Video transcoding control method adaptive to resolution and code rate
By deploying a multi-dimensional fusion decision system on the client side, and utilizing a linear Kalman filter and a resource constraint model, real-time optimization of video transcoding parameters was achieved, solving the problems of response lag and resource waste in existing technologies, and improving video smoothness and device battery life.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-03-13
Smart Images
Figure CN121665072A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer technology, specifically relating to a video transcoding control method that adapts to resolution and bitrate. Background Technology
[0002] With the widespread adoption of video streaming services, Adaptive Bitrate (ABR) technology has become a core means of ensuring user experience. The basic principle of this technology is that the client dynamically selects video segments with different bitrates for playback based on historical network throughput to cope with bandwidth fluctuations.
[0003] However, existing ABR mechanisms still have significant shortcomings in practical applications. First, most schemes rely on long historical observation windows (e.g., 5-10 seconds) for bandwidth prediction, resulting in a delayed response to sudden network fluctuations and easily causing stuttering in scenarios with instantaneous signal changes on the order of hundreds of milliseconds. Second, existing transcoding control often decouples resolution and bitrate, reducing the bitrate while maintaining the original resolution when bandwidth decreases, causing significant blockiness and image quality degradation in high-dynamic content at low bitrates. Furthermore, existing methods generally lack awareness of the device's own status (e.g., CPU temperature, battery level), failing to adaptively reduce computational load when device resources are strained, thus affecting battery life and stability.
[0004] Although some improvement solutions have attempted to introduce more complex prediction models or quality assessment mechanisms, problems such as high inference latency, large computational overhead, or reliance on cloud collaboration still exist, making it difficult to achieve unified control with low latency, lightweight and adaptive capabilities on the terminal side.
[0005] Therefore, existing technologies have limitations in dealing with short-cycle network fluctuations, achieving resolution-bitrate collaborative optimization, and adapting to terminal resource status. There is an urgent need for a transcoding control method that can achieve millisecond-level response, content awareness, and device adaptation on the client side. Summary of the Invention
[0006] To address the aforementioned technical problems, this invention provides a video transcoding control method that adapts to resolution and bitrate. This invention achieves real-time, joint, and adaptive optimization of video transcoding parameters by deploying a multi-dimensional fusion decision system on the client side, consisting of network state prediction, video content complexity analysis, and device resource constraint modeling.
[0007] The method proposed in this invention first uses a linear Kalman filter to predict the instantaneous network throughput at the millisecond level, in order to overcome the response lag caused by relying on long-period historical data in the prior art; secondly, it analyzes the spatial and temporal complexity of the video stream to be transcoded in real time and establishes a quantitative relationship between content characteristics and encoding resource consumption; at the same time, it establishes a resource constraint model based on the device hardware status, incorporating physical parameters such as central processing unit temperature and battery power into the decision boundary.
[0008] Finally, on a pre-built bitrate-resolution-quality multidimensional mapping manifold, a constrained optimization search is performed based on the predicted network bandwidth, video content complexity, and current device resource constraints to determine a bitrate and resolution combination that maximizes visual quality while ensuring smoothness and does not exceed the device's carrying capacity, and then instructs the transcoding engine to execute it.
[0009] An adaptive resolution and bitrate video transcoding control method, which includes: Through the client's built-in multi-sensor interface, a multi-dimensional state vector is periodically collected, which includes network status data, video content data, and device physical status data. Based on network state data, a state-space model is used to predict the available network bandwidth within a future transmission time window, generating a predicted bandwidth value. Real-time calculations are performed on video content data to extract content feature parameters that characterize the dynamics and texture complexity of video images; Based on the device's physical status data, the client hardware operating status is mapped to a preset resource constraint level; In a pre-built multi-dimensional quality assessment model that associates transcoding resolution, transcoding bitrate, video content features, and encoded video quality, an optimal combination of target resolution and target bitrate is solved by combining predicted bandwidth values, content feature parameters, and resource constraint levels. The optimal combination is sent as a transcoding instruction to the video transcoding engine to control the transcoding process of the next video data segment.
