Human body target quality grade determination method and system, storage medium and program product

Abstract

The application provides a human target quality grade determination method and system, a storage medium and a program product, and relates to the field of artificial intelligence. The method comprises the following steps: selecting several representative frame images of a human target from a video frame sequence; extracting a plurality of global score factors and local score factors of preset human key parts from each representative frame image to obtain a score factor set of each representative frame image; combining auxiliary information corresponding to each representative frame image to calculate correlation coefficients between the score factors and construct a correlation coefficient matrix to perform weight distribution on the score factors to obtain a weight distribution result; and performing weighted calculation on the score factor set based on the weight distribution result to obtain a quality score result of the human target on each representative frame image and an overall quality score result of the human target in the video. The method can dynamically adjust the weights of the score factors and reduce the quality grade determination deviation caused by weight imbalance.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method, system, storage medium, and program product for determining the quality level of human targets. Background Technology

[0002] With the rapid development of applications such as video surveillance, intelligent security, and public safety, higher demands are being placed on the automatic assessment of the image quality of human targets in videos. The quality score of human targets not only provides fundamental data for subsequent visual intelligent processing such as face recognition and behavior analysis, but also directly affects the accuracy of the overall system analysis results. Therefore, how to effectively assess the visual information quality of human targets in videos has become an important issue of concern in related fields.

[0003] In related technologies, for human targets appearing in videos or images, quality analysis is generally performed by extracting information from multiple quality-related scoring factors such as the sharpness, occlusion, and brightness of the human target, and then obtaining the quality level through weighted scoring. Regarding the weighting of scoring factors, there are implementation methods that use fixed weights for each scoring factor and perform a weighted sum, as well as methods that dynamically adjust the weights of scoring factors based on some prior data, scene attributes, or changes in scoring factors.

[0004] However, the weighting methods used for scoring factors in related technologies, whether employing fixed weights or dynamically adjusting them based on prior data and scenario attributes, fail to fully reflect the interrelationships and influence of different scoring factors on the actual quality performance of human targets. Therefore, in complex or variable application scenarios, the importance of individual scoring factors may be overestimated or underestimated, leading to discrepancies between the quality scores of some human targets and their actual quality levels. Summary of the Invention

[0005] This application provides a method, system, storage medium, and program product for determining the human body mass level, which addresses the problem that the human body mass level calculated by the weight setting method of scoring factors in related technologies may deviate significantly from the actual human body mass level of the human target.

[0006] In a first aspect, this application provides a method for determining the target quality level of a human body, applied to a human body quality scoring system, the method comprising: According to the preset representative sampling rules, several representative frame images of human targets are selected from the video frame sequence. The representative frame images include the appearance frame, disappearance frame, the frame with the lowest clarity, the frame with the highest occlusion, and the motion state change frame of the human target. Multiple global scoring factors and preset local scoring factors for key human body parts are extracted from each representative frame image to obtain a set of scoring factors for each representative frame image. The set of scoring factors includes multiple scoring factors, including pose angle, completeness, sharpness, and brightness. Based on the set of scoring factors and the auxiliary information corresponding to each representative frame image, the correlation coefficient between each scoring factor is calculated using statistical analysis methods to obtain a set of correlation coefficients. The correlation coefficients are used to characterize the correlation or information dependence between the scoring factors. Based on the set of correlation coefficients, a correlation coefficient matrix is ​​constructed to assign weights to each scoring factor, and the weight assignment results are obtained. Based on the weight allocation result, the set of scoring factors is weighted and calculated to obtain the quality score result of the human target on each representative frame image; The quality scores of the human target in the video are fused based on the temporal fusion method to obtain the overall quality score of the human target in the video.

[0007] Through the above embodiments, the system analyzes the scoring factors of multiple representative frames containing human targets in a video, determines the correlation coefficients between different types of scoring factors, assigns weights to different types of scoring factors based on the correlation coefficients, and then weights them to obtain the overall quality score of the human target in the video. The weights determined by this method can dynamically reflect the true contribution of each scoring factor to the quality of the human target in the actual scene. The human target score obtained based on this dynamic weighting can address the quality level deviation caused by the imbalance of scoring factor weights in related technologies, achieving a high degree of consistency between the human target quality level assessment and actual performance, and improving the accuracy and reliability of the scoring results.

[0008] In some embodiments, the step of calculating the correlation coefficient set between each scoring factor based on the set of scoring factors and the auxiliary information corresponding to each representative frame image using statistical analysis methods specifically includes: The scoring factor values ​​corresponding to each category of scoring factors in each set of scoring factors are extracted sequentially according to the order of representative frame images to obtain the scoring factor sequence; Obtain the rating factor sequences of any two types of rating factors in all representative frame images to obtain rating factor sequence pairs. The rating factor sequence pairs are combinations of the value sequences of the two types of rating factors in different frame images. By removing the parameter values ​​of auxiliary information that exceed the preset normal range threshold, the corrected scoring factor sequence pairs are obtained. The correlation coefficients of the corrected scoring factor sequence pairs were calculated using statistical analysis methods. By summing all the correlation coefficients, a set of correlation coefficients is obtained, which includes the correlation coefficients between any two types of rating factors.

[0009] Through the above embodiments, the system can eliminate noise interference by serialization and outlier removal, making the correlation coefficient calculation more realistically reflect the actual dependence between different scoring factors. The resulting set of correlation coefficients can alleviate the weight imbalance problem caused by the inaccurate reflection of the mutual influence between scoring factors, thus improving the accuracy and consistency of human target quality level assessment.

[0010] In some embodiments, the step of constructing a correlation coefficient matrix based on the correlation coefficient set, assigning weights to each scoring factor, and obtaining the weight assignment result specifically includes: Construct a correlation coefficient matrix based on the correlation coefficients among the rating factors in the aforementioned correlation coefficient set; Calculate the sum of the absolute values ​​of the correlation coefficients of each rating factor with the other rating factors in the correlation coefficient matrix to obtain the total correlation measure of each rating factor; The total relevance measure of each scoring factor is reverse-quantified and normalized using a preset weighting rule to obtain the weighting result.

[0011] Through the above embodiments, the system objectively quantifies the actual correlation between scoring factors, highlighting scoring factors with strong independence and significant contributions to overall quality judgment, while suppressing the weights of redundant or strongly correlated factors. This dynamically allocated weight can address the accumulation of errors under traditional static or subjective weight settings, improving the scientific rigor and adaptability of human target quality level assessment results.

