A progressive head pose estimation method based on a probability diffusion model

Abstract

This invention provides a progressive head pose estimation method based on a probability diffusion model, comprising: establishing a head pose estimation loop, the loop including the following steps: detecting the human head pose based on a first probability diffusion model and a second color image to obtain a human head pose detection result, which is then used as a second head pose detection result; optimizing the first probability diffusion model based on the second color image and the second head pose detection result to obtain a second probability diffusion model; upon completion of the current loop, inputting the second probability diffusion model from the current loop as the first probability diffusion model for the next loop. This invention addresses the technical problem of low accuracy in head pose estimation due to the inability to obtain complete 3D head data. By proposing a progressive head pose estimation method, it continuously improves the accuracy of the 3D head data, thereby enhancing the accuracy of the head pose estimation results.

Description

Technical Field

[0001] This invention relates to the field of data recognition technology, and more specifically, to a progressive head pose estimation method based on a probability diffusion model. Background Technology

[0002] As people become more health-conscious, their demands for office desks are also increasing. Currently, incorrect posture while using office desks often leads to cervical spondylosis and other health problems. Therefore, based on these reasons, a type of yoga desk has been developed to detect, identify, and record the user's posture. Users can adjust their posture based on the detected information. In the field of data recognition and processing for human posture estimation, complete three-dimensional head data has a decisive impact on the accuracy of head posture estimation at extreme angles.

[0003] However, in the actual data processing of existing yoga tables, it is difficult to obtain complete 3D head data of users in advance, resulting in insufficient 3D head data for posture estimation, which ultimately leads to inaccurate estimation results of human posture. Summary of the Invention

[0004] The problem addressed by this invention is the technical issue of low accuracy in head pose estimation due to the inability to obtain complete 3D head data of the user. By proposing a progressive head pose estimation method based on a probability diffusion model, the 3D head data can be continuously improved, thereby enhancing the accuracy of head pose estimation results.

[0005] To address the aforementioned problems, this invention provides a progressive head pose estimation method based on a probability diffusion model. This method utilizes a probability diffusion model to estimate the head pose of a human body. The method includes: establishing a head pose estimation loop, which comprises the following steps: Step S100: Obtaining the probability diffusion model at a first time moment, as the first probability diffusion model; Step S200: Obtaining a color image of the human head at a second time moment, as the second color image; Step S300: Detecting the human head pose based on the first probability diffusion model and the second color image, obtaining the human head pose detection result at the second time moment, and using it as the second head pose detection result; Step S400: Inputting the second head pose detection result into a head pose estimation algorithm to obtain the human head pose estimation result at the second time moment, and outputting it; Step S500: Optimizing the first probability diffusion model based on the second color image and the second head pose detection result, obtaining the probability diffusion model at the second time moment, as the second probability diffusion model; After the current loop is completed, the second probability diffusion model under this loop is used as the first probability diffusion model for the next loop and input into the next loop; The second time moment is obtained after the first time moment passes through a target time period.

[0006] Compared to existing technologies, the technical effects achieved by this solution are as follows: The human body is constantly in motion, and its state changes almost every second. Therefore, a target time period of 1 second is optimally chosen. Furthermore, when reconstructing the head's 3D data through posture detection, the probability diffusion model from the previous second and the image acquired at the current time are combined. This allows the head's 3D data to continuously improve over time, resulting in more accurate head posture estimation after being incorporated into the head posture estimation algorithm. In the entire posture estimation method, the probability diffusion model from the previous time step is continuously used for detection and estimation at each moment, and this model is constantly updated for use in the next moment. This continuous learning and improvement of the probability diffusion model allows the entire estimation method to learn and optimize over time, leading to increasingly accurate estimation results.