[0010] As one embodiment of the present invention, a multi-dimensional state vector is periodically collected through the multi-sensor interface built into the client, specifically including: By using the network transport layer interface with a sampling period of 100ns, the received timestamp and round-trip time of the transmission control protocol acknowledgment message are obtained and used to calculate the instantaneous network throughput. The original video frame data to be transcoded is obtained through the buffer at the front end of the video decoder and used as input for video content analysis. The system obtains the real-time core temperature and operating frequency of the central processing unit, as well as the remaining battery percentage reported by the battery management unit, through the operating system kernel interface.
[0011] As one embodiment of the present invention, based on network state data, a state-space model is used to predict the available network bandwidth within a future transmission time window, specifically including: A linear Kalman filter is used as the state-space model; The state vector of the Kalman filter is defined as a two-dimensional vector, whose components are the bandwidth estimate and bandwidth change rate at the current moment, respectively. The state transition matrix of the Kalman filter is set to a second-order square matrix with 1s on the main diagonal, the sampling time interval in the upper right corner, and 0s for the remaining elements, in order to describe the physical process of the bandwidth changing linearly with time. The observation matrix of the Kalman filter is set to a matrix with one row and two columns, where the first element is 1 and the second element is 0, representing the direct observation bandwidth value; The instantaneous network throughput is input into the Kalman filter as an observation. The filter iterates through two steps: prediction and update. It outputs a smoothed estimate of the current bandwidth and a predicted bandwidth value for the next 200ms to 500ms time window. The predicted bandwidth value is the predicted bandwidth value.
[0012] As one embodiment of the present invention, real-time calculation is performed on video content data to extract content feature parameters characterizing the dynamics and texture complexity of the video frame, specifically including: For the first keyframe in each image group obtained from the front-end buffer of the video decoder, the Sobel operator is applied for edge detection, and the spatial standard deviation of the pixel gradient magnitude of the entire frame is calculated as a spatial information index characterizing the texture complexity of the image. The mean of the absolute values of the pixel-by-pixel differences between the key frame and the two consecutive predicted frames is calculated, and the standard deviation of this mean in the time dimension is used as the temporal information index characterizing the intensity of motion in the image. The spatial information index and the temporal information index are combined into a two-dimensional content feature vector, which is used as the content feature parameter.
[0013] As one embodiment of the present invention, the client hardware operating status is mapped to a preset resource constraint level based on device physical status data, specifically including: Define three resource constraint levels: normal operation level, performance control level, and power priority level; When the CPU core temperature is below 85°C and the remaining battery power is above 15%, the device status is mapped to the normal operating level. When the CPU core temperature is higher than or equal to 85°C, the device state is mapped to a performance control level regardless of the battery level. When the remaining battery power is less than or equal to 15%, the device status will be mapped to the power priority level regardless of the core temperature.
[0014] As one embodiment of the present invention, in a pre-constructed multi-dimensional quality assessment model that associates transcoding resolution, transcoding bitrate, video content features, and encoded video quality, an optimal combination of target resolution and target bitrate is solved by combining predicted bandwidth values, content feature parameters, and resource constraint levels. Specifically, this includes: The multidimensional quality assessment model is a four-dimensional lookup table stored locally on the client. The four dimensions of the lookup table are resolution, bit rate, spatial information index, and temporal information index. The values stored in the table are the video multi-method evaluation fusion scores calculated through large-scale offline testing. When the device is in normal operating condition, in the four-dimensional lookup table, using the two-dimensional slices determined by the content feature parameters as the search space, find all "resolution-bitrate" grid points that satisfy that their corresponding bitrate is lower than 90% of the predicted bandwidth value, and select the grid point that makes the video multi-method evaluation fusion score the highest, whose corresponding resolution and bitrate are the optimal combination. When the device is in the performance restraint level, the search space is limited to a subset with a resolution of no more than 1920×1080 pixels, and computationally complex coding options are disabled. Then, the same search process as in the normal operation level is performed. When the device is in the power priority level, the search space is further limited to a subset with a resolution of no more than 1280×720 pixels and a bit rate of no more than 1Mbit / s, and then the same search process as the normal operation level is performed.
