information, the formula is used:

[0014] ;

[0015] Get the visual feature extraction results ,in, is the convolutional layer weight, is the input video frame data, is the bias term, is the activation function;

[0016] S113: Analyze the visual feature extraction results, including the texture, edge and color distribution of the video frame, to obtain a video frame analysis result.

[0017] The present invention is improved in that the step of acquiring the characteristic data mapping is specifically as follows:

[0018] S121: Dynamically adjust the size and shape of the convolution kernel according to the changes in video content, using the formula:

[0019] ;

[0020] Get the adjusted convolution kernel parameters ,in, Indicates the rate of change of the speed of the object in the video. is the adjustment factor, and Represent the size and shape of the original convolution kernel respectively;

[0021] S122: Based on the adjusted convolution kernel parameters, the video frame is analyzed to match the changes in video features to obtain feature data mapping.

[0022] The present invention is improved in that the steps of refining the local and global information are specifically as follows:

[0023] S211: Based on the feature data mapping, integrating feature data of multi-resolution video streams, unifying feature formats and scales, and obtaining a feature data set;

[0024] S212: Based on the feature data set, refine the feature expression, using the formula:

[0025] ;

[0026] Get optimized feature details ,in, is a local feature, is a global feature, It is a coefficient used to adjust local and global features.

[0027] The present invention is improved in that the step of acquiring the scene adaptation information is specifically as follows:

[0028] S221: adjusting tracking parameters according to different lighting and dynamic scene conditions, where the dynamic scene conditions include object movement, environmental changes, and scene changes, and automatically adjusting exposure and contrast parameters by analyzing real-time video stream data to obtain environmental adjustment parameters;

[0029] S222: Based on the environment adjustment parameters, the response to the dynamic scene is optimized using the formula:

[0030] ;

[0031] Get scene adaptation information ,in, Represents the feature change caused by dynamic scene changes, Represents the ambient lighting characteristics, is the adjustment factor, is the balance coefficient.

[0032] The present invention is improved in that the step of memorizing the time series information is specifically as follows:

[0033] S311: Based on the scene adaptation information, using a recursive neural network, performing video sequence processing, optimizing the performance of capturing the time-dependent characteristics of the video, and obtaining an initial model configuration;

[0034] S312: Based on the initial model configuration, the video sequence data is input into a recursive neural network using the formula:

[0035] ;

[0036] Memorize and analyze time series information, where Indicates at time The memory state, is the memory state of the previous moment, is the feature input of the current frame, is the forgetting coefficient.

[0037] The present invention is improved in that the step of obtaining the detail recognition configuration is specifically as follows:

[0038] S321: Perform frame-by-frame analysis on the video frames to capture color changes and texture differences in each frame to obtain a basic visual feature set;

[0039] S322: Based on the basic visual feature set, feature refinement is performed using the formula:

[0040] ;

[0041] Gradually refine the feature parsing process to obtain refined visual features ,in, represents the edge strength extracted from the video frame, Indicates the amount of color change, Represents the complexity of texture information;

[0042] S323: Based on the refined visual features, feature fusion and optimization are performed to obtain detail recognition configuration.

[0043] The present invention is improved in that the step of obtaining the pixel tracking model is specifically as follows:

[0044] S411: Based on the detail recognition configuration, adjusting the recursive neural network parameters, matching the newly received video stream data, and generating an initial configuration of the network;

[0045] S412: Based on the initial configuration of the network, adjust the tracking parameters using the formula:

[0046] ;

[0047] Get new tracking parameters ,in, Indicates historical parameter settings. is the adjustment factor, and Represent the newly received and historical video feature data respectively;

[0048] S413: Continuously track changes in the video content according to the new tracking parameters, and refine the tracking responsiveness of future video frames to obtain a pixel tracking model.