[0007] In one embodiment of the present invention, if the head pose estimation loop is in the initial state, the steps are as follows: Step S110: Obtain a color image of the human head at the first moment and use it as the first color image; Step S120: Detect the human head pose according to the original probability diffusion model and the first color image to obtain the human head pose detection result at the first moment and use it as the first head pose detection result; Step S130: Input the first head pose detection result into the head pose estimation algorithm to obtain the human head pose estimation result at the first moment and output it; Use the original probability diffusion model as the first probability diffusion model and use it in the second moment, where the original probability diffusion model is the probability diffusion model set at the beginning of the method; If the head pose estimation loop is not in the initial state, then Step S120 is as follows: Detect the human head pose according to the first color image and the second probability diffusion model obtained in the previous loop to obtain the human head pose detection result at the first moment and use it as the first head pose detection result.

[0008] Compared to existing technologies, the technical effects achieved by this solution are as follows: the probability diffusion model itself has the ability to construct the required data from noise, and can be used to improve and construct 3D head data. To continuously update and optimize the probability diffusion model, this solution uses the data from the current moment to optimize and generate the probability diffusion model from the previous moment. After optimization, the probability diffusion model for the current moment is obtained, and this model is then used to obtain the detection results for future moments, thus achieving progressive improvement and iteration of the data. After obtaining the human head pose detection results at the second moment, the probability diffusion model is further optimized based on the results, updating and improving the data information of the probability diffusion model to obtain more accurate and complete head detection results for the next moment.

[0009] In one embodiment of the present invention, the method further includes: the progressive head pose estimation method based on the probability diffusion model stops execution in response to a stop command; after the progressive head pose estimation method based on the probability diffusion model stops execution, the latest output human pose head detection result and the probability diffusion model are stored as stored data.

[0010] Compared with existing technologies, the technical effects achieved by this solution are as follows: This solution is an estimation method that learns progressively over time. When a stop instruction is given to the method at a certain point, it is necessary to control the storage of the latest output estimation results and probability diffusion model. The stored data is used for the next execution of the progressive head pose estimation method.

[0011] In one embodiment of the present invention, step S500: optimizing the first probability diffusion model based on the second color image and the second head pose detection result to obtain the probability diffusion model at the second time step, which serves as the second probability diffusion model, includes: obtaining an inverse Markov chain based on the first probability diffusion model; constructing a loss function for each link in the inverse Markov chain by maximizing the likelihood function and minimizing the KL divergence; extracting parameters from the second color image and the second head pose detection result as optimization parameters; inputting the optimization parameters and random noise into the first probability diffusion model, and training the first probability diffusion model using a gradient descent algorithm based on the loss function to obtain the second probability diffusion model.

[0012] Compared with existing technologies, the technical effect achieved by adopting this technical solution is that by using the optimization parameters and loss function at the second time step to optimize the first probability diffusion model, the optimization parameter data at the second time step can be updated, making the results more accurate.

[0013] In one embodiment of the present invention, obtaining an inverse Markov chain based on a first probability diffusion model includes: a first process of transforming a hidden space random vector in the first probability diffusion model into a key point in the model space; modeling the first process to obtain an inverse Markov random process and forming an inverse Markov chain; wherein the conditional probability and transition function of each link in the inverse Markov chain are determined according to the parameters in the first probability diffusion model.

[0014] Compared with existing technologies, the technical effects achieved by this solution are as follows: In the process of obtaining the inverse Markov chain based on the first probability diffusion model, specifically by modeling the process of transforming the hidden space random vector into the model space key points, the conditional probability and transition function of each link in the process of obtaining the inverse Markov chain are determined according to the parameters in the first probability diffusion model. At this time, the inverse process is obtained on the basis of the first probability diffusion model. Therefore, the first probability diffusion model serves as the optimization basis, so that the originally optimized model no longer uses outdated parameters and data.

[0015] In one embodiment of the present invention, the head pose of a human body is detected based on a first probability diffusion model and a second color image to obtain a head pose detection result at a second time point, which is then used as the second head pose detection result. This includes: determining whether the second color image includes human face information; if it includes human face information, the pose of the human face is detected to obtain a first detection result, which is then used as the second head pose detection result; if it does not include human face information, the pose of the human head is detected to obtain a second detection result, which is then used as the second head pose detection result.