[0015] In summary, this application includes at least one of the following beneficial technical effects: 1. By employing a lightweight linear Kalman filter to perform millisecond-level bandwidth prediction on the client side, this invention reduces the decision response time from hundreds of milliseconds required by existing technologies that rely on historical data to less than 50ms. This effectively addresses sudden changes in network signals in scenarios such as subway tunnel entrances and exits, and significantly reduces video stuttering.
[0016] 2. This invention abandons the existing paradigm of decoupling resolution and bitrate control. By constructing a multi-dimensional bitrate-resolution-quality model that includes content complexity, it ensures that the selected transcoding parameter combination is the optimal match for the current video content under any network conditions. This avoids block artifacts caused by high resolution and low bitrate or resource waste caused by low resolution and high bitrate, and improves the actual viewing experience quality for users.
[0017] 3. This invention is the first to incorporate the physical states of the device, such as CPU temperature and battery level, as hard constraints for decision-making, establishing a resource constraint model. When the device is under high load or low battery, it automatically reduces the transcoding intensity, thereby avoiding performance degradation due to overheating or service interruption due to rapid battery depletion, and extending the effective usage time of the mobile device. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the overall technical solution of the adaptive resolution and bitrate video transcoding control method of the present invention. Detailed Implementation
[0019] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on specific implementation methods of the present invention.
[0020] Reference Figure 1 This invention provides a video transcoding control method that adapts to resolution and bitrate. Its core lies in constructing a multi-dimensional fusion decision system deployed locally on the client. By synchronously sensing three types of heterogeneous information—network status, video content characteristics, and device physical resources—it completes the joint solution of the optimal transcoding parameters within the future transmission window within a millisecond timescale. This method overcomes the shortcomings of traditional adaptive bitrate technology, such as relying on long-period historical data for bandwidth prediction, decoupling resolution and bitrate control, and ignoring device operating status feedback. It achieves a dynamic balance between smoothness, image quality, and energy consumption.
[0021] An adaptive resolution and bitrate video transcoding control method includes the following steps: S1, through the multi-sensor interface built into the client, periodically collects data to form a multi-dimensional state vector, which includes network status data, video content data and device physical status data. S2, based on network state data, uses a state-space model to predict the available network bandwidth within a future transmission time window and generates a predicted bandwidth value. S3 performs real-time calculations on video content data to extract content feature parameters that characterize the dynamics and texture complexity of the video images; S4, based on the device physical status data, maps the client hardware operating status to a preset resource constraint level; S5, in a pre-built multi-dimensional quality assessment model that associates transcoding resolution, transcoding bitrate, video content features and encoded video quality, combines predicted bandwidth values, content feature parameters and resource constraint levels to solve for an optimal combination of target resolution and target bitrate. S6 sends the optimal combination as a transcoding instruction to the video transcoding engine to control the transcoding process of the next video data segment.
[0022] In step S1, a multi-dimensional state vector is periodically collected through the multi-sensor interface built into the client. Its construction is the input basis of the entire decision system, and its completeness and timeliness directly determine the accuracy of subsequent prediction and optimization.
[0023] Specifically, the acquisition process executes three sub-operations simultaneously with a fixed sampling period of 100ns.
[0024] The first sub-operation is to obtain the reception timestamp and round-trip time of the Transmission Control Protocol (TCP) acknowledgment message through the network transport layer interface, and use the socket buffer monitoring mechanism provided by the operating system kernel to capture the arrival time of the acknowledgment message corresponding to each downlink data packet in real time, and calculate the single round-trip delay by combining it with the original sending timestamp of the data packet.
[0025] Based on this, the total number of payload bytes corresponding to the five most recent acknowledgment messages is counted using a sliding window mechanism, and then divided by the time span covered by these messages to calculate the instantaneous network throughput. This throughput value, as a core indicator of network status, has high noise characteristics but retains the original details of network fluctuations, making it suitable for subsequent filtering processing.
[0026] The second sub-operation is to obtain the original video frame data to be transcoded through the buffer of the front end of the video decoder. The video decoder refers to the module on the client side used to receive the original streaming media data and perform preliminary parsing. Its front end buffer stores the original image group sequence that has not yet entered the transcoding pipeline. In each sampling period, it locks the first keyframe and the two subsequent predicted frames in the current image group and copies its complete pixel matrix to a dedicated memory area for subsequent content complexity analysis.
