Cloud game interaction method and system based on human posture recognition

Abstract

The application relates to the technical field of data processing, and discloses a cloud game interaction method and system based on human body posture recognition, which comprises the following steps: acquiring a plurality of image sequences of a target area game player in real time, segmenting the preprocessed plurality of image sequences, and generating a human body area occlusion mask; the target area is an interactive platform moved by the game player; a precise human body area occlusion mask is generated by fusing a segmentation probability matrix, a complexity area graph and a boundary resource adaptive matrix, boundary deviation caused by clothes is corrected, the generated skeleton point occlusion weight graph is used to exclude the interference of clothes shaking, the skeleton points are calibrated through a skeleton point adaptive adjustment factor, a skeleton point dynamic feature vector is generated, active operation and clothes shaking are distinguished through posture motion intention entropy, a game interaction instruction set is generated in combination with a dynamic activation threshold, the accuracy of human body posture recognition is ensured, and the somatosensory interaction of cloud games is improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a cloud gaming interaction method and system based on human posture recognition. Background Technology

[0002] Currently, with the development of the Internet, the gaming industry has also ushered in rapid growth. Among them, cloud gaming, as a new form of gaming, is based on cloud computing and can run without high-performance gaming devices, making it popular among gamers. Cloud gaming is mainly controlled by mouse, keyboard, and game controller. In order to combine the control method with the player's body posture, a camera is usually used to capture changes in human posture and map these changes to standard game inputs in the cloud, thereby enriching the content of motion-sensing games and realizing the interactivity of cloud gaming.

[0003] However, the above interaction methods still have the following defects in practical applications: Since most motion-sensing cloud games require repeated movement, and in winter, game players wear more clothes, when analyzing and recognizing the game player's human posture, some skeletal points of the human body are often obscured by clothing, resulting in positioning deviations and inaccurate posture recognition, which affects the game's interaction. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a cloud gaming interaction method and system based on human posture recognition, thus solving the aforementioned problems.

[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0006] Cloud gaming interaction methods based on human pose recognition include:

[0007] Step S1 is used to acquire multi-frame image sequences of game players in the target area in real time, segment the preprocessed multi-frame image sequences, and generate human body region occlusion mask. The target area is the interactive platform where game players move.

[0008] Step S2 is used to perform occlusion analysis on the human body region in the preprocessed multi-frame image sequence based on the human body region occlusion mask, identify multiple skeletal points, and generate a skeletal point occlusion weight map.

[0009] Step S3 is used to filter different interference regions in the occlusion weight map of skeletal points and generate a skeletal point association map;

[0010] Step S4 is used to adjust the bone points in the interference area of ​​the bone point association map according to the bone point occlusion weight map, and generate the dynamic feature vector of the bone points.

[0011] Step S5 is used to analyze the dynamic feature vectors of the skeletal points, generate a game interaction instruction set that matches the motion posture of multiple skeletal points, and send the game interaction instruction set to the cloud gaming client for execution to complete the cloud gaming interaction.

[0012] Furthermore, the preprocessed multi-frame image sequence is segmented to generate human body region occlusion masks, including:

[0013] The preprocessed multi-frame image sequence is segmented to generate a segmentation probability matrix;

[0014] Texture and depth fusion analysis is performed on the human body region in the segmentation probability matrix to generate a complexity region map.

[0015] Furthermore, the preprocessed multi-frame image sequence is segmented to generate a human body region occlusion mask, which also includes:

[0016] Based on the complexity region map, the boundary pixels at different spatial scales are analyzed to obtain the boundary resource adaptation matrix;

[0017] The segmentation probability matrix and the boundary resource adaptation matrix are fused to obtain the human body region occlusion mask.

[0018] Furthermore, based on the human body region occlusion mask, occlusion analysis is performed on the human body region in the preprocessed multi-frame image sequence, and multiple skeletal points are identified to generate a skeletal point occlusion weight map, including:

[0019] Feature extraction is performed on the human body region occlusion mask to identify multiple skeletal points and generate an occlusion parse tensor.

[0020] Based on the occlusion resolution tensor, the occlusion effect of each pixel on its neighboring skeletal points is evaluated, and a spatial occlusion density field is generated.

[0021] Based on the changes in human body region morphology in the segmentation probability matrix, the temporal stability of each region in the spatial occlusion density field is calculated, and a temporal stability matrix is ​​generated.

[0022] Furthermore, based on the human body region occlusion mask, occlusion analysis is performed on the human body region in the preprocessed multi-frame image sequence, and multiple skeletal points are identified to generate a skeletal point occlusion weight map. This also includes:

[0023] The spatial occlusion density field is mapped to multiple skeleton points to generate the perturbation probability value of the skeleton points;

[0024] By reverse-analyzing the temporal stability matrix and the perturbation probability values ​​of the skeleton points, an occlusion weight map of the skeleton points is generated.

[0025] Furthermore, different interference regions in the occlusion weight map of skeletal points are filtered out to generate a skeletal point association map, including:

[0026] Gradient analysis is performed on the occlusion weight map of the skeleton points, and calibration is performed in combination with the temporal stability matrix to generate the skeleton point interference gradient entropy.

[0027] Based on the gradient entropy of skeleton point interference, high and low interference regions are divided, and a dual-domain interference discrimination mask is generated.

[0028] Furthermore, the process of filtering different interference regions in the skeletal point occlusion weight map and generating a skeletal point association map also includes:

[0029] Based on the dual-domain interference discrimination mask, reliability scores are performed on multiple skeleton points to generate skeleton point confidence bias.

[0030] Based on the confidence bias of the skeleton points, the spatial association strength and motion coordination among multiple skeleton points are calculated to generate the skeleton point association potential.

[0031] By fusing the dual-domain interference discrimination mask and the skeletal point association potential, a skeletal point association graph is generated.

[0032] Furthermore, based on the occlusion weight map of the skeleton points, the skeleton points in the interference region of the skeleton point association map are adjusted to generate dynamic feature vectors of the skeleton points, including:

[0033] Based on the dual-domain interference discrimination mask and the skeletal point occlusion weight map, the skeletal points in the high and low interference regions of the skeletal point association map are analyzed to generate skeletal point adaptation adjustment factors.

[0034] Based on the bone point adaptation adjustment factor and the bone point association potential value in the bone point association graph, the bone points are calibrated to generate dynamic feature vectors of the bone points.