[0016] Compared with existing technologies, the technical effects achieved by adopting this technical solution are as follows: first determining whether it contains facial information, and then performing detection, can save detection steps, avoid obtaining unnecessary data, and improve detection speed and efficiency.

[0017] In one embodiment of the present invention, if human face information is included, the pose of the human face is detected to obtain a first detection result, and the first detection result is used as a second head pose detection result, including: using Hopenet to perform key point detection on the human face to obtain facial key point information; obtaining the first detection result based on the facial key point information; and outputting the first detection result as a second head pose detection result.

[0018] Compared with existing technologies, the technical advantages of this solution are as follows: Facial pose detection is particularly important in human head pose detection. This solution utilizes Hopenet to solve the problem of 2D-to-3D correspondence through key points on the face. This method has higher accuracy and is more accurate than traditional detection methods.

[0019] In one embodiment of the present invention, if facial information is not included, the pose of the human head is detected to obtain a second detection result, and the second detection result is used as the second head pose detection result. This includes: obtaining the top normal vector of the human head and the three-dimensional point data of the human head based on the second color image, and using the top normal vector as preliminary pose information; detecting the human head pose based on a first probability diffusion model to obtain point cloud voxel data of the human head; minimizing the projection error between the point cloud voxel data and the three-dimensional point data based on the preliminary pose information to obtain the second detection result; and outputting the second detection result as the second head pose detection result.

[0020] Compared to existing technologies, this technical solution achieves the following advantages: when a face cannot be recognized, only the human head can be identified and detected. When detecting head pose, the top normal vector of the human head is relatively easy to obtain and can be used as preliminary pose information. After obtaining the preliminary pose information, to obtain more refined head pose information, it is necessary to continuously refine it using a first probability diffusion model to obtain more accurate pose detection results. Specifically, the refinement method involves minimizing the projection error between the point cloud voxel data and the 3D point data. Minimization improves data accuracy and yields more refined detection results.

[0021] In one embodiment of the present invention, the head pose of a human body is detected according to a first probability diffusion model to obtain point cloud voxel data of the human head, including: removing Gaussian noise from the point cloud voxel data by combining a denoising diffusion model; and / or training the first probability diffusion model according to a conditional likelihood function based on the second norm to obtain noise-removed point cloud voxel data.

[0022] Compared with existing technologies, the technical effects achieved by this solution are as follows: In the process of reconstructing human head pose data using a probability diffusion model, two denoising methods can be selected to reduce noise interference. One method utilizes a denoising diffusion model, while the other uses a conditional likelihood function based on the L2 norm. These two methods ultimately yield more refined point cloud voxel data, further improving the accuracy of head pose estimation results. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings to be used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is one of the schematic diagrams illustrating the steps of a progressive head pose estimation method provided by the present invention;

[0025] Figure 2 The second schematic diagram illustrates the steps of a progressive head pose estimation method provided by the present invention.

[0026] Figure 3 The third schematic diagram illustrates the steps of a progressive head pose estimation method provided by this invention. Detailed Implementation

[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0028] Example 1

[0029] In one specific embodiment, see Figure 1-2 This paper proposes a progressive head pose estimation method based on a probability diffusion model. This method uses a probability diffusion model to estimate the head pose of the human body. The method includes:

[0030] Establish a head pose estimation loop, which includes the following steps:

[0031] Step S100: Obtain the probability diffusion model at the first time step, as the first probability diffusion model;

[0032] Step S200: Obtain a color image of the human head at the second time point and use it as the second color image;

[0033] Step S300: Detect the head pose of the human body according to the first probability diffusion model and the second color image to obtain the head pose detection result at the second time point, and use it as the second head pose detection result.