[0027] The third sub-operation involves obtaining the real-time core temperature, operating frequency, and remaining battery percentage reported by the battery management unit from the operating system kernel interface. The core temperature is read by the chip's built-in digital temperature sensor, the operating frequency is dynamically reported by the performance control subsystem, and the remaining battery percentage is calculated by the power management integrated circuit using the coulomb counting method. At the end of each sampling period, these three types of data are encapsulated into a structured multidimensional state vector. This multidimensional state vector contains scalar values and image data blocks, serving as a unified input source for subsequent processing modules.
[0028] In step S2, based on network state data, a state-space model is used to predict the available network bandwidth within a future transmission time window, generating a predicted bandwidth value. This step aims to overcome the response lag problem caused by the reliance on long historical windows in traditional solutions.
[0029] This invention uses a linear Kalman filter as the core implementation of the state-space model, and its lightweight characteristics ensure that prediction latency of less than 50ms can be achieved even on low-end mobile devices.
[0030] The state vector of a Kalman filter is defined as a two-dimensional column vector. Its first component represents the bandwidth estimate at the current moment, in bits per second (bit / s), and the second component represents the bandwidth change rate, in bits per second². The state transition matrix is set as a second-order square matrix, with all elements on the main diagonal being 1, the upper right element being the sampling time interval of 0.1s, and the remaining elements being 0. This accurately describes the physical assumption that the bandwidth evolves linearly over time. The observation matrix is set as a matrix with one row and two columns, with the first element being 1 and the second element being 0. This indicates that only the bandwidth value itself can be directly observed, and its rate of change cannot be directly measured.
[0031] At the beginning of each sampling period, the Kalman filter first performs a prediction step: using the state estimate and state transition matrix of the previous period, it calculates the prior state estimate; at the same time, it updates the prior error covariance according to the process noise covariance matrix.
[0032] Subsequently, the instantaneous network throughput calculated in step S1 is used as the input observation value to perform the update step: Calculate the Kalman gain, combine the observations with the prior estimate to generate the posterior state estimate, and update the posterior error covariance.
[0033] After this iterative process, the Kalman filter outputs a smoothed estimate of the current bandwidth. Based on this estimate and the bandwidth change rate, the predicted bandwidth value for the next 200ms to 500ms time window is extrapolated. This predicted bandwidth value is the predicted bandwidth value, and its mathematical expression is as follows:
[0034] in, This is the bandwidth estimate for the current moment. Its rate of change, The predicted time offset is set to a value ranging from 0.2s to 0.5s, which reflects the basic principle of linear extrapolation and ensures the physical rationality of the prediction results within a short time scale.
[0035] In step S3, the video content data is calculated in real time to extract content feature parameters that characterize the dynamics and texture complexity of the video image, thus solving the problem of unbalanced bitrate allocation caused by ignoring the differences in video content in traditional solutions.
[0036] The specific implementation is divided into two parallel sub-processes.
[0037] The first subprocess quantizes the spatial complexity by applying the Sobel operator to the keyframes obtained from the front-end buffer of the video decoder for edge detection. The Sobel operator consists of two 3×3 convolution kernels, which are used to detect gradients in the horizontal and vertical directions, respectively.
[0038] For each pixel location, calculate its horizontal gradient. Vertical gradient Thus, the gradient magnitude is obtained. Then, the standard deviation of the gradient magnitude of all pixels in the whole frame is calculated. This standard deviation is the spatial information index, which is used to measure the richness of texture details in the image. High texture scenes, such as urban street scenes or dense vegetation, have a significantly higher spatial information index than low texture scenes, such as solid color backgrounds or skies.
[0039] The second subprocess quantifies the temporal complexity by calculating the absolute value of the pixel-by-pixel difference between the keyframe and the two subsequent consecutive prediction frames.
[0040] For each frame pair, iterate through all pixel positions, calculate the absolute value of the difference between the corresponding pixel brightness values, and obtain the average value for the entire frame. Then, treat the average value of these two inter-frame differences as a time series of length two, and calculate its standard deviation. This standard deviation value is the temporal information index, which is used to measure the intensity of motion and time-varying stability of the scene. Fast-moving scenes such as sports events or vehicle chases have a higher temporal information index, while static scenes such as conference speeches or landscape time-lapse photography have a lower value.