[0049] Compared with the prior art, the advantages and positive effects of the present invention are:

[0050] In the present invention, high-definition video streams are processed by combining deep convolution and recursive neural networks to achieve fine tracking and efficient analysis. The accuracy and speed of pixel tracking are significantly improved by refining local and global information in high-definition video streams and adjusting dynamic parameters. Especially in rapidly changing video scenes, the color and texture changes of pixels can be accurately extracted and analyzed through frame-by-frame analysis, and the pixel tracking performance can be optimized. When processing large-scale high-definition video data, real-time adjustments can be made to ensure accurate and timely tracking results. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A flowchart of a high-definition video stream pixel tracking training method based on AI application is proposed for the present invention;

[0052] Figure 2 A flowchart of analyzing the texture, edge and color distribution of a video frame in the present invention;

[0053] Figure 3 It is a flowchart of obtaining feature data mapping in the present invention;

[0054] Figure 4 It is a detailed flow chart of local and global information in the present invention;

[0055] Figure 5 This is a flowchart for obtaining scene adaptation information in the present invention;

[0056] Figure 6 It is a memory flow chart of time series information in the present invention;

[0057] Figure 7 A flowchart for obtaining the detail recognition configuration in the present invention;

[0058] Figure 8 This is a flowchart for obtaining the pixel tracking model in the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0060] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating positions or positional relationships, are based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0061] Example:

[0062] See also Figure 1 The present invention provides a technical solution: a high-definition video stream pixel tracking training method based on AI application, comprising the following steps:

[0063] S1: Based on the set video input parameters, the DCNN network is used to analyze the texture, edge and color distribution of the video frame, and the convolution kernel size and shape are adjusted to match the changes in video features to obtain feature data mapping;

[0064] Matching video feature changes refers to the features within the video frame, such as texture, edge, and color distribution, which may change significantly in different video scenes and time points. For example, in the transition from day to night, from indoors to outdoors, or from static scenes to dynamic scenes. Convolution kernels are the basic elements for feature extraction. Larger convolution kernels may be more suitable for capturing texture information in larger areas, while smaller convolution kernels are suitable for capturing subtle edge details. Matching video feature changes requires not only identifying the features in the video frame, but also adjusting its own parameters to adapt to feature changes. The adjustment can be performed dynamically, that is, the network can automatically adjust the processing method of subsequent frames based on the feature changes identified in the previous frame or frames.

[0065] S2: Based on feature data mapping, it integrates the features of multi-resolution video streams, refines the local and global information in the video streams, and adjusts the feature tracking parameters to match the changes in lighting and dynamic scenes to generate scene adaptation information;

[0066] In multi-resolution video streams, detailed features can be extracted from high-resolution videos, while more representative global features can be extracted from low-resolution videos. Local information refers to the detailed features of smaller areas in video frames, such as object edges, textures, and color changes. Local details are very important for understanding specific elements in the scene (such as faces, vehicles, text, etc.). Global information involves the overall scene understanding of video frames, such as the layout of the scene, lighting conditions, and background elements, which helps to build the context of the video content so that local information can be correctly parsed in a larger framework. The adjustment of feature tracking parameters includes dynamically selecting parameters in the tracking algorithm such as window size, tracking speed, and response threshold to optimize the tracking effect.

[0067] S3: Based on the scene adaptation information, a recursive neural network is used to memorize the time series information of the video, perform frame-by-frame analysis on the video frames, extract fine-grained features of color changes and texture differences, and gradually refine the feature parsing process to generate detail recognition configurations;

[0068] S4: Based on the detail recognition configuration, adjust the recursive neural network parameters to match the newly received video stream data, adjust the tracking parameters through a continuous recursive feature optimization process, refine the tracking responsiveness to future video frames, and generate a pixel tracking model.

[0069] The feature optimization process refers to adjusting the weights and parameters in the network to better capture and track pixel changes in the video. Tracking parameters include settings that control how feature tracking responds to changes in the video, such as tracking speed, accuracy, and sensing range. By optimizing parameters, targets can be tracked more accurately and respond to factors such as changes in ambient lighting and scene switching.

[0070] Feature data mapping includes texture hierarchy, color spectrum, and edge intensity. Scene adaptation information includes illumination adaptation, dynamic change adaptability, and resolution matching index. Detail recognition configuration includes color sensitivity, texture perception, and dynamic capture capability. The pixel tracking model includes tracking stability, pixel response speed, and error rate control.

[0071] See also Figure 2 , the analysis steps for the texture, edge and color distribution of the video frame are as follows:

[0072] S111: Based on the set video input parameters, use the DCNN network to receive the video frame, adjust and match the resolution and frame rate of the video stream, and obtain DCNN configuration information;

[0073] When configuring a deep convolutional neural network model, you need to set the model's input layer parameters. The parameters need to be adjusted according to the specific resolution and frame rate of the video. In order to adapt to different video stream inputs, the size and scale of the input layer need to be dynamically calculated, which involves preliminary analysis of the original video data, including detecting the resolution and average frame rate of the video frames, and then adjusting the network configuration based on the data so that the model can receive and process input streams of different resolutions and frame rates. In order to ensure the consistency and efficiency of the model processing, the size and step size of the first-layer convolution kernel of the model are adapted at the same time to ensure that the area covered by the video frame can be maximized while retaining important visual information. Through this process, it can be ensured that the configured network model can adapt to various input conditions and generate model configuration results.