[0035] Furthermore, the dynamic feature vectors of the skeletal points are analyzed to generate a set of game interaction instructions that match the motion postures of multiple skeletal points, including:

[0036] The temporal motion trajectory is calculated from the dynamic feature vectors of the skeleton points, and the collaborative intention of each skeleton point's motion is analyzed by integrating the associated potential values ​​of the skeleton points to generate the posture motion intention entropy.

[0037] Analyze the attitude motion intention entropy and dual-domain interference discrimination mask to generate a game interaction instruction set.

[0038] Furthermore, a cloud gaming interaction system based on human posture recognition, applied to the aforementioned cloud gaming interaction method based on human posture recognition, includes:

[0039] The data processing unit is used to acquire multi-frame image sequences of game players in the target area in real time, segment the pre-processed multi-frame image sequences, and generate human body region occlusion masks. The target area is the interactive platform where game players move.

[0040] The data analysis unit is used to perform occlusion analysis on the human body region in the preprocessed multi-frame image sequence based on the human body region occlusion mask, and to identify multiple skeletal points and generate a skeletal point occlusion weight map.

[0041] An interactive recognition unit is used to filter different interference regions in the occlusion weight map of skeletal points and generate a skeletal point association map.

[0042] The interactive adjustment unit is used to adjust the bone points in the interference area of ​​the bone point association map according to the bone point occlusion weight map, and generate the dynamic feature vector of the bone points.

[0043] The data interaction unit is used to analyze the dynamic feature vectors of skeletal points, generate a game interaction instruction set that matches the motion posture of multiple skeletal points, and send the game interaction instruction set to the cloud gaming client for execution to complete the cloud gaming interaction.

[0044] In summary, the present invention has the following main beneficial effects:

[0045] Step S1 generates an accurate human body occlusion mask by fusing the segmentation probability matrix, complexity region map, and boundary resource adaptation matrix, clearly defining the human body boundary and effectively reducing the interference of clothing regions on skeletal point recognition. It also corrects the human body boundary deviation caused by clothing occlusion. Step S2 accurately identifies skeletal points and generates a skeletal point occlusion weight map based on the occlusion analytic tensor, spatial occlusion density field, and temporal stability matrix, avoiding local biases in single-skeletal-point occlusion analysis and eliminating interference from clothing movement on skeletal point localization. Simultaneously, Step S3 utilizes the skeletal point interference gradient entropy... Step S4 generates a skeletal point association map by using a dual-domain interference discrimination mask, skeletal point confidence bias, and skeletal point association potential value to distinguish the association relationship between high and low interference regions and skeletal points. Step S5 completes skeletal point differential calibration based on skeletal point adaptation adjustment factor to generate dynamic feature vectors of skeletal points, accurately reflecting the player's posture changes. Finally, step S5 distinguishes between active operation and passive clothing shaking by posture motion intention entropy and generates a precise game interaction command set. This process ensures accurate skeletal point positioning and posture recognition accuracy, while preventing accidental command triggering and optimizing the cloud gaming motion experience. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the steps of the cloud gaming interaction method based on human posture recognition according to the present invention.

[0047] Figure 2 This is a schematic diagram of the cloud gaming interaction system based on human posture recognition of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] refer to Figure 1 and Figure 2 Cloud gaming interaction methods based on human pose recognition include:

[0050] Step S1 is used to acquire multi-frame image sequences of game players in the target area in real time, segment the preprocessed multi-frame image sequences, and generate human body region occlusion mask. The target area is the interactive platform where game players move.

[0051] Step S2 is used to perform occlusion analysis on the human body region in the preprocessed multi-frame image sequence based on the human body region occlusion mask, identify multiple skeletal points, and generate a skeletal point occlusion weight map.

[0052] Step S3 is used to filter different interference regions in the occlusion weight map of skeletal points and generate a skeletal point association map;

[0053] Step S4 is used to adjust the bone points in the interference area of ​​the bone point association map according to the bone point occlusion weight map, and generate the dynamic feature vector of the bone points.

[0054] Step S5 is used to analyze the dynamic feature vectors of the skeletal points, generate a game interaction instruction set that matches the motion posture of multiple skeletal points, and send the game interaction instruction set to the cloud gaming client for execution to complete the cloud gaming interaction.

[0055] In one embodiment, the preprocessed multi-frame image sequence is segmented to generate a human body region occlusion mask, including:

[0056] The preprocessed multi-frame image sequence is analyzed to generate a segmentation probability matrix, specifically including:

[0057] For multiple frames of preprocessed images, the images are divided into non-overlapping local sub-regions of square size, with the size of each sub-region set to a fixed 32×32 pixels. The sub-regions are divided sequentially from left to right and from top to bottom according to the image pixels to ensure that the local sub-regions do not overlap and completely cover the entire image.

[0058] For each local sub-region, extract its average gray value in 5 consecutive frames, such as frames 1-5, frames 2-6, etc.; calculate the absolute value of the difference between the average gray values ​​of adjacent frames in turn, and then sum all the absolute values ​​of the difference and divide by the number of inter-frame intervals to obtain the inter-frame gray value change coefficient of the local sub-region.

[0059] For each local sub-region, based on its gray-level average value of 5 consecutive frames, calculate the difference between its gray-level average value and the gray-level average value of its four neighboring local sub-regions (up, down, left, and right), and calculate the standard deviation of these four gray-level average differences. The smaller the standard deviation, the smoother the gray-level change in the neighborhood and the stronger the continuity.

[0060] Select local sub-regions whose inter-frame grayscale change coefficient is less than the mean of the inter-frame grayscale change coefficients of all local sub-regions in the multi-frame image, and whose standard deviation is less than the mean of the standard deviation of all local sub-regions in the multi-frame image. These local sub-regions are then designated as human body sub-regions; otherwise, they are designated as background regions.

[0061] For each identified human body sub-region, calculate the difference between its grayscale mean and the four neighboring human body sub-regions above, below, left, and right, and form a difference dataset. Remove the first 5% and the last 5% of the grayscale mean differences in the difference dataset, and calculate the arithmetic mean of the remaining 90% of the grayscale mean differences to obtain the continuous feature threshold of human body grayscale.

[0062] Starting with the human body sub-region, check the adjacent local sub-regions above, below, left and right in sequence. If the difference in grayscale mean between the adjacent local sub-region and the human body sub-region is less than the continuous feature threshold, and the standard deviation of the grayscale in the neighborhood of the human body sub-region is less than the standard deviation of the grayscale in the neighborhood of the local sub-region, then it is determined that the merging condition is met, and the local sub-region is taken as the human body sub-region. This step is iterated until no new regions are added, and a preliminary binary region map of the human body is generated.