[0041] Finally, the spatial information index and the temporal information index are merged into a two-dimensional content feature vector, which is used as a content feature parameter and input into the subsequent joint decision-making module.
[0042] In step S4, the client hardware operating status is mapped to a preset resource constraint level based on the device physical status data. By introducing a device self-protection mechanism, overheating or sudden power drop caused by continuous high load transcoding is prevented.
[0043] Three mutually exclusive resource constraint levels are preset: normal operation level, performance control level, and power priority level.
[0044] The mapping rules are based on two hard thresholds: whether the CPU core temperature reaches or exceeds 85°C, and whether the remaining battery power drops to or falls below 15%.
[0045] When the core temperature is below 85°C and the battery level is above 15%, the device is in normal operating condition, at which point all computing resources can be used for high-quality transcoding.
[0046] When the core temperature reaches or exceeds 85°C, regardless of the battery level, the system immediately switches to a performance throttling level, forcibly limiting the resolution limit and disabling high-complexity coding tools. These include, but are not limited to, disabling bidirectional prediction frames (B-frames), limiting the number of reference frames to a single frame, disabling adaptive context binary arithmetic coding (CABAC), reducing the search range or complexity of the motion vector prediction (IMVP) mode, and high-complexity modes in rate-distortion optimization (RDO), in order to reduce the encoder's computational load and power consumption.
[0047] When the battery level drops to or below 15%, regardless of the core temperature, the system switches to power priority, further compressing the resolution and bitrate limits to maximize battery life.
[0048] As can be seen, the mapping process is implemented by a finite state machine, and its state transitions are driven only by the two threshold conditions mentioned above, ensuring the determinism and real-time performance of the decision logic.
[0049] In step S5, an optimal combination of target resolution and target bitrate is solved by combining predicted bandwidth value, content feature parameters and resource constraint level in a pre-built multi-dimensional quality assessment model that associates transcoding resolution, transcoding bitrate, video content features and encoded video quality.
[0050] The multidimensional quality assessment model is stored in the client's local flash memory in the form of a four-dimensional lookup table. Its four dimensions are resolution, bit rate, spatial information index, and temporal information index. The value stored in each grid point of the four-dimensional lookup table is the video multi-method evaluation fusion score, which is obtained by weighted fusion of offline large-scale subjective testing and objective indicators, and comprehensively reflects the human eye's perception and evaluation of video quality under specific transcoding parameters.
[0051] The pre-construction process of the multidimensional quality assessment model (four-dimensional lookup table) includes: Representative video sequences were selected, covering video content with different resolutions, bit rates, spatial information indices, and temporal information indices; The above video sequences were encoded under different combinations of resolution and bitrate to simulate actual transcoding scenarios; Perform large-scale subjective quality assessment (such as MOS score) and / or objective quality index (such as PSNR, SSIIM) on the encoded video segments, and fuse them into a multi-method evaluation fusion score for the video segment; The resolution, bitrate, spatial information index, temporal information index, and corresponding video multi-method evaluation fusion score are stored as data points in a four-dimensional lookup table on the client side to complete the model construction.
[0052] The video multi-method evaluation fusion score is derived by comprehensively and weightedly fusing objective quality metrics, including but not limited to Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Multi-Scale Structural Similarity (MIS-SSIM), with offline large-scale user subjective quality assessment (MOS) data. Specifically, the calculation method includes an offline stage: preparing a dataset of video clips covering different resolutions, bitrates, and content complexities; inviting multiple users to subjectively score each video clip, collecting subjective quality assessment (MOS) data; simultaneously calculating the objective quality metrics for each video clip; and using machine learning regression models (such as support vector regression, neural network models, etc.) to train the mapping relationship between objective metrics and subjective scores, forming a fusion score model. During runtime, this fusion score model outputs a predicted video quality score based on real-time acquired content features and transcoding parameters.