[0034] Step S400: Input the second head pose detection result into the head pose estimation algorithm to obtain the human head pose estimation result at the second time point, and output it.

[0035] Step S500: Optimize the first probability diffusion model based on the second color image and the second head pose detection results to obtain the probability diffusion model at the second time step, which is used as the second probability diffusion model;

[0036] After the current loop is completed, the second probability diffusion model under that loop is used as the first probability diffusion model for the next loop and input into the next loop.

[0037] The second time point is obtained after the first time point passes through the target time period.

[0038] In this embodiment, the probabilistic diffusion model, also known as Diffusion Models, is based on assumptions. By finding and understanding each step of the diffusion process and the current state, it works by reversing the previous state step by step from an initial isotropic Gaussian noise distribution, thus obtaining the original state of the particle. The probabilistic diffusion model itself has the ability to construct the required data from noise, and therefore can be used to refine and construct the 3D head data. In step 200, in addition to acquiring color images, depth images can also be acquired for detection and estimation. The color image and depth image are RGB images and depth video frames read from the RGBD camera. Depth images, also known as range images, refer to images that use the distance values ​​of each point in the scene acquired by the image acquisition device as pixel values. They directly reflect the geometry of the visible surface of the scene. Using depth images can conveniently solve many problems in 3D target description.

[0039] Furthermore, probability diffusion models, color images, and depth images can all serve as information sources for reconstructing 3D head data. However, to ensure more accurate head pose estimation, the process of constructing 3D head data and performing pose detection in this case is a gradual one, continuously refined over time based on previously obtained data to obtain more complete pose detection results.

[0040] For details, see Figure 1 Let the first time point be T1 and the second time point be T2. We select the second time point as the current time point, and the first time point as the time within the target time period preceding the second time point. In this case, the target time period is 1 second, meaning the first time point is the second before the second time point. To obtain the second head pose detection result at the second time point, which is the head pose detection result at the current time point, to make the results more accurate, we need to utilize the probability diffusion model constructed in the previous second (i.e., the first probability diffusion model in step S100) and the latest image information, namely the second color image and / or the second depth image at the second time point.

[0041] In this embodiment, the human body is constantly moving, and its state is different from that of the previous second almost every second. Therefore, the target time period is preferably 1 second. When reconstructing the head 3D data during posture detection, the probability diffusion model of the previous second and the image acquired at the current time are combined to reconstruct the data. In this way, as time progresses, the head 3D data will be continuously improved, and the head posture estimation result obtained after being put into the head posture estimation algorithm will be more accurate.

[0042] In a specific embodiment, obtaining the probability diffusion model at a first time step and using it as the first probability diffusion model includes:

[0043] If the head pose estimation loop is in the initial state, the steps are as follows:

[0044] Step S110: Obtain a color image of the human head at the first moment and use it as the first color image;

[0045] Step S120: Detect the head pose of the human body based on the original probability diffusion model and the first color image to obtain the head pose detection result at the first moment, and use it as the first head pose detection result.

[0046] Step S130: Input the first head pose detection result into the head pose estimation algorithm to obtain the human head pose estimation result at the first moment, and output it;

[0047] The original probability diffusion model is used as the first probability diffusion model and applied in the second time step. The original probability diffusion model is the probability diffusion model set at the beginning of the method.

[0048] If the head pose estimation loop is in a non-initial state, then step S120 is: to detect the head pose of the human body based on the first color image and the second probability diffusion model obtained in the previous loop, to obtain the head pose detection result of the human body at the first moment, and to use it as the first head pose detection result.

[0049] In this embodiment, the probability diffusion model at the first moment is obtained based on the data at the first moment and before the first moment. The first color image and the first head pose detection result at the first moment are obtained. Based on this information, the probability diffusion model at the first moment can be obtained. At this time, the probability diffusion model can be learned and reconstructed based on the information at the first moment, forming the latest data information compared to before the first moment.