[0063] In the binary human body region image, the probability value of the pixels in the human body sub-region is set to 1, and the probability value of the pixels in other local sub-regions in the binary human body region image is set to 0. A two-dimensional matrix with the same size as the original image is created, and the probability values ​​in the image are filled into the corresponding positions of the matrix to obtain the segmentation probability matrix. The segmentation probability matrix is ​​used to define the basic human body boundary for skeletal point recognition and reduce the interference of clothing areas on the selection of skeletal point candidate points.

[0064] Texture and depth fusion analysis is performed on the human body region in the segmentation probability matrix to generate a complexity region map. Specifically, this includes: for all pixels in the human body sub-region of the segmentation probability matrix, extracting the pixel value of the corresponding position in the preprocessed image; calculating the sum of the interquartile range and absolute median difference of the gray values ​​of each pixel in its 5×5 neighborhood to obtain a local texture intensity measure; simultaneously calculating the entropy value of the gradient direction angle of all pixels in this neighborhood to obtain a structural heterogeneity measure; and calculating the cumulative difference value of the gray values ​​of the same pixel in three consecutive preprocessed images. After adding 1 to the cumulative difference value, a natural logarithmic transformation is performed with a constant e as the base to obtain a temporal depth stability measure.

[0065] The fusion complexity of a pixel is obtained by multiplying the local texture intensity, structural heterogeneity metric, and temporal depth stability metric. The fusion complexity is calculated for all pixels within the human body region. The fusion complexity is then filled into a graph with the same size as the segmentation probability matrix to obtain the complexity region map. The complexity region map is used to identify clothing-occluded sub-regions within the human body region, preventing clothing-occluded regions from being treated as extensions of the human body boundary, which would result in an excessively large range for skeletal point recognition.

[0066] In one embodiment, segmenting the preprocessed multi-frame image sequence to generate a human body region occlusion mask further includes:

[0067] Based on the complexity region map, the boundary pixels at different spatial scales are analyzed to obtain the boundary resource adaptation matrix. Specifically, the following steps are taken: all zero-value pixels adjacent to the human body sub-region in the complexity region map are taken as boundary pixels; for each boundary pixel, five pixels are extended outward along its normal direction, and a fusion complexity sequence consisting of the boundary pixel and its five extended pixels is collected.

[0068] Calculate the Pearson correlation coefficient between the fusion complexity sequence and the fusion complexity sequence of pixels inside two human sub-regions adjacent to the boundary pixel in the complexity region map, and take the minimum value of the two Pearson correlation coefficients as the continuity suppression factor of the boundary pixel.

[0069] The complexity region map is downsampled by a factor of 2 and 4 to obtain a medium-scale map and a coarse-scale map, respectively. For each boundary pixel in the complexity region map, its corresponding coordinate position in the medium-scale map and the coarse-scale map is found. The fusion complexity of all pixels in the 7x7 neighborhood centered on that point in each of the three maps is extracted, and the variance of the fusion complexity of all pixels in that neighborhood is calculated. The variance is used as the response value. The response values ​​of the three maps are combined to form the scale response vector corresponding to each boundary pixel. The scale response vectors of all boundary pixels are organized into a three-dimensional matrix according to their corresponding coordinates in the image. The third dimension stores the response values ​​of the three scales. This matrix is ​​the boundary resource adaptation matrix. The boundary resource adaptation matrix is ​​used to correct the human body boundary deviation caused by clothing occlusion, so that the human body region segmentation is closer to the player's real limb contour and avoids the appearance of skeletal point candidate points on the false boundary of clothing occlusion.

[0070] The segmentation probability matrix and the boundary resource adaptation matrix are fused to obtain a human body region occlusion mask. Specifically, the process includes: using the segmentation probability matrix as an initial binary mask, extracting edge pixels of all human body sub-regions from the initial binary mask, where an edge pixel is defined as a pixel with a value of 1 and at least one 4-neighbor or 8-neighbor pixel with a value of 0; for each such edge pixel in the initial binary mask, extracting the scale response vector of its corresponding position from the boundary resource adaptation matrix, calculating the arithmetic mean of the three response values ​​in the scale response vector, adding 1 to the average value, and then calculating its natural logarithm to the base e to obtain the edge factor of the edge pixel.

[0071] The median of the edge factors of all edge pixels is calculated to obtain the adaptive culling threshold. If the edge factor of an edge pixel is greater than the adaptive culling threshold, the value of the edge pixel is corrected to 0 in the final human body region occlusion mask; otherwise, its value is retained as 1. For all non-edge pixels in the initial mask, their values ​​are retained in the final mask, thereby generating an optimized human body region occlusion mask. The human body region occlusion mask is used to eliminate the interference of clothing areas and retain only the real human body area.

[0072] By dividing local sub-regions with fixed pixels and using a dual screening process combining inter-frame grayscale variation coefficients and the mean difference and standard deviation of neighboring grayscale values, the human body sub-regions are accurately located and iteratively merged. The generated segmentation probability matrix clearly delineates the human body boundaries, reducing interference from clothing areas on the selection of candidate skeletal points. At the same time, the complexity region map can accurately identify clothing-occluded sub-regions, avoiding false boundary extensions. The boundary resource adaptation matrix can correct boundary deviations through multi-scale response and continuity suppression factors. The final generated human body region occlusion mask can accurately preserve the real human body region, ensuring the accuracy and smoothness of motion control and optimizing the cloud gaming motion interaction experience.

[0073] In one embodiment, based on the human body region occlusion mask, occlusion analysis is performed on the human body region in the preprocessed multi-frame image sequence, and multiple skeletal points are identified to generate a skeletal point occlusion weight map, including:

[0074] Feature extraction is performed on the human body region occlusion mask to identify multiple skeletal points and generate an occlusion parse tensor. Specifically, a square kernel with a side length of 5 pixels is used to dilate the human body region occlusion mask to fill the broken parts caused by occlusion in the human body sub-region with a connection value of 1.

[0075] For each pixel in the expanded human body region occlusion mask, calculate the straight-line distance to the nearest non-human sub-region pixel with a pixel value of 0. Search for points that meet the following conditions as candidate bone points: within a 7x7 area around the pixel, the straight-line distance of the point is the largest, and the straight-line distance is greater than the median value of all straight-line distances after sorting all human body region occlusion masks.