[0053] During runtime, the joint decision-making module first locates the corresponding two-dimensional slice in the four-dimensional lookup table based on the current content feature parameters; then, based on the resource constraint level determined in step S4, it applies different search boundaries to the slice.
[0054] If it is the normal operation level, then search all grid points that meet the condition that the bit rate is lower than 90% of the predicted bandwidth value within the full resolution and bit rate range, and select the grid point with the highest video multi-method evaluation fusion score as the optimal solution. The 90% safety margin is set to reserve the margin for network fluctuations and avoid the actual transmission exceeding the limit due to prediction deviation.
[0055] If the performance restraint level is set, the search space is limited to a subset with a resolution no higher than 1920×1080 pixels, and configurations that enable computationally complex coding options such as bidirectional prediction or multiple reference frames are excluded.
[0056] If the priority is power consumption, the search space is further limited to a subset with a resolution of no more than 1280×720 pixels and a bit rate of no more than 1 Mbit / s.
[0057] In all cases, the search process employs a linear scanning strategy, traversing all legal grid points, comparing their quality scores, and ultimately determining the unique optimal combination.
[0058] In step S6, the optimal combination is sent as a transcoding instruction to the video transcoding engine to control the transcoding process of the next video data segment. The transcoding instruction includes the target resolution, target bitrate, and encoding toolset limitations implied by the resource constraint level.
[0059] Upon receiving the instruction, the video transcoding engine immediately reconfigures its encoding parameters, including adjusting the output size of the image scaling module, setting the target bit rate of the bitrate control module, and enabling or disabling specific encoding tools. The next video data segment is then encoded according to the new parameters and sent to the server or directly pushed to the playback end via the network transport layer. The entire control loop cycle is strictly controlled within 100ns, ensuring real-time response to changes in the network and content.
[0060] In summary, this embodiment constructs a closed-loop adaptive transcoding control system through the close collaboration of six steps from S1 to S6. The entire perception, prediction, analysis and decision-making process is completed locally on the client side without relying on server-side models or cloud training, which meets the requirements of edge computing architecture. Its core innovation lies in integrating network prediction, content parsing and device constraints into a unified optimization framework, realizing the joint, real-time and adaptive adjustment of transcoding parameters.
[0061] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.
[0062] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A video transcoding control method that adapts to resolution and bitrate, characterized in that, include: Through the multi-sensor interface built into the client, a multi-dimensional state vector is periodically collected, which includes network status data, video content data, and device physical status data. Based on network state data, a linear Kalman filter is used as the state space model to predict the available network bandwidth within the future transmission time window and generate a predicted bandwidth value. Real-time calculations are performed on video content data to extract spatial information index and temporal information index, which characterize the dynamics and texture complexity of video images, as content feature parameters. Based on the device's physical status data, the client's central processing unit core temperature and remaining battery power are mapped to a preset resource constraint level; In a pre-constructed multidimensional quality assessment model that uses a four-dimensional lookup table to associate transcoding resolution, transcoding bitrate, video content features, and encoded video quality, the optimal combination of target resolution and target bitrate is solved by combining predicted bandwidth values, content feature parameters, and resource constraint levels. The optimal combination is sent as a transcoding instruction to the video transcoding engine to control the transcoding process of the next video data segment.
2. The video transcoding control method with adaptive resolution and bitrate according to claim 1, characterized in that, Through the client's built-in multi-sensor interface, a multi-dimensional state vector is periodically collected, including: By using the network transport layer interface with a sampling period of 100ns, the receiving timestamp and round-trip time of the transmission control protocol acknowledgment message are obtained and used to calculate the instantaneous network throughput. The original video frame data to be transcoded is obtained through the buffer at the front end of the video decoder and used as input for video content analysis. The system obtains the real-time core temperature and operating frequency of the central processing unit, as well as the remaining battery percentage reported by the battery management unit, through the operating system kernel interface.