[0012] In some embodiments, after the step of performing reverse quantization and normalization on the total relevance measure of each scoring factor using a preset weight allocation rule to obtain the weight allocation result, the method further includes: Collect auxiliary task labels related to human target quality, including human behavior category labels, human attribute labels, and human tracking stability labels; A target function is constructed based on the auxiliary task labels and the weight allocation results. The target function includes performance indicators of the scoring results and performance indicators of the auxiliary labels. Obtain the weights of the scoring factors corresponding to the objective function when the objective function value reaches the preset optimal range, and obtain the optimized weight allocation result.

[0013] Through the above embodiments, based on the initial weight allocation, the system further collects auxiliary task labels related to human body target quality and combines these auxiliary labels with the performance indicators of the scoring results to construct an objective function. By optimizing the objective function, the system ensures that the weights of the scoring factors not only consider the quality score itself but also the performance of the auxiliary tasks, resulting in an optimized weight allocation. This method, by integrating multi-source label information, can dynamically adjust the weight allocation to adapt to complex real-world scenarios, addressing the inadequacies and biases caused by single-factor weight configurations, thereby improving the overall performance and generalization ability of human body quality assessment.

[0014] In some embodiments, the step of fusing the quality score results of the human target on all representative frame images based on the temporal fusion method to obtain the overall quality score result of the human target in the video specifically includes: The quality score results of each representative frame image are obtained sequentially to obtain a quality score result sequence, wherein the representative frame images are arranged in chronological order; Calculate the difference between two adjacent quality score results in the quality score result sequence to obtain the score change sequence; Obtain the absolute value of each item in the score change sequence to obtain the score change magnitude sequence; Items in the score change range sequence that exceed a preset threshold are marked to obtain the marking results; Increase the weight of the quality score of the representative frame image corresponding to the labeling result in the overall quality score by a preset amplitude value, and obtain the preliminary quality score result corresponding to each representative frame image; The preliminary quality score results are weighted and fused based on a temporal fusion method to obtain the overall quality score result of the human target in the video.

[0015] Through the above embodiments, the system arranges the quality scores of each representative frame image in chronological order to form a score result sequence, calculates the score variation between adjacent frames, marks items with variation exceeding a preset threshold, and increases their weight in the overall quality score. Finally, the system fuses the weighted score results of each frame based on a temporal fusion method to obtain the overall quality score of the human target in the video. This method can highlight the impact of keyframes on the overall quality, weaken the interference of ordinary or abnormal frames, effectively cope with quality assessment deviations caused by score fluctuations or occasional anomalies, and thus improve the stability and accuracy of the overall quality score of the human target.

[0016] In some embodiments, the step of extracting multiple global scoring factors and preset local scoring factors of key human body parts from each representative frame image to obtain a set of scoring factors for each representative frame image further includes: Each representative frame image undergoes multi-stage spatial partitioning and downsampling to obtain feature data; Based on the aforementioned feature data, a global feature representation set covering the entire target area of ​​the human body and a local feature representation set locating key parts of the human body are extracted. Channel dimension weighted calculation and spatial adaptive pooling are performed on the global feature representation set and the local feature representation set respectively to obtain the global scoring factor set and the local scoring factor set; The global scoring factor set is combined with the local scoring factor set to obtain the scoring factor set for each representative frame image.

[0017] Through the above embodiments, the system performs multi-stage spatial partitioning and downsampling processing on each representative frame image, extracting global feature representations covering the entire human target region and local feature representations locating key parts of the human body. Then, channel weighting and spatial adaptive pooling are performed respectively, ultimately combining them into a set of scoring factors. This method, through joint modeling of global and local features, can meticulously capture the overall state and key details of the human target, thus providing more comprehensive and accurate basic data for correlation analysis and weight allocation of scoring factors. Therefore, the human target quality level evaluation can better reflect the actual quality level, reduce scoring errors caused by insufficient feature extraction, and achieve improved granularity and accuracy in the assessment.

[0018] In some embodiments, after the step of fusing the quality scores of the human target on all representative frame images based on the temporal fusion method to obtain the overall quality score of the human target in the video, the method further includes: The difference between the quality score of each representative frame image and the overall quality score of the human target in the video is calculated to obtain the difference sequence; Items in the difference sequence whose absolute value is greater than a preset threshold are marked with a deviation. The confidence score corresponding to the auxiliary information is used to determine the confidence weight of the target representative frame image corresponding to each deviation marker; The quality score of the representative frame image of the target is corrected by weighting the confidence score based on the confidence score, and the corrected score is obtained. Based on the corrected scoring results, the overall quality score of the human target in the video is re-fused to obtain the corrected overall quality score of the human target.

[0019] Through the above embodiments, after obtaining the overall quality score of the human target in the video, the system further calculates the difference between the score of each representative frame and the overall score. Items with an absolute difference greater than a threshold are marked as deviations, and the scores of these frames are weighted and corrected by combining the confidence score of auxiliary information. Finally, the system re-integrates the overall quality score based on the corrected score results to obtain the corrected overall quality score of the human target. This method, by dynamically identifying and correcting scoring deviations, can effectively overcome the influence of single-frame anomalies or scoring instability, improve the adaptive correction capability and final accuracy of the scoring results, and achieve a high degree of consistency between the human target quality level assessment results and actual performance.

[0020] Secondly, this application provides a human body mass indexing system, which includes one or more processors and a memory; The memory is coupled to the one or more processors. The memory is used to store computer program code, which includes computer instructions. The one or more processors call the computer instructions so that the human body mass scoring system can implement the human target mass level determination method provided in the above embodiments, which will not be described in detail here.

[0021] Thirdly, this application provides a computer-readable storage medium including instructions that, when executed on a human body mass indexing system, enable the human body mass indexing system to implement a human target quality level determination method provided in the above embodiments, which will not be elaborated here.

[0022] Fourthly, this application provides a computer program product that, when run on a human body quality scoring system, enables the human body quality scoring system to implement a human target quality level determination method provided in the above embodiments, which will not be elaborated here.

[0023] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. Representative human target frames are selected from the video using preset sampling rules. Global and local multidimensional scoring factors are extracted for each frame. Furthermore, statistical analysis methods are used to systematically quantify the correlation between these scoring factors, constructing a correlation coefficient matrix. Subsequently, a reverse quantization and normalization mechanism is employed to adaptively allocate the weights of each scoring factor, enabling the weights to dynamically reflect the true contribution of different factors to human target quality in real-world scenarios. This correlation-based weight allocation method overcomes the limitations of traditional static or subjective weight allocation, achieving a more scientific, objective, and scenario-adaptive assessment of human quality levels.