[0076] Calculate the mutual distance between all candidate bone points. If the mutual distance between two candidate bone points is within the range of 10 to 50 pixels, and at least 80% of the pixels on the connecting line segment belong to the human body sub-region, then establish a connection between the two candidate bone points. Starting from the candidate bone point with the largest mutual distance, connect the associated unused candidate bone points in sequence to form a tree-like skeleton structure. This process is repeated until it is no longer possible to connect the associated unused candidate bone points in sequence or 18 candidate bone points have been selected. Then, all candidate bone points are converted into bone points.

[0077] Centered on the coordinates of each skeletal point, a square region with sides of 11 pixels is selected. The percentage of pixels with a value of 0 in the human body occlusion mask within this square region is counted. The percentage is normalized to the 0-1 range and used as the weight. The direction vector of the line connecting each skeletal point to its parent node in the tree-like skeleton structure is found. The angle between the direction vector and the horizontal axis (positive direction of the X-axis) of the image is calculated as the angle corresponding to the direction of the skeletal point. If there is no parent node, the angle of its direction is recorded as 0. The two-dimensional coordinates, weights, mutual distances, and directions of each skeletal point are combined into a four-element feature, arranged in order from the torso of the game player to the extremities, which constitutes the occlusion resolution tensor. The occlusion resolution tensor is used to complete the preliminary identification of multiple skeletal points within the human body occlusion mask.

[0078] Based on the occlusion resolution tensor, the occlusion impact of each pixel on its neighboring skeletal points is evaluated, and a spatial occlusion density field is generated. Specifically, this includes: constructing a two-dimensional matrix with the same size as the occlusion mask of the human body region, and setting all elements in the two-dimensional matrix to zero; for each skeletal point in the occlusion resolution tensor, its coordinates, weights, mutual distances, and orientation angles are extracted, and the influence radius of the skeletal point is calculated: if its mutual distance with its parent node is greater than 0, the influence radius is twice the mutual distance; otherwise, the influence radius is 20 pixels, and the influence radius is limited to between 1 and 50 pixels.

[0079] Iterate through each pixel and calculate its Euclidean distance to the skeleton point. If the Euclidean distance is less than the influence radius, subtract the Euclidean distance from the influence radius and then divide by the influence radius to obtain the distance influence factor.

[0080] The occlusion impact value of this pixel on the current skeleton point is: the weight of the skeleton point multiplied by the distance influence factor; all occlusion impact values ​​are added together to obtain the occlusion density; the occlusion density is filled into the position of the corresponding pixel in the two-dimensional matrix. The two-dimensional matrix after filling is the spatial occlusion density field. The spatial occlusion density field is used to identify the spatial clustering area of ​​clothing occlusion, determine the joint interference of the area on multiple surrounding skeleton points, avoid analyzing the local deviation caused by the occlusion of only a single skeleton point, and improve the overall performance of pose analysis.

[0081] Based on the changes in human body region morphology in the segmentation probability matrix, the temporal stability of each region in the spatial occlusion density field is calculated, and a temporal stability matrix is ​​generated. Specifically, this includes: arranging the human body region occlusion masks corresponding to multiple frames of images according to the time of the frames to form a sequence of human body region occlusion masks; for the human body sub-region in the human body region occlusion mask of frame t in the sequence, finding the connected region that overlaps with its spatial position in the human body region occlusion mask of frame t+1; if the number of pixels in the overlapping part of two human body sub-regions accounts for more than 70% of the total number of pixels of the two, they are determined to be continuous trajectories of the same target, thereby establishing multiple human body region trajectories spanning frames, where t represents the time frame number, such as frame 1, etc.

[0082] For each human body region trajectory, calculate its three temporal morphological indices across multiple frames: the total number of pixels of the human body region trajectory in each frame, which is the area index; the ratio of the number of pixels at the boundary of the human body region trajectory to the total number of pixels, which is the boundary complexity index; and the aspect ratio of the minimum bounding rectangle of the human body region trajectory, which is the shape extensibility index. Arrange the three temporal morphological indices to form a temporal morphological index sequence, and calculate the standard deviation of the temporal morphological index sequence across all frames.

[0083] The total fluctuation value of each human body region trajectory is obtained by summing the three standard deviations of each trajectory. This total fluctuation value is then multiplied by negative two and the exponent is taken as the base of the natural constant e to obtain the temporal stability coefficient of the trajectory. The temporal stability coefficient of each human body region trajectory is used to replace all pixels covered by that trajectory in the spatial occlusion density field. If a pixel is covered by multiple human body region trajectories, the arithmetic mean of their temporal stability coefficients is used for replacement. This generates a temporal stability matrix with the same size as the spatial occlusion density field. The temporal stability matrix is ​​used to eliminate the interference of random swaying of clothing caused by limb movement on the positioning of skeletal points, making the skeletal point recognition more consistent with the continuous and stable movement trajectory of the player's limbs, and avoiding frequent jumps in skeletal point coordinates caused by clothing swaying.

[0084] In one embodiment, based on the human body region occlusion mask, occlusion analysis is performed on the human body region in the preprocessed multi-frame image sequence, and multiple skeletal points are identified to generate a skeletal point occlusion weight map. The method also includes:

[0085] Mapping the spatial occlusion density field to multiple skeleton points generates the perturbation probability value of the skeleton points. Specifically, for each skeleton point in the occlusion analytic tensor, extract all occlusion densities within a circular area with a radius of 15 pixels from the spatial occlusion density field, centered on its two-dimensional coordinates.

[0086] The spatial occlusion focus is obtained by calculating the difference between the third quartile and the first quartile of all occlusion densities within the circular region; at the same time, the temporal stability coefficients within the same circular region are extracted from the temporal stability matrix, and the arithmetic mean of these temporal stability coefficients is calculated to obtain the temporal reliability.

[0087] Divide the spatial occlusion focus by the temporal reliability, and then multiply by the weight of the skeleton point itself to obtain the intermediate interference index. Add the intermediate interference index to 1, calculate its logarithm to the natural constant 10, and normalize the final result to the 0-1 interval, which is the disturbance probability value of the skeleton point. The disturbance probability value is used to reflect the degree of interference of each skeleton point by clothing occlusion.

[0088] The temporal stability matrix and the perturbation probability values ​​of the skeleton points are analyzed in reverse to generate a skeleton point occlusion weight map. Specifically, this includes: creating a two-dimensional matrix with the same size as the human body region occlusion mask as the initial weight map and initializing all elements to zero; for each skeleton point, its perturbation probability value is used as the initial occlusion energy, centered on its coordinates, and directly replacing the corresponding single pixel value in the initial weight map.