[0074] S112: Based on the DCNN configuration information, the formula is:

[0075] ;

[0076] Get the visual feature extraction results ,in, is the convolutional layer weight, is the input video frame data, is the bias term, It is an activation function used to enhance nonlinear feature representation and sparse activation;

[0077] If you have the following data, The value is 0.5, The input pixel value is 100, The value is 0.01, and it is substituted into the formula:

[0078] Convolution operation, calculation ,Right now ;

[0079] Add bias, bias Add to the above result, that is ;

[0080] Apply the ReLU function, , the ReLU function is defined as ;

[0081] because is greater than 0, so ;

[0082] The output after ReLU activation is 50.01, showing the transformation from the input pixel value to the output value after processing through the convolution kernel and then to ReLU activation. This operation actually corresponds to extracting the features of the input video frame through a single convolution kernel, and then using the activation function to increase the nonlinearity in order to capture more complex patterns, indicating that the activation function output maintains the main features of the input and is the basis for the next step of processing.

[0083] S113: Analyze the visual feature extraction results, including the texture, edge, and color distribution of the video frame, to obtain a video frame analysis result;

[0084] In the visual feature extraction results, the analysis involves comprehensive processing of the extracted feature information, including texture, edge and color distribution. It is necessary to further statistically analyze the feature matrix output from the convolutional neural network. According to the spatial distribution and intensity of the features, cluster analysis is used to identify different feature areas, and the texture and edge features of each area are encoded. The color features are analyzed by calculating the color histogram in the area. This can not only describe the visual content of each video frame in detail, but also effectively classify and mark the video frames. The obtained video frame analysis results will be directly used for advanced parsing and indexing of video content.

[0085] See also Figure 3 , the specific steps for obtaining feature data mapping are:

[0086] S121: Dynamically adjust the size and shape of the convolution kernel according to the changes in video content, using the formula:

[0087] ;

[0088] Get the adjusted convolution kernel parameters ,in, Indicates the rate of change of the speed of the object in the video. is the adjustment coefficient to make the convolution kernel capture dynamic features more accurately. It represents the exponential decay factor based on the adjustment of motion changes, reflecting the influence of motion speed on the adjustment of convolution kernel. and Represent the size and shape of the original convolution kernel respectively;

[0089] set up Indicates the rate of change of the object's velocity, To adjust the coefficients to enhance the sensitivity of the convolution kernel to rapidly changing features, the original size of the convolution kernel is and shape Keeping it constant, calculate the adjustment factor:

[0090] ;

[0091] Then, calculate the new convolution kernel size:

[0092] ;

[0093] Since the convolution kernel size must be an integer, we get , the shape remains unchanged. This result shows that by adjusting the size and shape of the convolution kernel, it is possible to more accurately capture rapidly changing features in the video, such as fast-moving objects, thereby improving the accuracy of feature recognition in dynamic scenes.

[0094] S122: Analyze the video frame based on the adjusted convolution kernel parameters, match the changes of video features, and obtain feature data mapping;

[0095] The video frames are analyzed based on the adjusted convolution kernel parameters. For the feature matching and data mapping process in the video frames, the adaptability to dynamic changes in video content is enhanced by using new convolution kernel parameters. Further analysis includes edge detection, texture recognition, and color distribution statistics. Feature extraction is crucial for subsequent scene analysis and object recognition. After being processed by the mapping algorithm, the feature data is mapped into a structured feature space, which is specially designed for subsequent applications such as tracking, recognition or classification, to obtain the feature data mapping results.

[0096] See also Figure 4 , the refinement steps of local and global information are as follows:

[0097] S211: Based on feature data mapping, integrate feature data of multi-resolution video streams, unify feature formats and scales, and obtain feature data sets;

[0098] Integrate feature data from multi-resolution video streams, adjust the input of the video stream to capture data of different resolutions from low to high, and unify the data format and scale in order to create a feature dataset. The dataset will include information extracted from various video resolutions, involving decoding the original video stream and then performing feature extraction, such as edge detection, texture analysis, and color distribution, to generate a feature dataset containing a wide range of information, providing the necessary foundation for the next step of feature refinement and scene analysis.