[0024] 2. Building upon the initial weight allocation, the scheme further introduces multi-source auxiliary task labels such as behavioral categories, attribute features, and tracking stability. These labels are integrated into the objective function along with the performance indicators of the quality scoring results. The weights of the scoring factors are dynamically adjusted through the optimization of the objective function. This mechanism effectively integrates and utilizes external label information to drive weight configuration, significantly enhancing the scoring system's adaptability to complex real-world scenarios and diverse application needs, and greatly improving the overall performance and generalization ability of human body quality assessment.

[0025] 3. In the scoring fusion stage, a strategy of time-series change awareness and dynamic keyframe weight enhancement is adopted. Difference analysis and dynamic weighting are performed on the fluctuations in the scores of representative frames to highlight the impact of keyframes on the overall evaluation. At the same time, for abnormal fluctuations or deviations in the scores, the scores of abnormal frames are weighted and corrected and then fused again in combination with the confidence level of auxiliary information to form a final quality score with adaptive correction capabilities. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a method for determining the target quality level of a human body in an embodiment of this application; Figure 2 This is another flowchart illustrating a method for determining the target quality level of a human body in an embodiment of this application; Figure 3 This is a schematic diagram of the physical device structure of a human body quality scoring system in the embodiments of this application. Detailed Implementation

[0027] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0028] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0029] For ease of understanding, the method provided in this implementation is described in process below. Please refer to [link / reference]. Figure 1This is a flowchart illustrating a method for determining the quality level of a human target in an embodiment of this application.

[0030] S101. Based on the preset representative sampling rules, select several representative frame images of the human target from the video frame sequence.

[0031] Among them, the preset representative sampling rule refers to the standard or logic pre-set by the system to select key frames that can reflect different states of human targets from the video frame sequence; the representative frame image refers to the key frame selected from the video frame sequence that can reflect the quality characteristics of human targets in different states; the appearance frame refers to the frame in which the human target first appears in the video; the disappearance frame refers to the frame in which the human target last appears in the video; the frame with the lowest clarity refers to the frame in which the human target area image has the highest degree of blur; the frame with the highest occlusion degree refers to the frame in which the human target is covered by other objects to the highest extent; and the frame with a change in motion state refers to the frame in which the human target's motion speed, direction, or posture changes significantly.

[0032] Specifically, the human body quality scoring system selects representative frames from a series of video frames. Since the human target's state is similar across most frames in the video, directly analyzing all frames would lead to redundant calculations. Representative frames, however, cover the key states of the human target (such as appearance, disappearance, blurring, occlusion, and motion changes), sufficiently reflecting its overall quality characteristics. Based on preset rules, the system sequentially detects the video frame sequence, identifying the frame where the human target first appears (appearance frame), the frame where it last appears (disappearance frame), the frame with the lowest clarity (e.g., determined by image sharpness algorithms), the frame with the most severe occlusion (e.g., determined by target detection box overlap rate), and the frame with abrupt changes in motion (e.g., speed or direction changes exceeding a threshold). These frames are then aggregated to form a representative frame image set.

[0033] Optionally, the system traverses the video frame sequence frame by frame, using a human detection model (such as YOLOv8) to identify whether each frame contains a human target. The frame where the human target is first detected is recorded as the appearance frame, and the frame where it is last detected is recorded as the disappearance frame. For frames containing human targets, the Laplacian operator is used to calculate image sharpness (the lower the sharpness value, the lower the clarity), and the frame with the lowest sharpness value is selected as the frame with the lowest clarity. A semantic segmentation model is used to identify human regions and other object regions, and the proportion of human regions covered is calculated (the higher the proportion, the greater the occlusion), and the frame with the largest proportion is selected as the frame with the greatest occlusion. The motion trajectory of the human target is analyzed, and when the rate of change of human speed in adjacent frames exceeds 30% or the angle of change of direction exceeds 60°, the frame is marked as a motion state change frame. The above frames are summarized, and duplicate frames are removed to obtain a representative set of frame images.

[0034] S102. Extract multiple global scoring factors and preset local scoring factors of key human body parts from each representative frame image to obtain a set of scoring factors for each representative frame image.

[0035] Among them, global scoring factors refer to factors that reflect the overall quality characteristics of the entire human body target; local scoring factors refer to factors that reflect the quality characteristics of key parts of the human body.

[0036] Specifically, the human body quality scoring system extracts specific factors reflecting quality from each representative frame image, including both global features of the entire human body and local features of key areas. Global scoring factors (such as overall pose angle, completeness, sharpness, and brightness) are used to assess the overall quality of the human target, while local scoring factors (such as head sharpness and torso completeness) are used to assess the quality of key areas—because the quality of key areas (such as the head) has a greater impact on tasks such as face recognition. The system analyzes the frame images through image processing or deep learning models, quantifies these features into specific values ​​(such as a pose angle of 30° and completeness of 80%), and integrates them into a set of scoring factors for each frame.

[0037] Optionally, (1) preprocess each representative frame image (e.g., denoising, normalizing size); (2) use a human pose estimation model (e.g., MediaPipe) to extract key points of the human skeleton and calculate the angle between the skeletons (e.g., the angle between the torso and the thigh) as the pose angle; (3) calculate the completeness by the percentage of the overlapping area between the target detection box and the standard human template (the higher the overlap rate, the higher the completeness); (4) use the tenengrad gradient function to calculate the gradient value of the entire human body region. The higher the gradient value, the higher the clarity; (5) calculate the average gray value of the pixels in the human body region as the brightness; (6) locate the preset key parts (head, hands), repeat steps (4)-(5) to calculate the local clarity and brightness, and integrate global and local factors to form a set of scoring factors.

[0038] Optionally, for each representative frame image, the human body quality scoring system first performs multi-stage spatial partitioning, dividing the image into multiple sub-regions according to different scales (e.g., first dividing into an 8×8 large grid, then dividing the large grid into 4×4 small grids), and downsampling each sub-region using average pooling or max pooling (e.g., reducing a 64×64 sub-region to 32×32), obtaining feature data containing multi-scale spatial information; then, it extracts a global feature representation set from the feature data (integrating all sub-region features through global pooling to reflect the overall human posture, contour integrity, etc.), and simultaneously calls a human keypoint detection model to locate key parts such as the head and torso, extracting the corresponding sub-region feature shapes. The system generates a set of local feature representations (reflecting the sharpness and occlusion of key parts); then, it performs channel-dimensional weighted calculations (e.g., assigning higher weights to highly relevant channels using the SENet mechanism) and spatial adaptive pooling (using small window pooling to preserve details in high-activation regions and large window pooling to reduce redundancy in low-activation regions) on the global and local feature representation sets respectively, resulting in a quantified set of global scoring factors (e.g., overall integrity, global brightness) and a set of local scoring factors (e.g., head sharpness, torso occlusion); finally, the two sets are combined (e.g., feature stitching or weighted fusion) to form a set of scoring factors for each representative frame image containing multi-dimensional quality indicators.