[0089] The energy space diffusion process is performed in five rounds. In each round, each pixel in the initial weight map is traversed. If its value is greater than zero, it is regarded as an energy source. The product of the pixel value of the energy source and the temporal stability coefficient of its corresponding position in the temporal stability matrix is ​​calculated as the energy value that can be diffused in this round. This energy value is evenly distributed to the four adjacent pixels above, below, left, and right of the energy source pixel. The energy value received by each neighboring pixel is added to its pixel value in the initial weight map. After five rounds of diffusion, all pixels with a value greater than zero in the initial weight map have completed energy transfer.

[0090] Each pixel value in the initial weight map after diffusion is normalized to the 0-1 range and used as the occlusion weight. The normalized initial weight map is then the occlusion weight map of the skeleton points.

[0091] By generating an occlusion resolution tensor, preliminary accurate identification of skeletal points is achieved. The spatial occlusion density field, combined with the influence radius and distance factors, avoids local biases in single skeletal point occlusion analysis. The temporal stability matrix eliminates interference from clothing movement through three morphological indicators and temporal stability coefficients. At the same time, the skeletal point occlusion weight map generated through five rounds of energy diffusion accurately reflects the degree of occlusion interference, improves the accuracy of skeletal point positioning and posture recognition, ensures the continuity of motion control, and optimizes the cloud gaming interactive experience.

[0092] In one embodiment, filtering different interference regions in the skeletal point occlusion weight map to generate a skeletal point association map includes:

[0093] Gradient analysis is performed on the occlusion weight map of the skeleton points, and calibration is performed in conjunction with the temporal stability matrix to generate the interference gradient entropy of the skeleton points. Specifically, this includes: calculating the gradient value of each pixel in the horizontal and vertical directions in the occlusion weight map of the skeleton points; subtracting the weight of the pixel immediately to the right of the pixel from the weight of the pixel immediately to the left of the pixel, and the difference is the horizontal gradient value of the pixel; subtracting the weight of the pixel immediately below the pixel from the weight of the pixel immediately above the pixel, and the difference is the vertical gradient value of the pixel.

[0094] The original gradient magnitude of a pixel is obtained by summing the absolute values ​​of the gradients in the two directions. The temporal stability coefficients corresponding to each pixel are extracted from the temporal stability matrix, and the temporal stability coefficients are multiplied by the original gradient magnitudes to obtain the calibrated gradient magnitudes.

[0095] The calibrated gradient magnitudes of all pixels within a circular neighborhood of 5 pixels for each pixel are statistically analyzed, and the standard deviation of these calibrated gradient magnitudes is calculated to obtain the local gradient intensity heterogeneity.

[0096] Simultaneously, based on the horizontal and vertical gradient values ​​of each pixel, its gradient direction angle is calculated. The direction angle range from 0 degrees to 360 degrees is evenly divided into 8 consecutive intervals, each spanning 45 degrees, such as 0 degrees to 45 degrees, 45 degrees to 90 degrees, and so on. The interval to which the gradient direction angle of all pixels in the circular neighborhood belongs is counted, and the number of pixels in each interval is recorded. The number of pixels in each interval is divided by the total number of pixels in the neighborhood to obtain the direction distribution ratio of that interval. The sum of the squares of all direction distribution ratios is subtracted from 1 to obtain the neighborhood gradient direction dispersion. The local gradient intensity heterogeneity is multiplied by the neighborhood gradient direction dispersion, and then the logarithm is taken to the base of the natural constant e to obtain the skeleton point interference gradient entropy of the pixel. The skeleton point interference gradient entropy is used to reflect the degree of interference and accurately distinguish between high interference areas caused by clothing obstruction and low interference areas caused by normal human movement.

[0097] Based on the skeleton point interference gradient entropy, high and low interference regions are divided, and a dual-domain interference discrimination mask is generated. Specifically, the process includes: calculating the median of the skeleton point interference gradient entropy of all pixels to obtain the global baseline entropy value; and initially marking all pixels with a value greater than or equal to the global baseline entropy value as candidate high interference points.

[0098] For each candidate high interference point, calculate the number of pixels in a circular neighborhood with a radius of three pixels that are also marked as candidate high interference points. If this number is less than two, the point is considered an isolated misjudged point and its mark is removed.

[0099] Starting from any remaining candidate high interference point, use it as a seed point. Check the pixels in the eight-neighborhood of the seed point. If the skeleton point interference gradient entropy of the neighboring pixels is greater than or equal to 70% of the global baseline entropy value, and it is also a candidate high interference point, then include it in the candidate region. By iterating this process, until the candidate region can no longer be expanded, the candidate region with a total number of pixels greater than 20 is determined as the high interference region, and the remaining regions are classified as low interference regions.

[0100] Create a binary mask with the same size as the temporal stability matrix, set the pixel values ​​of all high-interference regions to 1, and set the pixel values ​​of all low-interference regions to 0. This will give you a dual-domain interference discrimination mask. The dual-domain interference discrimination mask is used to determine whether the environment in which the skeleton point is located is a high-interference region or a low-interference region, thus achieving differentiated processing of the skeleton point.

[0101] In one embodiment, the process of filtering different interference regions in the skeletal point occlusion weight map and generating a skeletal point association map further includes:

[0102] Based on the dual-domain interference discrimination mask, reliability scores are performed on multiple skeleton points to generate skeleton point confidence bias. Specifically, for each skeleton point in the occlusion parsing tensor, a square verification region with a side length of 7 pixels is delineated with its coordinates as the center.

[0103] Within the verification area, the proportion of high-interference pixels with a value of 1 in the dual-domain interference discrimination mask is counted. If the proportion exceeds 30%, the current coordinates of the skeleton point are determined to be in a high-interference environment. For the skeleton point that is determined to be high-interference, the time series of the corresponding coordinate points of the skeleton point in 5 consecutive frames of images is extracted, the Euclidean distance between the coordinates of two adjacent frames in the time series is calculated, and the median of the 4 Euclidean distances is used as the instantaneous displacement reference of the skeleton point.

[0104] Simultaneously, the temporal stability coefficient corresponding to the coordinate point is obtained from the temporal stability matrix. The instantaneous displacement reference is divided by the sum of the temporal stability coefficient and 0.1 to obtain the displacement anomaly index. Then, the proportion of high interference pixels is multiplied by the displacement anomaly index, and then a logarithmic transformation with the natural constant e as the base is performed to obtain the initial reliability score of the skeleton point.