[0099] S212: Based on the feature data set, refine the feature expression using the formula:

[0100] ;

[0101] Get optimized feature details ,in, is a local feature, is a global feature, It is a coefficient used to adjust local and global features.

[0102] set up The value of is 0.6, which represents the strength of local features. The value of is 0.4, which represents the strength of the global feature. The value is 0.7, which means that more emphasis is placed on local information. Substitute it into the formula for calculation:

[0103] ;

[0104] The results show that the integrated feature data tends to be local details, which can more accurately identify and parse key objects and events in the video when performing scene analysis, and improve the ability to handle dynamic scenes.

[0105] See also Figure 5 , the specific steps for obtaining scene adaptation information are:

[0106] S221: Adjust tracking parameters for different lighting and dynamic scene conditions. Dynamic scene conditions include object movement (such as the presence and movement of pedestrians, vehicles or other moving objects. The speed, direction and size of dynamic elements are important considerations), environmental changes (such as the alternation of day and night, instantaneous light changes caused by cloud cover, weather changes, etc.) and scene changes (changes in camera perspective, rapid changes in background, etc.). By analyzing real-time video stream data, the exposure and contrast parameters are automatically adjusted to obtain environmental adjustment parameters.

[0107] When adjusting the feature tracking parameters, it includes identifying the dynamic elements and ambient lighting conditions in the current scene, collecting data through on-site camera equipment, analyzing to obtain the initial parameters of motion features and ambient lighting, and quantifying the parameters through real-time analysis. For example, the motion speed is determined by the inter-frame difference algorithm, and the lighting conditions are obtained through image brightness analysis. According to the quantized parameters, the weight coefficients and adaptive strength parameters in feature tracking are further determined. The basis for setting weights and parameters includes the evaluation of scene dynamics and the measurement of ambient lighting intensity. To ensure the accuracy of parameter settings, each parameter must be calibrated through multiple scene simulations to reflect the feature changes under different scene conditions. Through the parameter adjustment process, the device can adjust the tracking process in real time and optimize the response to various scene conditions.

[0108] S222: Adjust parameters based on the environment to optimize the response to dynamic scenes, using the formula:

[0109] ;

[0110] Get scene adaptation information ,in, Represents the amount of feature change caused by dynamic scene changes, including changes in motion speed or direction, etc. Represents the ambient lighting characteristics, including factors such as light intensity and light quality. is the adjustment coefficient used to adjust the sensitivity of scene adaptation information. is the balance coefficient, which is used to determine the dynamic scene feature changes With ambient lighting Relative contribution to the scene;

[0111] If you have the following data, Represents feature changes induced by dynamic scenes, such as moving objects or camera motion, represents the ambient lighting characteristics, such as natural light or indoor lighting, represents the weight of the dynamic scene, Adjust the adaptation strength to enhance or reduce the overall scene adaptation effect. The calculation process is:

[0112] ;

[0113] Calculated A weighted adaptation strength is shown, in which the contribution of dynamic scenes to scene adaptation is greater than that of ambient lighting, indicating that dynamic factors are given priority to adjust scene adaptation, which is especially effective for scenes with frequent dynamic changes.

[0114] See also Figure 6 , the specific steps for memorizing time series information are:

[0115] S311: Based on the scene adaptation information, a recursive neural network is used to process the video sequence, optimize the performance of capturing the time-dependent characteristics of the video, and obtain an initial model configuration;

[0116] Initialize the basic configuration of the recursive neural network and set the input layer of the network to receive video data. The configuration includes that the number of input nodes is equal to the feature dimension of the video frame. In addition, the number of hidden layer nodes of the network needs to be set, which depends on the complexity of the video content and the depth of time dependency that needs to be captured. For example, in order to effectively process time series information, selecting an appropriate number of hidden layer nodes can enhance the network's memory capacity and ability to handle complex dynamic changes. In addition, appropriately adjust the learning rate and activation function to ensure that the network can quickly and effectively update its state when receiving new video frames to adapt to the continuity and variability of the video content, provide a basis for the next step of deep time series analysis, and generate network initialization results.