[0039] S103. Based on the set of scoring factors and the auxiliary information corresponding to each representative frame image, the correlation coefficients between the scoring factors are calculated using statistical analysis methods to obtain the set of correlation coefficients.

[0040] Among them, auxiliary information refers to additional data related to representative frame images that may affect the accuracy of scoring factors (such as camera focal length, lighting conditions, image resolution, etc.); statistical analysis methods refer to mathematical methods used to analyze the degree of correlation between variables (such as Pearson correlation coefficient, Spearman rank correlation coefficient, etc.); correlation coefficient refers to a numerical value that quantifies the degree of linear correlation between two variables (ranging from -1 to 1, with a larger absolute value indicating a stronger correlation).

[0041] Specifically, the human body quality scoring system calculates the correlation between different scoring factors because these factors are not independent (e.g., abnormal brightness may cause sharpness distortion), and their interrelationships affect the overall assessment of human body quality. The system first incorporates auxiliary information (e.g., removing abnormal brightness values ​​caused by excessive lighting) to eliminate noise interference; then, it uses statistical analysis methods (e.g., Pearson correlation coefficient) to calculate the correlation between the values ​​of any two types of scoring factors (e.g., sharpness and occlusion, brightness and pose angle) across all representative frames; finally, it summarizes the correlation coefficients of all pairs of factors to form a correlation coefficient set. This set clearly reflects which factors are closely related (e.g., occlusion is negatively correlated with integrity) and which factors are relatively independent (e.g., pose angle and brightness have no significant correlation), providing a basis for subsequent weight allocation based on correlation.

[0042] Optionally, the system extracts the numerical values ​​of each factor (e.g., sharpness: 0.6, 0.8...; occlusion: 0.3, 0.5...) from the set of scoring factors for each representative frame, and collects the corresponding auxiliary information (e.g., illumination: 50 lux, 200 lux...); sets a normal range for auxiliary information (e.g., illumination 30-300 lux), and removes the scoring factor values ​​of frames that exceed the range to avoid interference from abnormal data; for any two types of scoring factors (e.g., sharpness and occlusion), it extracts their values ​​in the valid frames. The numerical sequences (e.g., sharpness sequence: [0.6, 0.8, 0.5], occlusion sequence: [0.3, 0.2, 0.6]) are used to calculate the correlation coefficient of the sequence pair using the Pearson correlation coefficient formula (formula: r=[nΣxy-(Σx)(Σy)] / sqrt[(nΣx²-(Σx)²)(nΣy²-(Σy)²)], where n is the number of valid frames, and x and y are the values ​​of the two factors). The above correlation coefficient calculation is repeated for all possible scoring factor pairs, and the correlation coefficient set is obtained by summing them up.

[0043] S104. Construct a correlation coefficient matrix based on the correlation coefficient set, assign weights to each scoring factor, and obtain the weight assignment results.

[0044] The correlation coefficient matrix is ​​a matrix constructed with the rating factors as rows and columns and the correlation coefficients of the corresponding two types of rating factors as elements. It is used to visually present the degree of correlation between all rating factors. The weight allocation is the process of assigning a proportion of each rating factor in the total score according to the contribution of the rating factors to the overall quality. The weight allocation result is the set of weight values ​​corresponding to each rating factor after allocation (the sum of the weight values ​​is usually 1).

[0045] Specifically, the human body quality scoring system constructs a correlation coefficient matrix using a set of correlation coefficients: the rows and columns of the matrix correspond to all scoring factors (such as posture angle, completeness, sharpness, brightness, etc.), and the element in the i-th row and j-th column of the matrix is ​​the correlation coefficient between the i-th and j-th scoring factors. Subsequently, the system evaluates each scoring factor by analyzing this matrix and calculating the total correlation metric. Furthermore, a higher total correlation metric (the sum of the absolute values ​​of the correlation coefficients of a factor with all other factors) indicates a stronger correlation between that factor and other factors (potentially indicating information redundancy), and a smaller unique contribution to the overall quality assessment; conversely, a lower total correlation metric indicates stronger independence of the factor and a greater unique contribution. Based on this, the system employs reverse quantization (high total correlation results in lower weight, and low total correlation results in higher weight) and normalization (making the sum of all weights equal to 1) to ultimately obtain the weight of each scoring factor, forming the weight allocation result.

[0046] Optionally, assuming the rating factors are A (posture angle), B (completeness), C (sharpness), and D (brightness), a 4×4 correlation coefficient matrix M is constructed based on the correlation coefficient set, where M[i][j] is the correlation coefficient between factors i and j; the total correlation measure of each factor is calculated: total correlation of A = |M[A][B]| + |M[A][C]| + |M[A][D]|, and the total correlation of B, C, and D is calculated similarly; reverse quantization processing: set the initial weight value = 1 / total correlation measure value (the higher the total correlation, the lower the initial weight); normalization processing: divide all initial weights by the sum of weights to obtain the final weight (e.g., weight of A = initial weight of A / (initial weight of A + initial weight of B + C + initial weight of D)), forming the weight allocation result.

[0047] S105. Based on the weight allocation results, perform weighted calculation on the set of scoring factors to obtain the quality score results of human targets on each representative frame image.

[0048] Specifically, for each representative frame image, the human body quality scoring system extracts the quantified values ​​of all scoring factors (such as pose angle 85°, integrity 90%, sharpness 0.8, brightness 0.7, etc.) from its set of scoring factors, and matches them with the corresponding weights obtained in step S104 (such as pose angle weight 0.2, integrity 0.3, sharpness 0.3, brightness 0.2). Then, the system obtains the comprehensive score for that frame through weighted calculation (i.e., summing the quantified value of each factor multiplied by its corresponding weight): for example, if the factor values ​​of a frame are pose angle 0.8 (normalized), integrity 0.9, sharpness 0.7, and brightness 0.6, with weights of 0.2, 0.3, 0.3, and 0.2 respectively, then the quality score = 0.8 × 0.2 + 0.9 × 0.3 + 0.7 × 0.3 + 0.6 × 0.2 = 0.16 + 0.27 + 0.21 + 0.12 = 0.76. This process ensures that the score for each frame reflects the actual contribution of each factor.