[0105] For skeletal points not identified as high-interference, their initial reliability score is directly set to 1. Based on the human tree-like skeleton structure, the Euclidean distance between each skeletal point and its parent node coordinates is calculated. The Euclidean distance is compared with the corresponding mutual distance in the occlusion resolution tensor. If the deviation is greater than 5 pixels, the initial reliability score is multiplied by 0.7 for reduction, and the reciprocal of the reduced score is taken to obtain the confidence bias of each skeletal point. Otherwise, the reciprocal of the initial reliability score is directly used as the skeletal point confidence bias. The skeletal point confidence bias is used to reflect the location reliability of each skeletal point.

[0106] Based on the confidence bias of the skeleton points, the spatial association strength and motion coordination between multiple skeleton points are calculated to generate the skeleton point association potential value. Specifically, for any two skeleton points, the Euclidean distance between the two skeleton points is calculated, the Euclidean distance is subtracted from the product of the corresponding skeleton point confidence biases, and the calculation result is normalized to the 0-1 interval, which is the spatial association strength between the two skeleton points.

[0107] Extract the coordinate sequence of two skeleton points over 7 consecutive frames. Calculate the displacement vector of each skeleton point in each frame relative to its own coordinates in the previous frame. Calculate the average of the dot product of the two displacement vectors over these 7 frames to obtain the original motion covariance. Calculate the harmonic mean of the confidence biases of the two skeleton points to obtain the bias harmonic term. Divide the original motion covariance by the sum of the bias harmonic term and 1. Then, transform the quotient using an S-shaped function to obtain the motion coordination between the two skeleton points.

[0108] Multiplying the calculated spatial association strength by the motion synergy and normalizing the result to the 0-1 interval yields the skeletal point association potential value between two skeletal points. The skeletal point association potential value is used to reflect the natural association and motion synergy between skeletal points.

[0109] The fusion of a dual-domain interference discrimination mask and skeletal point association potentials generates a skeletal point association map. Specifically, this involves: creating a three-channel color image of the same size as the dual-domain interference discrimination mask as the association map background; filling the background with dark gray pixels marked as low-interference regions in the dual-domain interference discrimination mask; filling the background with light gray pixels in high-interference regions; and then mapping the coordinates of multiple skeletal points in the occlusion resolution tensor to the association map background.

[0110] Using each coordinate point as the center, iterate through all pairs of bone points. If the bone point correlation potential between two bone points is greater than 0.5, draw a connecting line between the centers of the two bone points. If the bone point correlation potential between two bone points is between 0.51 and 0.8, the line color is a gradient from dark blue to cyan; if the bone point correlation potential between two bone points is greater than 0.8, the line color is a gradient from yellow to orange.

[0111] Calculate the proportion of pixels belonging to the high interference region among all pixels passed through by each connecting line. If this proportion exceeds 50%, modify the style of the entire connecting line to a dashed line with 30% transparency. The resulting fused image is a skeletal point association map that simultaneously marks the high and low interference regions, the positions of multiple skeletal points, and the relationships between the skeletal points.

[0112] By accurately distinguishing high-interference areas caused by clothing occlusion from low-interference areas caused by normal human movement through the gradient entropy of skeletal point interference, the mask marking region type is distinguished through dual-domain interference. The positional confidence of each skeletal point is reflected by the confidence bias of the skeletal point, and the spatial correlation and motion coordination between skeletal points are reflected by the correlation potential value of the skeletal points. The final generated skeletal point correlation map marks the high and low interference areas, the position of the skeletal points and the correlation relationship. The correlation strength is presented by differentiated lines, which further corrects the occlusion bias and improves the coherence and accuracy of pose recognition.

[0113] In one embodiment, the bone points in the interference region of the bone point association map are adjusted according to the bone point occlusion weight map to generate a dynamic feature vector of the bone points, including:

[0114] Based on the dual-domain interference discrimination mask and the skeletal point occlusion weight map, the skeletal points in the high and low interference regions of the skeletal point association map are analyzed to generate skeletal point adaptation adjustment factors. Specifically, for each skeletal point, all pixels in a circular area with a radius of 9 pixels are extracted with its coordinates as the center. The proportion of pixels in the circular area that belong to the high interference region in the dual-domain interference discrimination mask is counted and recorded as the high interference density.

[0115] Next, extract the occlusion weights of all pixels within the same circular region from the skeletal point occlusion weight map, calculate the difference between the 90th percentile and the 10th percentile of these occlusion weights to obtain the local weight range; and simultaneously calculate the geometric mean of the occlusion weights of all pixels within the circular region.

[0116] Multiply the high interference density by the local weight range to obtain the base adjustment amount. Add the reciprocal of the geometric mean to the base adjustment amount and normalize the calculation result to the 0-1 interval to obtain the bone point adaptation adjustment factor. The bone point adaptation adjustment factor is used to reflect the adaptation adjustment range of each bone point, avoiding indiscriminate adjustment of bone points with different occlusion interference.

[0117] Based on the bone point adaptation adjustment factor and the bone point association potential value in the bone point association graph, the bone points are calibrated to generate a dynamic feature vector of the bone points. Specifically, this includes: creating a dynamic feature list containing multiple entries; for each bone point in the occlusion parsing tensor, its coordinates, weights, and orientations are stored as basic feature triples in the dynamic feature list.

[0118] For each skeleton point in the dynamic feature list, calculate the difference between its position and the average position of its coordinates in the previous 5 consecutive frames to generate a historical trajectory perturbation vector. Multiply the historical trajectory perturbation vector by the skeleton point adaptation adjustment factor of that skeleton point. If the skeleton point adaptation adjustment factor > 0.5, then multiply the historical trajectory perturbation vector by the skeleton point adaptation adjustment factor itself for enhancement. If the skeleton point adaptation adjustment factor ≤ 0.5, then multiply the historical trajectory perturbation vector by the square root of the skeleton point adaptation adjustment factor for suppression. Add the corrected historical trajectory perturbation vector to the original coordinates of the current skeleton point to obtain the first corrected coordinates.

[0119] Traverse all bone points, find the three other bone points with the highest associated potential value of the current bone point, calculate the geometric center of the first corrected coordinates of these three bone points, calculate the vector from the first corrected coordinates of the current bone point to the geometric center, multiply this vector by the average of the associated potential values ​​of these three bone points to obtain a traction vector, add this traction vector to the first corrected coordinates of the current bone point to obtain the final calibration coordinates.