[0117] S312: Based on the initial model configuration, the video sequence data is input into the recurrent neural network using the formula:

[0118] ;

[0119] Memorize and analyze time series information, where Indicates at time The memory state, is the memory state of the previous moment, is the feature input of the current frame, is the forgetting coefficient, which is used to control the proportion of the previous memory state retained by the network when new information arrives;

[0120] If the video frame is at time Features for , the previous time point Memory status for , forgetting coefficient Set to 0.6, according to the formula, the new memory state The calculation is as follows:

[0121] ;

[0122] This result shows that the memory state of the new calculation It more closely reflects the reasonable fusion of the characteristics of the current video frame and the previous memory, and provides a more coherent basis for time series analysis, so that the dynamic changes of video content can be tracked more accurately.

[0123] See also Figure 7 , the specific steps for obtaining the detail recognition configuration are:

[0124] S321: Perform frame-by-frame analysis on the video frames, capture the color changes and texture differences in each frame, obtain a basic visual feature set, and record features such as edge intensity and color spectrum distribution;

[0125] Performing the task of analyzing video frames frame by frame and identifying color changes and texture differences in each frame involves rapid parsing of video data streams to capture key visual information, such as edge clarity changes and color transition patterns. Through the data, a basic visual feature set is constructed, which contains the main visual elements that affect the expression of video content, such as edge intensity and color spectrum distribution. The purpose is to encode basic information so that it can be used for more refined visual feature analysis in subsequent steps. Through this process, not only the efficiency of data processing is improved, but also a basis for deep analysis and feature extraction is provided.

[0126] S322: Based on the basic visual feature set, feature refinement is performed using the formula:

[0127] ;

[0128] Gradually refine the feature parsing process to obtain refined visual features ,in, represents the edge strength extracted from the video frame, Indicates the amount of color change, Represents the complexity of texture information;

[0129] If the following data are available, the edge strength 5, the amount of color change 9, the complexity of texture information is 3, calculate the edge strength part:

[0130] ;

[0131] Calculate the color change part:

[0132] ;

[0133] Calculate the complexity of texture information:

[0134] ;

[0135] Add the above results to get the refined visual features:

[0136] ;

[0137] This result shows that the visual features processed by the refinement It is 4.38, indicating that the data after feature extraction and processing has higher analysis and recognition value.

[0138] S323: Based on the refined visual features, feature fusion and optimization are performed to obtain detail recognition configuration;

[0139] Based on the refined visual features, further feature fusion and optimization operations are performed, involving the combination of multiple feature dimensions, including comprehensive analysis of edge strength, texture complexity and color spectrum. Through deep learning or machine learning algorithms, visual information is optimized to form an efficient and reliable detail recognition configuration, which can be applied to deep analysis of video content, such as automatic labeling and scene recognition, providing strong data support for subsequent visual tasks.

[0140] See also Figure 8 , the specific steps for obtaining the pixel tracking model are:

[0141] S411: Based on the detail recognition configuration, adjust the recursive neural network parameters, including the number of layers and the number of neurons in each layer, match the newly received video stream data, and generate an initial configuration of the network;

[0142] The implementation of detail recognition configuration involves adjusting the number of layers and the number of neurons in each layer of the recursive neural network (RNN) to match the newly received video stream data. The number of layers and neurons of the network is determined based on the characteristics of the input data and the processing requirements to ensure that the network can effectively process the received video data. The initial configuration of the network is generated to ensure that the network has a good processing foundation when receiving new input data to achieve efficient video data processing.

[0143] S412: Based on the initial configuration of the network, adjust the tracking parameters using the formula:

[0144] ;

[0145] Get new tracking parameters ,in, Indicates historical parameter settings. is an adjustment factor used to control the speed of parameter update, It is used to process the difference between new and old features. data, and Represent the newly received and historical video feature data respectively;

[0146] If you have the following data ,but:

[0147] ;

[0148] Assumptions (Sigmoid function), then:

[0149] ;

[0150] The result shows that the new tracking parameter is 0.5622, which reflects the smooth transition from old features to new features and enhances the model's adaptability to changes in video content.

[0151] S413: continuously tracking changes in the video content according to the new tracking parameters, and refining the tracking responsiveness of future video frames to obtain a pixel tracking model;

[0152] According to the new tracking parameters, changes in the video content are continuously tracked, and the tracking responsiveness of future video frames is continuously refined by adjusting the tracking parameters. This involves analyzing the dynamic changes of the video content to ensure tracking accuracy and efficiency. The resulting pixel tracking model can more accurately predict and track key pixels and objects in the video content, thereby improving the performance of the video surveillance and analysis system.