[0049] S106. Based on the temporal fusion method, the quality score results of the human target on all representative frame images are fused to obtain the overall quality score result of the human target in the video.

[0050] Among them, the temporal fusion method refers to a method that considers the temporal order of representative frame images and performs weighted fusion of their quality scores (emphasizing that frames at different time points have different impacts on the overall quality).

[0051] Specifically, the human body quality scoring system first arranges the quality scores of all representative frames in chronological order (e.g., appearance frame → motion change frame → lowest sharpness frame → maximum occlusion frame → disappearance frame), forming a time series. Since frames at different times have varying impacts on overall quality (e.g., critical state frames are more important than ordinary frames), the system employs a temporal fusion method: for example, assigning higher weights to frames closer to the middle of the video (assuming middle-section frames are more stable), or increasing the weights of critical frames such as motion change frames and frames with maximum occlusion (because they reflect extreme quality states). Through weighted fusion (e.g., weight × the sum of the corresponding frame scores), the overall quality score is obtained. This result integrates the quality characteristics of the human target in different states and at different times, providing a more comprehensive reflection of the human target's quality level throughout the entire video.

[0052] Furthermore, after obtaining the overall quality score of the human target in the video, the human quality scoring system first calculates the difference between the quality score of each representative frame and the overall quality score, forming a difference sequence to reflect the degree of deviation between the single-frame score and the overall score. For example, if the scores of five frames are 80, 75, 90, 65, and 85, and the overall score is 79, the corresponding difference sequence is [1, -4, 11, -14, 6]. Next, the absolute value of each element in the difference sequence is taken and compared with a preset threshold (e.g., 10). Items with absolute values ​​greater than the threshold are marked as deviation markers. For example, in the above sequence, the absolute values ​​of 11 and 14 are greater than 10, corresponding to the marking of frames 3 and 4 as significantly deviation frames. Subsequently, auxiliary information and confidence scores of these target frames are extracted (e.g., frame 3). Because the confidence level for abnormal lighting is 0.4, and the confidence level for frame 4 is 0.8 due to information reliability, the confidence scores are normalized to obtain confidence weights (0.33 for frame 3 and 0.67 for frame 4). Then, the target frame scores are weighted and corrected according to the confidence weights. Frames with scores higher than the overall score are corrected downwards, and frames with scores lower than the overall score are corrected upwards. The correction amount is (overall score - original score) × confidence weight. For example, frame 3 is corrected from 90 to approximately 86.37, and frame 4 is corrected from 65 to approximately 74.38. Finally, the corrected target frame scores are time-series fused with the original scores of other uncorrected frames (e.g., using average weighting) to obtain the corrected overall quality score result, for example, approximately 80.15. This result eliminates the significant deviation of abnormal frames and is more in line with the actual quality level of human targets.

[0053] In the above embodiments, the system analyzes the scoring factors of multiple representative frames containing human targets in a video, determines the correlation coefficients between different types of scoring factors, assigns weights to different types of scoring factors based on the correlation coefficients, and then weights them to obtain the overall quality score of the human target in the video. The weights determined by this method can dynamically reflect the true contribution of each scoring factor to the quality of the human target in the actual scene. The human target score obtained based on this dynamic weighting can address the quality level deviation caused by the imbalance of scoring factor weights in related technologies, achieve a high degree of consistency between the human target quality level assessment and actual performance, and improve the accuracy and reliability of the scoring results.

[0054] The following provides a more detailed description of the process of the method provided in this implementation. Please refer to [link / reference]. Figure 2 This is another flowchart illustrating a method for determining the quality level of a human target in an embodiment of this application.

[0055] S201. Extract the rating factor values ​​corresponding to each type of rating factor in each set of rating factors in the order of representative frame images to obtain the rating factor sequence.

[0056] Specifically, the human body quality scoring system first determines the temporal order of all representative frame images (e.g., arranged by frame number from smallest to largest to ensure consistency with the video playback order). Then, for each set of scoring factors, the system extracts the scoring factor values ​​by category: for example, it extracts "sharpness" values ​​from the scoring factor set of "frames with appearance," extracts "sharpness" values ​​from the scoring factor set of "frames with changes in motion," and so on, until all "sharpness" category values ​​have been extracted from all representative frames and arranged chronologically to form a "sharpness" scoring factor sequence. This process needs to cover all scoring factor categories (such as pose angle, integrity, brightness, etc.), ultimately resulting in multiple scoring factor sequences consistent with the number of scoring factor categories.

[0057] S202. Obtain the rating factor sequences of any two types of rating factors in all representative frame images to obtain rating factor sequence pairs.

[0058] Specifically, the human body quality scoring system selects all possible pairwise combinations from all scoring factor sequences: assuming there are n types of scoring factors, the number of combinations is n×(n-1) / 2 (to avoid duplicate pairings). For example, if there are three types of factors: "pose angle (A), integrity (B), and sharpness (C)," three sequence pairs (A, B), (A, C), and (B, C) need to be generated. Each sequence pair contains two sequences of the same length (because they are extracted from the same representative frame image), and the values ​​at corresponding positions in the sequences come from the values ​​of the two types of factors in the same frame image (e.g., in sequence pair (C, B), the i-th element is the sharpness value and integrity value of the i-th frame).

[0059] S203. Remove the combination of values ​​of auxiliary information corresponding to parameter values ​​that exceed the preset normal range threshold to obtain the corrected scoring factor sequence pairs.

[0060] Specifically, the human body quality scoring system acquires auxiliary information parameter values ​​corresponding to each representative frame image (e.g., the illumination intensity of a frame is 1500 lux, exceeding the preset normal range of 100-1000 lux). Subsequently, the system checks the auxiliary information of the frame corresponding to each value sequence combination in the scoring factor sequence pair: if the auxiliary information parameter value of a frame exceeds a preset threshold, the corresponding numerical combination of that frame in the sequence pair (e.g., a sharpness value of 0.3 and a brightness value of 0.9) is determined to be an "outlier combination." The system removes all outlier combinations from the sequence pair while maintaining the temporal order of the remaining values ​​(e.g., if the original sequence pair length is 5, after removing 2 outliers, the sequence pair length becomes 3).