[0120] The final calibration coordinates, weights, orientations, and bone point adaptation adjustment factors of each bone point are arranged in order to form a 4-dimensional sub-vector. All the 4-dimensional sub-vectors of bone points are connected in a predetermined order from the torso to the extremities to form a 72-dimensional dynamic feature vector of bone points.

[0121] In one embodiment, the dynamic feature vectors of the skeletal points are analyzed to generate a set of game interaction instructions that match the motion postures of multiple skeletal points, including:

[0122] The dynamic feature vectors of skeleton points are used to calculate the temporal motion trajectory, and the associated potential values ​​of skeleton points are fused to analyze the collaborative intention of each skeleton point's motion, generating the posture motion intention entropy. Specifically, this includes: extracting the dynamic feature vectors of skeleton points from the current and past 8 consecutive frames; for each skeleton point, obtaining the sequence of its final calibration coordinates from these 8 frames; calculating the displacement vector between two adjacent frames in the sequence; and calculating the magnitude and coefficient of variation of these 7 displacement vectors to obtain the individual motion fluctuation of the skeleton point.

[0123] For any two skeleton points, check if their skeleton point association potential is greater than 0.5. If it is greater than 0.5, perform a dot product operation on the displacement vectors of the two skeleton points in this frame to obtain the instantaneous cooperative projection value. Count all skeleton point pairs that meet the conditions, calculate the average value of all instantaneous cooperative projection values ​​to obtain the intra-frame cooperative strength. At the same time, calculate the standard deviation of the absolute value of the difference between the direction angles of the displacement vectors of these skeleton point pairs in this frame as the direction dispersion.

[0124] The mean of the individual motion fluctuation of all bone points in the current frame is used as the motion reference; the intra-frame coordination intensity is divided by the directional dispersion to obtain the coordination index; the coordination index is subtracted from the motion reference, and the absolute value of the calculation result is normalized to the 0-1 interval to obtain the posture motion intention entropy. The posture motion intention entropy is used to accurately distinguish between the player's explicit active game operation and the vague passive shaking caused by clothing obstruction.

[0125] Analyze the posture motion intention entropy and dual-domain interference discrimination mask to generate a game interaction instruction set, specifically including: predefining a game instruction library, in which each instruction, such as the jump instruction, the left movement instruction, etc., is bound to the human posture change mode. For example, the trigger mode of the jump instruction is that the coordinates of the bone points of both ankles in the current frame are simultaneously greater than 1.2 times the average coordinates of their respective bone points in the previous 3 frames.

[0126] Based on multiple skeletal points in the posture change pattern of the instruction, the coordinates of the corresponding skeletal points in multiple consecutive frames are extracted to form an original coordinate sequence, and the mean of the original coordinate sequence is calculated to obtain a dynamic comparison benchmark.

[0127] For example, for the jump command, it is defined as the coordinates of the bone points of both ankles in the current frame, which are greater than 1.2 times the average coordinates of their respective bone points in the previous 3 frames. Then, the coordinates of the current frame are divided by the dynamic comparison benchmark to obtain the actual multiple. The actual multiple is normalized to the 0-1 range to obtain the original compliance.

[0128] Identify all skeletal points involved in the human posture change pattern of the instruction, and count the proportion of these skeletal points whose coordinates fall in the high interference region of the dual-domain interference discrimination mask, which is denoted as the high interference ratio; subtract the high interference ratio from 1 and multiply by 0.6 to obtain the region penalty coefficient.

[0129] Calculate the geometric mean of the bone point adaptation adjustment factors of the involved bone points to obtain the reliability coefficient; multiply the original conformity, the region penalty coefficient and the reliability coefficient to obtain the comprehensive trigger strength of the instruction in the current frame.

[0130] Multiply the attitude motion intent entropy by 0.35 and add 0.25 to obtain the dynamic activation threshold of the current frame. This allows the dynamic activation threshold to automatically increase when the attitude motion intent entropy is high to prevent false triggering, and to decrease when the attitude motion intent entropy is low to facilitate sensitive triggering.

[0131] An instruction is activated only when its overall trigger strength is greater than the dynamic activation threshold; otherwise, it is not activated. All activated game instructions are sorted from high to low according to their overall trigger strength to form the game interaction instruction set for the current frame, and then sent to the cloud gaming client for execution in real time, thereby completing the closed-loop interaction from human posture to game control.

[0132] Differential adjustments are achieved through skeletal point adaptation adjustment factors to avoid errors from indiscriminate calibration. Combined with skeletal point correlation potential value calibration coordinates, dynamic feature vectors of skeletal points are generated to accurately reflect the player's posture. Based on the posture motion intention entropy, the player's active operation and the passive swaying of clothing are distinguished. The final generated game interaction command set is sorted and issued according to the comprehensive trigger intensity. This not only corrects the skeletal point deviation caused by occlusion, but also prevents the command from being triggered erroneously, ensuring the accuracy of posture recognition and the continuity of control, and improving the player's haptic experience in cloud gaming.

[0133] In one embodiment, the cloud gaming interaction system based on human posture recognition, applied to the aforementioned cloud gaming interaction method based on human posture recognition, includes:

[0134] The data processing unit is used to acquire multi-frame image sequences of game players in the target area in real time, segment the pre-processed multi-frame image sequences, and generate human body region occlusion masks. The target area is the interactive platform where game players move.

[0135] The data analysis unit is used to perform occlusion analysis on the human body region in the preprocessed multi-frame image sequence based on the human body region occlusion mask, and to identify multiple skeletal points and generate a skeletal point occlusion weight map.

[0136] An interactive recognition unit is used to filter different interference regions in the occlusion weight map of skeletal points and generate a skeletal point association map.

[0137] The interactive adjustment unit is used to adjust the bone points in the interference area of ​​the bone point association map according to the bone point occlusion weight map, and generate the dynamic feature vector of the bone points.

[0138] The data interaction unit is used to analyze the dynamic feature vectors of skeletal points, generate a game interaction instruction set that matches the motion posture of multiple skeletal points, and send the game interaction instruction set to the cloud gaming client for execution to complete the cloud gaming interaction.