[0050] Furthermore, when obtaining the first probability diffusion model based on the first color image and the first head pose detection result at the first moment, it is optimized based on the probability diffusion model before the first moment, that is, the probability diffusion model corresponding to the previous second before the first moment.

[0051] When the loop is the initial loop, the first probability diffusion model is obtained based on the original probability diffusion model. When the loop is not the initial loop, the first probability diffusion model is obtained based on the second probability diffusion model obtained in the previous loop. Therefore, the probability diffusion model is constantly learning and iterating over time. However, when the probability diffusion model is used for detection for the first time, the original probability diffusion model used is a model that is manually added to the system.

[0052] In this embodiment, the probability diffusion model itself has the function of constructing the required data from noise, and can be used to improve and construct the head 3D data. In order to continuously update and optimize the probability diffusion model, this scheme uses the data at the current moment to optimize and generate the probability diffusion model at the previous moment, and after optimization, obtains the probability diffusion model at the current moment. Then, it continues to use the probability diffusion model at the current moment to obtain the detection results at future moments, thus ultimately achieving progressive improvement and iteration of the data.

[0053] In one specific embodiment, the method further includes:

[0054] The progressive head pose estimation method based on the probability diffusion model stops execution after responding to a stop command;

[0055] After the progressive head pose estimation method based on the probability diffusion model stops executing, the latest output of human pose head detection results and probability diffusion model are stored as storage data.

[0056] In this embodiment, the proposed method employs a progressive learning estimation approach based on the passage of time. When a stop instruction is given to the method at a certain point, it is necessary to control the storage of the latest output estimation result and probability diffusion model. The stored data is then used for the next execution of the progressive head pose estimation method.

[0057]

Example 2

[0058] In a specific embodiment, step S500: Optimize the first probability diffusion model based on the second color image and the second head pose detection result to obtain the probability diffusion model at the second time step, which serves as the second probability diffusion model, including:

[0059] The inverse Markov chain is obtained based on the first probability diffusion model;

[0060] The loss function for each link in the inverse Markov chain is constructed by maximizing the likelihood function and minimizing the KL divergence.

[0061] Parameters from the second color image and the second head pose detection results are extracted as optimization parameters;

[0062] The optimization parameters and random noise are input into the first probability diffusion model, and the gradient descent algorithm is used to train the first probability diffusion model according to the loss function to obtain the second probability diffusion model.

[0063] In this embodiment, when optimizing the first probability diffusion model, the probability diffusion model contains a forward Markov chain and a reverse Markov chain. First, the reverse Markov chain is obtained through the first probability diffusion model. Then, the loss function of each link in the reverse Markov chain is used to construct the model. Finally, the first probability diffusion model is optimized using the loss function and optimization parameters to obtain the second probability diffusion model.

[0064] In this embodiment, the inverse Markov chain is obtained according to the first probability diffusion model, including:

[0065] The process of transforming the hidden space random vectors in the first probability diffusion model into key points in the model space is referred to as the first process;

[0066] The first process is modeled to obtain the inverse Markov random process, and an inverse Markov chain is formed.

[0067] In the reverse Markov chain, the conditional probability and transition function of each link are determined based on the parameters in the first probability diffusion model.

[0068] In this embodiment, the process of obtaining the inverse Markov chain based on the first probability diffusion model specifically involves modeling the process of transforming the hidden space random vector into the model space key points to obtain the inverse Markov chain. In the process of obtaining the inverse Markov chain, the conditional probability and transition function of each link are determined according to the parameters in the first probability diffusion model. At this time, the inverse process is obtained on the basis of the first probability diffusion model. Therefore, the first probability diffusion model serves as the optimization basis, so that the originally optimized model no longer uses outdated parameters and data.

[0069]

Example 3

[0070] In one specific embodiment, see Figure 3 The head pose of the human body is detected based on the first probability diffusion model and the second color image to obtain the head pose detection result at the second time point, which is used as the second head pose detection result, including:

[0071] Determine whether the second color image includes human facial information;

[0072] If human facial information is included, the posture of the human face is detected to obtain the first detection result, and the first detection result is used as the second head posture detection result.