[0061] S204. Calculate the correlation coefficients of the corrected scoring factor sequence pairs according to statistical analysis methods, and summarize all correlation coefficients to obtain the correlation coefficient set.

[0062] Specifically, the human body quality scoring system calculates correlation coefficients for each corrected pair of scoring factors using appropriate statistical analysis methods. For example, if the values ​​in the pair exhibit a linear relationship (e.g., sharpness increases linearly with brightness), the Pearson correlation coefficient is used; if the relationship is monotonic but non-linear (e.g., occlusion decreases monotonically with integrity), the Spearman rank correlation coefficient is used. During the calculation, the system can perform significance tests based on the length of the pair (i.e., the number of effective frames) (e.g., using a t-test to determine if the correlation coefficient is statistically significant). Subsequently, the system summarizes the correlation coefficients of all pairs of pairs (including significant and insignificant correlation coefficients that passed the test) to form a set of correlation coefficients. For example, the correlation coefficient for "pose angle-integrity" is 0.2, the correlation coefficient for "sharpness-brightness" is 0.8, and the correlation coefficient for "occlusion-integrity" is -0.9, etc.

[0063] S205. Construct a correlation coefficient matrix based on the correlation coefficients between the rating factors in the correlation coefficient set.

[0064] Specifically, the body quality scoring system uses the categories of all scoring factors (such as posture angle, completeness, sharpness, brightness, etc.) as row and column indices respectively in the correlation coefficient matrix (the order of rows and columns is consistent). Then, the system extracts the correlation coefficient for each pair of scoring factors from the correlation coefficient set and fills the corresponding positions in the matrix: for example, if the correlation coefficient between "posture angle" and "completeness" is 0.3, then the element value in the "posture angle" row and "completeness" column of the matrix is ​​0.3, and the element value in the "completeness" row and "posture angle" column is also 0.3 (due to the symmetry of correlation). For the diagonal positions (where a factor is correlated with itself), a uniform 1 is filled (indicating perfect correlation).

[0065] S206. Calculate the sum of the absolute values ​​of the correlation coefficients of each rating factor with the other rating factors in the correlation coefficient matrix to obtain the total correlation measure of each rating factor.

[0066] Specifically, the body quality scoring system calculates the sum of the absolute values ​​of the correlation coefficients between each scoring factor (i.e., each row of the matrix) and all other scoring factors in the correlation coefficient matrix. For example, if the elements in the "Clarity" row of the matrix are [0 (with itself), 0.2 (with pose angle), 0.5 (with completeness), -0.3 (with brightness)], then the absolute values ​​are [0, 0.2, 0.5, 0.3]. Summing these values ​​yields a total of 1.0, meaning the total correlation measure for "Clarity" is 1.0. This value reflects the overall strength of the association between this factor and other factors: a higher total correlation measure indicates a higher degree of information overlap between this factor and other factors (potentially redundancy); conversely, a lower value indicates a stronger independence of this factor and a greater unique contribution to the overall quality assessment.

[0067] Optionally, the system iterates through each row of the correlation coefficient matrix (corresponding to one rating factor); for each row, the diagonal elements (corresponding to themselves with a correlation coefficient of 1) are removed, and the absolute values ​​of the remaining elements are taken; the sum of the absolute values ​​is calculated and used as the total correlation measure of the rating factor corresponding to that row; the above calculation process of the total correlation measure is repeated until the total correlation measure of all rating factors is obtained (e.g., attitude angle: 0.8, completeness: 1.2, sharpness: 1.0, brightness: 0.6).

[0068] S207. Using the preset weight allocation rules, the total relevance measure of each scoring factor is reverse-quantified and normalized to obtain the weight allocation result.

[0069] Specifically, the body mass indexing system transforms the total correlation metric into intermediate weights through reverse quantification: for example, if the total correlation metric of a factor is 1.0, the intermediate weight can be set to 1 / 1.0 = 1; if the total correlation metric is 2.0, the intermediate weight is set to 1 / 2.0 = 0.5 (reflecting the inverse relationship). The purpose of this step is to highlight factors with strong independence (low total correlation → high intermediate weight). Subsequently, the system performs normalization: all intermediate weights are added together to obtain a sum, and then each intermediate weight is divided by the sum to make the final weight sum equal to 1 (e.g., if the intermediate weights are [1, 0.5, 0.8, 1.2], the sum is 3.5, and the normalized weights are [1 / 3.5≈0.29, 0.5 / 3.5≈0.14, 0.8 / 3.5≈0.23, 1.2 / 3.5≈0.34]).

[0070] S208. Collect auxiliary task labels related to human target quality, and construct an objective function based on the auxiliary task labels and weight allocation results.

[0071] Among them, auxiliary task labels related to human target quality refer to the label information of other tasks related to human targets besides human quality scoring. They are used to constrain weight allocation from multiple dimensions, including human behavior category labels (labels that describe the types of human target actions, such as walking, running, standing, etc.), human attribute labels (labels that describe the inherent characteristics of human targets, such as gender, age, clothing color, etc.), and human tracking stability labels (labels that represent the continuity and accuracy of the trajectory of human targets when they are tracked in the video, such as the degree of tracking box jitter, number of times they are lost, etc.). Specifically, the human body quality scoring system collects auxiliary task labels: for human behavior category labels, these can be obtained by classifying human targets in videos through action recognition models; for human attribute labels, features such as gender and age can be extracted through attribute recognition models; and for human tracking stability labels, these can be calculated by analyzing the changes in the tracking box position output by the human tracking algorithm (such as the overlap rate and center offset of tracking boxes in adjacent frames). Subsequently, the system constructs an objective function based on these auxiliary task labels and the weight allocation results of step S207. The purpose of the objective function is to combine the "performance of the scoring result" with the "performance of the auxiliary task." For example, if the weight allocation is reasonable, the error between the quality scoring result and the actual quality (scoring performance index) should be small, and the accuracy of the frame images selected based on this weight when used for behavior recognition (auxiliary label performance index) should be high. The specific form of the objective function can be set according to requirements, such as weighted summation of the two performance indicators (the weights can be adjusted according to the importance of the task), so that a smaller (or larger) function value indicates better overall performance.

[0072] Optionally, the system can obtain auxiliary task labels (such as behavior categories and key attributes) of some representative frames through manual annotation as training data; calculate the quality score of each frame based on the weight allocation result of step S207, and select high-scoring frames for training the auxiliary task model (such as using high-scoring frames to train a behavior classifier); define the performance index of the scoring result as the consistency (standard deviation) of the quality scores of high-scoring frames on the test set, and the performance index of the auxiliary label as the F1 score of the auxiliary task model on high-scoring frames; construct the objective function as "(1-consistency)+(1-F1 score)", so that the smaller the function value, the better the weight can balance the scoring stability and the auxiliary task effect.