[0139] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cloud game interaction method based on human body posture recognition, characterized in that, include: Step S1: Acquire multi-frame image sequences of game players in the target area in real time, segment the preprocessed multi-frame image sequences, and generate human body region occlusion masks, including: The preprocessed multi-frame image sequence is segmented to generate a segmentation probability matrix. Specifically, the process includes: dividing the preprocessed multi-frame images into local sub-regions, calculating the inter-frame grayscale change coefficient of the local sub-regions, dividing the local sub-regions into human body sub-regions and background regions based on the inter-frame grayscale change coefficients, generating a preliminary human body binary region map, and obtaining the segmentation probability matrix based on the preliminary human body binary region map. The texture and depth fusion analysis of the human body region in the segmentation probability matrix is ​​performed to generate a complexity region map. Specifically, this includes: extracting and analyzing all pixels that are human body sub-regions in the segmentation probability matrix to obtain local texture intensity measure, structural heterogeneity measure and temporal depth stability measure, and calculating the three to obtain the complexity region map. Based on the complexity region map, the boundary pixels at different spatial scales are analyzed to obtain the boundary resource adaptation matrix. Specifically, this includes: identifying the boundary pixels in the complexity region map and analyzing them to obtain the fusion complexity sequence; then calculating the fusion complexity sequence and the complexity region map to generate the scale response vector; and combining the scale response vectors to form the boundary resource adaptation matrix. The segmentation probability matrix and the boundary resource adaptation matrix are fused to obtain the human body region occlusion mask. Specifically, the segmentation probability matrix is ​​used as the initial binary mask to extract and analyze the edge pixels of the human body sub-region, obtain the edge factor and adaptive culling threshold of the edge pixel, compare the edge factor and adaptive culling threshold, correct the value of the edge pixel, and generate the human body region occlusion mask. The target area is an interactive platform for mobile game players; Step S2: Based on the human body region occlusion mask, perform occlusion analysis on the human body region in the preprocessed multi-frame image sequence, identify multiple skeletal points, and generate a skeletal point occlusion weight map. Step S3: Filter different interference regions in the occlusion weight map of bone points and generate a bone point association map; Step S4: Adjust the bone points in the interference area of ​​the bone point association map according to the bone point occlusion weight map to generate a dynamic feature vector of bone points. Step S5: Analyze the dynamic feature vectors of the skeletal points, generate a game interaction instruction set that matches the motion posture of multiple skeletal points, and send the game interaction instruction set to the cloud gaming client for execution to complete the cloud gaming interaction.

2. The cloud game interaction method based on human body posture recognition according to claim 1, characterized in that, Based on the human body region occlusion mask, occlusion analysis is performed on the human body region in the preprocessed multi-frame image sequence, and multiple skeletal points are identified to generate a skeletal point occlusion weight map, including: Feature extraction is performed on the human body region occlusion mask to identify multiple skeletal points and generate an occlusion parse tensor. Based on the occlusion resolution tensor, the occlusion effect of each pixel on its neighboring skeletal points is evaluated, and a spatial occlusion density field is generated. Based on the changes in human body region morphology in the segmentation probability matrix, the temporal stability of each region in the spatial occlusion density field is calculated, and a temporal stability matrix is ​​generated.

3. The cloud game interaction method based on human body posture recognition according to claim 2, characterized in that, Based on the human body region occlusion mask, occlusion analysis is performed on the human body region in the preprocessed multi-frame image sequence, and multiple skeletal points are identified to generate a skeletal point occlusion weight map. This also includes: The spatial occlusion density field is mapped to multiple skeleton points to generate the perturbation probability value of the skeleton points; By reverse-analyzing the temporal stability matrix and the perturbation probability values ​​of the skeleton points, an occlusion weight map of the skeleton points is generated.

4. The cloud gaming interaction method based on human pose recognition according to claim 3, characterized in that, Filter different interference regions in the occlusion weight map of skeletal points to generate a skeletal point association map, including: Gradient analysis is performed on the occlusion weight map of the skeleton points, and calibration is performed in combination with the temporal stability matrix to generate the skeleton point interference gradient entropy. Based on the gradient entropy of skeleton point interference, high and low interference regions are divided, and a dual-domain interference discrimination mask is generated.

5. The cloud game interaction method based on human body posture recognition according to claim 4, characterized in that, The process of filtering different interference regions in the occlusion weight map of skeletal points and generating a skeletal point association map also includes: Based on the dual-domain interference discrimination mask, reliability scores are performed on multiple skeleton points to generate skeleton point confidence bias. Based on the confidence bias of the skeleton points, the spatial association strength and motion coordination among multiple skeleton points are calculated to generate the skeleton point association potential. By fusing the dual-domain interference discrimination mask and the skeletal point association potential, a skeletal point association graph is generated.

6. The cloud game interaction method based on human body posture recognition according to claim 5, characterized in that, Based on the occlusion weight map of the skeleton points, the skeleton points in the interference region of the skeleton point association map are adjusted to generate dynamic feature vectors of the skeleton points, including: Based on the dual-domain interference discrimination mask and the skeletal point occlusion weight map, the skeletal points in the high and low interference regions of the skeletal point association map are analyzed to generate skeletal point adaptation adjustment factors. Based on the bone point adaptation adjustment factor and the bone point association potential value in the bone point association graph, the bone points are calibrated to generate dynamic feature vectors of the bone points.

7. The cloud game interaction method based on human body posture recognition according to claim 6, characterized in that, The dynamic feature vectors of skeletal points are analyzed to generate a set of game interaction instructions that match the motion postures of multiple skeletal points, including: The temporal motion trajectory is calculated from the dynamic feature vectors of the skeleton points, and the collaborative intention of each skeleton point's motion is analyzed by integrating the associated potential values ​​of the skeleton points to generate the posture motion intention entropy. Analyze the attitude motion intention entropy and dual-domain interference discrimination mask to generate a game interaction instruction set.

8. A cloud gaming interaction system based on human posture recognition, applied to the cloud gaming interaction method based on human posture recognition as described in any one of claims 1-7, characterized in that, include: The data processing unit is used to acquire multi-frame image sequences of game players in the target area in real time, segment the pre-processed multi-frame image sequences, and generate human body region occlusion masks. The target area is the interactive platform where game players move. The data analysis unit is used to perform occlusion analysis on the human body region in the preprocessed multi-frame image sequence based on the human body region occlusion mask, and to identify multiple skeletal points and generate a skeletal point occlusion weight map. An interactive recognition unit is used to filter different interference regions in the occlusion weight map of skeletal points and generate a skeletal point association map. The interactive adjustment unit is used to adjust the bone points in the interference area of ​​the bone point association map according to the bone point occlusion weight map, and generate the dynamic feature vector of the bone points. The data interaction unit is used to analyze the dynamic feature vectors of skeletal points, generate a game interaction instruction set that matches the motion posture of multiple skeletal points, and send the game interaction instruction set to the cloud gaming client for execution to complete the cloud gaming interaction.

Images