3. The video transcoding control method with adaptive resolution and bitrate according to claim 1, characterized in that, Based on network state data, a linear Kalman filter is used as the state-space model to predict the available network bandwidth within future transmission time windows, including: The state vector of the Kalman filter is defined as a two-dimensional vector, whose components are the bandwidth estimate and bandwidth change rate at the current moment, respectively. The state transition matrix of the Kalman filter is set to be a second-order square matrix with 1s on the main diagonal, the sampling time interval in the upper right corner, and 0s for the rest. The observation matrix of the Kalman filter is set to be a matrix with one row and two columns, where the first element is 1 and the second element is 0; The instantaneous network throughput is input into the Kalman filter as an observation. The filter iterates through two steps: prediction and update. It outputs a smoothed estimate of the current bandwidth and a predicted bandwidth value for the next 200ms to 500ms time window. The predicted bandwidth value is the predicted bandwidth value.
4. The video transcoding control method with adaptive resolution and bitrate according to claim 1, characterized in that, Real-time calculations are performed on video content data to extract spatial and temporal information indices, which characterize the dynamics and texture complexity of video images, as content feature parameters, including: For the first keyframe in the image group obtained from the front-end buffer of the video decoder, the Sobel operator is applied for edge detection, and the spatial standard deviation of the pixel gradient magnitude of the entire frame is calculated as a spatial information index characterizing the texture complexity of the image. The mean of the absolute values of the pixel-by-pixel differences between the keyframe and the two subsequent consecutive predicted frames is calculated, and the standard deviation of the mean in the time dimension is used as the temporal information index characterizing the intensity of motion in the image.
5. The video transcoding control method with adaptive resolution and bitrate according to claim 1, characterized in that, Based on the device's physical status data, the client's CPU core temperature and remaining battery power are mapped to preset resource constraint levels, including: Define three resource constraint levels: normal operation level, performance control level, and power priority level; When the CPU core temperature is below 85°C and the remaining battery power is above 15%, the device status is mapped to the normal operating level. When the CPU core temperature is higher than or equal to 85°C, the device state is mapped to a performance control level regardless of the battery level. When the remaining battery power is less than or equal to 15%, the device status will be mapped to the power priority level regardless of the core temperature.
6. The video transcoding control method with adaptive resolution and bitrate according to claim 1, characterized in that, In a multidimensional quality assessment model in the form of a four-dimensional lookup table, by combining predicted bandwidth values, content feature parameters, and resource constraint levels, an optimal combination of target resolution and target bitrate is found, including: The multidimensional quality assessment model is a four-dimensional lookup table stored locally on the client. The four dimensions of the lookup table are resolution, bit rate, spatial information index, and temporal information index. The values stored in the lookup table are the video multi-method evaluation fusion scores calculated through large-scale offline testing. When the device is in normal operating condition, in the four-dimensional lookup table, using the two-dimensional slices determined by the content feature parameters as the search space, we search for all "resolution-bitrate" grid points that satisfy that their corresponding bitrate is lower than 90% of the predicted bandwidth value, and select the grid point that makes the video multi-method evaluation fusion score the highest, whose corresponding resolution and bitrate are the optimal combination. When the device is in the performance restraint level, the search space is limited to a subset with a resolution of no more than 1920×1080 pixels, and computationally complex coding options are disabled. Then, the same search process as in the normal operation level is performed. When the device is in the power priority level, the search space is further limited to a subset with a resolution of no more than 1280×720 pixels and a bit rate of no more than 1Mbit / s, and then the same search process as the normal operation level is performed.
7. The video transcoding control method with adaptive resolution and bitrate according to claim 3, characterized in that, The linear Kalman filter performs a prediction step and an update step in each sampling period. The prediction step uses the state estimate and state transition matrix of the previous period to calculate the prior state estimate. The update step combines the instantaneous network throughput observation and Kalman gain to generate the posterior state estimate, and extrapolates the bandwidth prediction value for the next 200ms to 500ms time window based on the posterior state estimate.
8. The video transcoding control method for adaptive resolution and bitrate according to claim 1 or 6, characterized in that, The optimal combination is sent as a transcoding instruction to the video transcoding engine, including: The target resolution, target bitrate, and the encoding toolset limitations implied by the resource constraint level are encapsulated into transcoding instructions; After receiving the transcoding instruction, the video transcoding engine reconfigures the output size of the image scaling module, the target bit rate of the bitrate control module, and the enabling status of specific encoding tools. The next video data segment is encoded and transmitted according to the reconfigured parameters.