[0073] It is understandable that other methods can be used to collect auxiliary task labels and construct the objective function, such as introducing more types of auxiliary labels (such as human posture complexity labels), which are not limited here.

[0074] S209. Obtain the weights of the scoring factors corresponding to the objective function when the objective function value reaches the preset optimal range, and obtain the optimized weight allocation result.

[0075] Specifically, the human quality assessment system adjusts the weights of each scoring factor through optimization algorithms (such as gradient descent, genetic algorithms, etc.) while continuously calculating the value of the objective function: if the objective function value corresponding to the current weight does not fall into the preset optimal range, the weights are adjusted (e.g., increasing the weights of scoring factors that help improve the performance of auxiliary tasks and decreasing the weights of scoring factors that inhibit them); if the objective function value enters the optimal range, the optimization stops, and the weight value at this time is the optimized weight allocation result.

[0076] Optionally, the system sets the search range for the weights (e.g., each weight is between 0 and 1, and the sum is 1), with the preset optimal range being the objective function value being in the top 5% of all possible values; multiple initial weights are randomly generated using a genetic algorithm (including the results of step S207), and the objective function value of each weight is calculated as the fitness; the next generation of weights is generated through selection (retaining weights with high fitness), crossover (combining two weights to generate new weights), and mutation (randomly fine-tuning weights); the above steps are repeated until there is a weight set in a certain generation whose objective function value falls into the top 5%, and the weight with the highest fitness is selected as the optimized result.

[0077] It is understandable that other methods can be used to achieve weight optimization, such as particle swarm optimization, but this is not a limitation here.

[0078] S210. Sequentially obtain the quality score results of each representative frame image to obtain a quality score result sequence, and calculate the difference between two adjacent quality score results to obtain a score change sequence.

[0079] Specifically, the human body quality scoring system extracts the quality scores of representative frames in the video according to their temporal order (e.g., frame numbers from smallest to largest). For example, the scores of five representative frames are arranged in chronological order as [80, 75, 90, 65, 85], forming a quality score result sequence. Subsequently, the system calculates the difference between two adjacent scores in the sequence: the difference between the second and first frames is 75-80=-5, the difference between the third and second frames is 90-75=15, the difference between the fourth and third frames is 65-90=-25, and the difference between the fifth and fourth frames is 85-65=20, thus obtaining the score change sequence [-5, 15, -25, 20].

[0080] S211. Obtain the absolute value of each item in the score change sequence to obtain the score change range sequence, and mark the items in the score change range sequence that are greater than a preset threshold to obtain the marking result.

[0081] Specifically, the human body quality scoring system takes the absolute value of each difference in the score change sequence and converts it into a score change amplitude sequence. For example, the absolute value sequence of the score change sequence [-5, 15, -25, 20] is [5, 15, 25, 20]. Then, the system compares each element in this sequence with a preset threshold (e.g., 10): element 5 < 10 (unmarked), 15 > 10 (marked), 25 > 10 (marked), 20 > 10 (marked), resulting in the labeling result "the frames corresponding to items 2, 3, and 4 are frames with significant changes." It is important to note that the i-th item in the score change amplitude sequence corresponds to the change between the (i+1)-th and i-th frames in the quality score result sequence. Therefore, the second item in the labeling result actually corresponds to the third frame in the original sequence (the third frame in chronological order).

[0082] S212. Increase the weight of the quality score result of the representative frame image corresponding to the marking result by a preset amplitude value to obtain the preliminary quality score result corresponding to each representative frame image.

[0083] Specifically, the human body quality scoring system assigns an initial weight to each representative frame image (e.g., an average weight; if there are 5 frames, the initial weight is 0.2 for each frame). Then, based on the labeling results of step S211, the weights of significantly changed frames are increased: for example, the preset amplitude is "original weight × 1.2". Frames 3, 4, and 5 are significantly changed frames, and their initial weight of 0.2 is increased to 0.24, while the weights of the remaining frames remain at 0.2. At this point, the system needs to renormalize the weights of all frames (ensuring the sum of the weights is 1): Total weight = 0.2 (frame 1) + 0.2 (frame 2) + 0.24 (frame 3) + 0.24 (frame 4) + 0.24 (frame 5) = 1.12. After normalization, the weight = weight of each frame / 1.12, resulting in: Frame 1: 0.179, Frame 2: 0.179, Frame 3: 0.214, Frame 4: 0.214, Frame 5: 0.214. Finally, the quality score of each frame is multiplied by the adjusted weight to obtain the preliminary quality score (e.g., the preliminary score of frame 3 is 90 × 0.214 ≈ 19.26). S213. The preliminary quality score results are weighted and fused based on the temporal fusion method to obtain the overall quality score result of the human target in the video.

[0084] Specifically, the human body quality scoring system uses a temporal fusion method to perform a weighted summation of the preliminary quality scores of all representative frames. For example, the preliminary quality scores of the five frames are [80×0.179≈14.32, 75×0.179≈13.43, 90×0.214≈19.26, 65×0.214≈13.91, 85×0.214≈18.19]. Summing these values ​​gives a total of 14.32+13.43+19.26+13.91+18.19≈79.11, which is the overall quality score.

[0085] The human body mass scoring system of this invention is applied to electronic devices. Figure 3 A schematic diagram of the architecture of an electronic device suitable for implementing embodiments of the present invention is shown.

[0086] It should be noted that, Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0087] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by instructions (computer programs), or by instructions (computer programs) controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. The electronic device of this embodiment includes a storage medium and a processor, wherein the storage medium stores multiple instructions that can be loaded by the processor to execute any step of the method provided in the embodiments of the present invention.

[0088] Specifically, the storage medium and the processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more signal lines. The storage medium stores computer-executable instructions that implement data access control methods, including at least one software functional module that can be stored in the storage medium in the form of software or firmware. The processor executes various functional applications and data processing by running the software program and module stored in the storage medium. The storage medium can be, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The storage medium stores the program, and the processor executes the program after receiving the execution instructions.