[0073] If facial information is not included, the posture of the human head is detected to obtain a second detection result, which is then used as the second head posture detection result.

[0074] In this embodiment, when performing pose detection on the entire head, the face is a crucial detection area. Compared to detecting the entire head, detecting only the face saves detection steps and data processing time, requiring less data and allowing for a more concise estimation result. Therefore, the step first determines whether the image information contains a human face. If it does, the face information is used for detection and serves as the second head pose detection result; otherwise, the head information is detected, and the detection result is used as the second head pose detection result.

[0075] In this embodiment, the detection is performed by first determining whether facial information is present, which saves detection steps, avoids obtaining unnecessary data, and improves detection speed and efficiency.

[0076] In a specific embodiment, if human facial information is included, the pose of the human face is detected to obtain a first detection result, and the first detection result is used as a second head pose detection result, including:

[0077] Hopenet is used to detect key points on the human face to obtain facial key point information;

[0078] The first detection result is obtained based on facial key point information;

[0079] The first detection result is output as the second head pose detection result.

[0080] In this embodiment, facial pose detection is particularly important in human head pose detection. This solution utilizes Hopenet to solve the problem of 2D-to-3D correspondence by identifying key points on the face. This method offers higher accuracy compared to traditional detection methods.

[0081] In one specific embodiment, if facial information is not included, the pose of the human head is detected to obtain a second detection result, and this second detection result is used as the second head pose detection result, including:

[0082] The top normal vector of the human head and the three-dimensional point data of the human head are obtained from the second color image, and the top normal vector is used as the initial posture information.

[0083] The head pose of the human body is detected using the first probability diffusion model to obtain point cloud voxel data of the human head.

[0084] Based on the preliminary attitude information, the projection error between the point cloud voxel data and the 3D point data is minimized to obtain the second detection result.

[0085] The second detection result is output as the second head pose detection result.

[0086] In this embodiment, when a face cannot be recognized, only the human head can be identified and detected. When detecting head pose, the top normal vector of the human head is relatively easy to obtain and can be temporarily used as preliminary pose information. After obtaining the preliminary pose information, in order to obtain more refined head pose information, it is necessary to continuously refine it using a first probability diffusion model to obtain more accurate pose detection results. Specifically, the refinement method involves minimizing the projection error between the point cloud voxel data and the 3D point data. Minimization improves data accuracy and yields more refined detection results.

[0087] In one specific embodiment, the head pose of the human body is detected according to a first probability diffusion model to obtain point cloud voxel data of the human head, including:

[0088] By combining a denoising diffusion model, Gaussian noise is removed from point cloud voxel data;

[0089] and / or

[0090] The first probability diffusion model is trained using the conditional likelihood function based on the L2 norm to obtain point cloud voxel data after noise removal.

[0091] In this embodiment, during the data reconstruction of human head pose using the probability diffusion model, two denoising methods can be selected to reduce noise interference. One method utilizes a denoising diffusion model, while the other uses a conditional likelihood function based on the L2 norm. Both methods ultimately yield more refined point cloud voxel data, further improving the accuracy of head pose estimation results.

[0092]

Example 4

[0093] In one specific embodiment, the progressive head pose estimation method based on the probability diffusion model utilizes a progressive head pose estimation system, which includes: an acquisition module for acquiring the probability diffusion model, a color image, and a depth image; a pose detection module for detecting the human head pose and obtaining the head pose detection result; and a head pose algorithm module for estimating the human head pose and obtaining the human head pose estimation result.

[0094] In this embodiment, the corresponding modules are used to complete the corresponding method steps in this case, so that the progressive head pose estimation method proposed in this case has the support and utilization of system modules.