[0153] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A high-definition video stream pixel tracking training method based on AI application, characterized in that: The following steps are involved: Based on the set video input parameters, the DCNN network is used to analyze the texture, edge and color distribution of the video frame, and the convolution kernel size and shape are adjusted to match the changes in video features to obtain feature data mapping; Based on the feature data mapping, the features of the multi-resolution video stream are integrated, the local and global information in the video stream is refined, and the feature tracking parameters are adjusted to match the changes of the illumination and dynamic scenes, and the scene adaptation information is generated; Based on the scene adaptation information, a recursive neural network is used to memorize the time series information of the video, perform frame-by-frame analysis on the video frames, extract fine-grained features of color changes and texture differences, and gradually refine the feature parsing process to generate a detail recognition configuration; Based on the detail recognition configuration, adjusting recursive neural network parameters to match newly received video stream data, adjusting tracking parameters through a continuous recursive feature optimization process, refining tracking responsiveness to future video frames, and generating a pixel tracking model; The steps of analyzing the texture, edge and color distribution of the video frame are specifically as follows: Based on the set video input parameters, the DCNN network is used to receive video frames, adjust and match the resolution and frame rate of the video stream, and obtain DCNN configuration information; Based on the DCNN configuration information, the formula is adopted: ； Get the visual feature extraction results ,in, is the convolutional layer weight, is the input video frame data, is the bias term, is the activation function; Analyzing the visual feature extraction results, including texture, edge and color distribution of the video frame, to obtain a video frame analysis result; The steps of acquiring the feature data mapping are specifically as follows: According to the changes in video content, the size and shape of the convolution kernel are dynamically adjusted using the formula: ； Get the adjusted convolution kernel parameters ,in, Indicates the rate of change of the speed of the object in the video. is the adjustment factor, and Represent the size and shape of the original convolution kernel respectively; Based on the adjusted convolution kernel parameters, the video frames are analyzed to match the changes in video features to obtain feature data mapping; The steps of refining the local and global information are specifically as follows: Based on the feature data mapping, the feature data of multi-resolution video streams are integrated, the feature format and scale are unified, and a feature data set is obtained; Based on the feature data set, the feature expression is refined and the formula is adopted: ； Get optimized feature details ,in, is a local feature, is a global feature, It is the coefficient used to adjust the local and global features; The steps of acquiring the scene adaptation information are specifically as follows: Adjust tracking parameters for different lighting and dynamic scene conditions. Dynamic scene conditions include object movement, environmental changes, and scene changes. By analyzing real-time video stream data, automatically adjust exposure and contrast parameters to obtain environmental adjustment parameters. Based on the environment adjustment parameters, the response to dynamic scenes is optimized using the formula: ； Get scene adaptation information ,in, Represents the feature change caused by dynamic scene changes, Represents the ambient lighting characteristics, is the adjustment factor, is the balance coefficient; The specific steps of memorizing the time series information are: Based on the scene adaptation information, a recursive neural network is used to process the video sequence, optimize the performance of capturing the time-dependent characteristics of the video, and obtain an initial model configuration; Based on the initial model configuration, the video sequence data is input into the recurrent neural network using the formula: ； Memorize and analyze time series information, where Indicates at time The memory state, is the memory state of the previous moment, is the feature input of the current frame, is the forgetting coefficient; The steps for obtaining the detail recognition configuration are specifically as follows: Performing frame-by-frame analysis on the video frames to capture color changes and texture differences in each frame to obtain a basic visual feature set; Based on the basic visual feature set, feature refinement is performed using the formula: ； Gradually refine the feature parsing process to obtain refined visual features ,in, represents the edge strength extracted from the video frame, Indicates the amount of color change, Represents the complexity of texture information; Based on the refined visual features, feature fusion and optimization are performed to obtain a detail recognition configuration; The steps for obtaining the pixel tracking model are specifically as follows: Based on the detail recognition configuration, adjusting the recursive neural network parameters to match the newly received video stream data and generate an initial configuration of the network; Based on the initial configuration of the network, the tracking parameters are adjusted using the formula: ； Get new tracking parameters ,in, Indicates historical parameter settings. is the adjustment factor, and Represent the newly received and historical video feature data respectively, It is used to process the difference between new and old features. data; According to the new tracking parameters, changes in the video content are continuously tracked, and the tracking responsiveness of future video frames is refined to obtain a pixel tracking model.

Images