[0089] Furthermore, the software programs and modules within the aforementioned storage medium may also include an operating system, which may include various software components and / or drivers for managing system tasks (e.g., memory management, storage device control, power management, etc.) and can communicate with various hardware or software components to provide an operating environment for other software components. The processor may be an integrated circuit chip with signal processing capabilities. The aforementioned processor may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc., which can implement or execute the methods, steps, and logic flowcharts disclosed in this embodiment. The general-purpose processor may be a microprocessor or any conventional processor.

[0090] Since the instructions stored in the storage medium can execute the steps in any of the methods provided in the embodiments of the present invention, the beneficial effects of any of the methods provided in the embodiments of the present invention can be achieved, as detailed in the preceding embodiments, and will not be repeated here.

[0091] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for determining the target quality level of the human body, applied to a human body quality scoring system, characterized in that, The method includes: According to the preset representative sampling rules, several representative frame images of human targets are selected from the video frame sequence. The representative frame images include the appearance frame, disappearance frame, the frame with the lowest clarity, the frame with the highest occlusion, and the motion state change frame of the human target. Multiple global scoring factors and preset local scoring factors for key human body parts are extracted from each representative frame image to obtain a set of scoring factors for each representative frame image. The set of scoring factors includes multiple scoring factors, including pose angle, completeness, sharpness, and brightness. Based on the set of scoring factors and the auxiliary information corresponding to each representative frame image, the correlation coefficient between each scoring factor is calculated using statistical analysis methods to obtain a set of correlation coefficients. The correlation coefficients are used to characterize the correlation or information dependence between the scoring factors. The step of calculating the correlation coefficient set between each rating factor based on the set of rating factors and the auxiliary information corresponding to each representative frame image, using statistical analysis methods, includes: The scoring factor values ​​corresponding to each category of scoring factors in each set of scoring factors are extracted sequentially according to the order of representative frame images to obtain the scoring factor sequence; Obtain the rating factor sequences of any two types of rating factors in all representative frame images to obtain rating factor sequence pairs, wherein the rating factor sequence pairs are combinations of the value sequences of the two types of rating factors in different frame images; By removing the parameter values ​​of auxiliary information that exceed the preset normal range threshold, the corrected scoring factor sequence pairs are obtained. The correlation coefficients of the corrected scoring factor sequence pairs were calculated using statistical analysis methods. Summarize all correlation coefficients to obtain a set of correlation coefficients, which includes the correlation coefficients between any two types of rating factors; Based on the set of correlation coefficients, a correlation coefficient matrix is ​​constructed to assign weights to each scoring factor, and the weight assignment results are obtained. Based on the weight allocation result, the set of scoring factors is weighted and calculated to obtain the quality score result of the human target on each representative frame image; The quality scores of the human target in the video are fused based on the temporal fusion method to obtain the overall quality score of the human target in the video.

2. The method according to claim 1, characterized in that, The step of constructing a correlation coefficient matrix based on the correlation coefficient set, assigning weights to each scoring factor, and obtaining the weight assignment results specifically includes: Construct a correlation coefficient matrix based on the correlation coefficients among the rating factors in the aforementioned correlation coefficient set; Calculate the sum of the absolute values ​​of the correlation coefficients of each rating factor with the other rating factors in the correlation coefficient matrix to obtain the total correlation measure of each rating factor; The total relevance measure of each scoring factor is reverse-quantified and normalized using a preset weighting rule to obtain the weighting result.

3. The method according to claim 2, characterized in that, After the step of performing reverse quantification and normalization on the total relevance measure of each scoring factor using a preset weight allocation rule to obtain the weight allocation result, the method further includes: Collect auxiliary task labels related to human target quality, including human behavior category labels, human attribute labels, and human tracking stability labels; A target function is constructed based on the auxiliary task labels and the weight allocation results. The target function includes performance indicators of the scoring results and performance indicators of the auxiliary labels. Obtain the weights of the scoring factors corresponding to the objective function when the objective function value reaches the preset optimal range, and obtain the optimized weight allocation result.

4. The method according to claim 1, characterized in that, The step of fusing the quality scores of the human target across all representative frame images using the temporal fusion method to obtain the overall quality score of the human target in the video specifically includes: The quality score results of each representative frame image are obtained sequentially to obtain a quality score result sequence, wherein the representative frame images are arranged in chronological order; Calculate the difference between two adjacent quality score results in the quality score result sequence to obtain the score change sequence; Obtain the absolute value of each item in the score change sequence to obtain the score change magnitude sequence; Items in the score change range sequence that exceed a preset threshold are marked to obtain the marking results; Increase the weight of the quality score of the representative frame image corresponding to the labeling result in the overall quality score by a preset amplitude value, and obtain the preliminary quality score result corresponding to each representative frame image; The preliminary quality score results are weighted and fused based on a temporal fusion method to obtain the overall quality score result of the human target in the video.

5. The method according to claim 1, characterized in that, The step of extracting multiple global scoring factors and preset local scoring factors of key human body parts from each representative frame image to obtain a set of scoring factors for each representative frame image further includes: Each representative frame image undergoes multi-stage spatial partitioning and downsampling to obtain feature data; Based on the aforementioned feature data, a global feature representation set covering the entire target area of ​​the human body and a local feature representation set locating key parts of the human body are extracted. Channel dimension weighted calculation and spatial adaptive pooling are performed on the global feature representation set and the local feature representation set respectively to obtain the global scoring factor set and the local scoring factor set; The global scoring factor set is combined with the local scoring factor set to obtain the scoring factor set for each representative frame image.

6. The method according to claim 1, characterized in that, After the step of fusing the quality scores of the human target across all representative frame images using the temporal fusion method to obtain the overall quality score of the human target in the video, the method further includes: The difference between the quality score of each representative frame image and the overall quality score of the human target in the video is calculated to obtain the difference sequence; Items in the difference sequence whose absolute value is greater than a preset threshold are marked with a deviation. The confidence score corresponding to the auxiliary information is used to determine the confidence weight of the target representative frame image corresponding to each deviation marker; The quality score of the representative frame image of the target is corrected by weighting the confidence score based on the confidence score, and the corrected score is obtained. Based on the corrected scoring results, the overall quality score of the human target in the video is re-fused to obtain the corrected overall quality score of the human target.

7. A human body mass assessment system, characterized in that, The human quality scoring system includes: one or more processors and memory; The memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the body mass indexing system to perform the method as described in any one of claims 1-6.

8. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on the body mass indexing system, the body mass indexing system performs the method as described in any one of claims 1-6.

9. A computer program product, characterized in that, When the computer program product is run on the human body mass indexing system, the human body mass indexing system performs the method as described in any one of claims 1-6.

Images