[0095] In one specific embodiment, a progressive head pose estimation method based on a probability diffusion model also involves a progressive head pose estimation device, which includes a processor and a memory. The memory is used to store at least one program, which, when executed by the processor, causes the processor to implement the progressive head pose estimation method as described in the above embodiment.

[0096] In one specific embodiment, a computer-readable storage medium is also included, which stores a computer program that, when executed by a processor, implements the progressive head pose estimation method as described in the above embodiments.

[0097] Example 5

[0098] In this solution, a progressive head posture estimation method is applied to a table, specifically a yoga table. When a user is studying or working at the table, their posture needs to be collected and detected to correct their movement. Therefore, the method proposed in this study is used to collect and detect the user's posture. Furthermore, since the user is not a completely stationary object, detecting head posture while the user is moving is quite difficult. However, in this solution, through multiple data collection and optimization, the user's head posture data can be continuously improved and supplemented, making the estimation results more accurate and enhancing the yoga table's effectiveness in detecting and correcting user posture.

[0099] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A progressive head pose estimation method based on a probability diffusion model, characterized in that, The method uses a probability diffusion model to estimate the head pose of the human body, and the method includes: Establish a head pose estimation loop, which includes the following steps: Step S100: Obtain the probability diffusion model at the first time step, as the first probability diffusion model; Step S200: Obtain a color image of the human head at the second time point and use it as the second color image; Step S300: Detect the head pose of the human body according to the first probability diffusion model and the second color image to obtain the human head pose detection result at the second time point, and use it as the second head pose detection result. Step S400: Input the second head pose detection result into the head pose estimation algorithm to obtain the human head pose estimation result at the second time point, and output it; Step S500: Optimize the first probability diffusion model based on the second color image and the second head pose detection result to obtain the probability diffusion model at the second time point, which is used as the second probability diffusion model; After the current loop is completed, the second probability diffusion model under that loop is used as the first probability diffusion model for the next loop and input into the next loop. The second time is obtained by passing the first time point through the target time period; Step S500: Optimize the first probability diffusion model based on the second color image and the second head pose detection result to obtain the probability diffusion model at the second time point, which serves as the second probability diffusion model, including: The inverse Markov chain is obtained based on the first probability diffusion model; The loss function for each link in the inverse Markov chain is constructed by maximizing the likelihood function and minimizing the KL divergence. Parameters are extracted from the second color image and the second head pose detection result as optimization parameters; The optimization parameters and random noise are input into the first probability diffusion model, and the gradient descent algorithm is used to train the first probability diffusion model according to the loss function to obtain the second probability diffusion model.

2. The progressive head pose estimation method based on the probability diffusion model according to claim 1, characterized in that, If the head pose estimation loop is in its initial state, the steps are as follows: Step S110: Obtain a color image of the human head at the first moment and use it as the first color image; Step S120: Detect the head pose of the human body according to the original probability diffusion model and the first color image to obtain the head pose detection result of the human body at the first time moment, and use it as the first head pose detection result. Step S130: Input the first head pose detection result into the head pose estimation algorithm to obtain the human head pose estimation result at the first moment, and output it; The original probability diffusion model is used as the first probability diffusion model and applied in the first time step. The original probability diffusion model is the probability diffusion model set at the beginning of the method. If the head pose estimation loop is in a non-initial state, then step S120 is: to detect the head pose of the human body based on the first color image and the second probability diffusion model obtained in the previous loop, to obtain the human head pose detection result at the first moment, and to use it as the first head pose detection result.

3. The progressive head pose estimation method based on the probability diffusion model according to claim 2, characterized in that, The method further includes: The progressive head pose estimation method based on the probability diffusion model stops execution in response to a stop command; After the progressive head pose estimation method based on the probability diffusion model stops executing, the latest output of human pose head detection results and probability diffusion model are stored as storage data.

4. The progressive head pose estimation method based on the probability diffusion model according to claim 1, characterized in that, The process of obtaining the inverse Markov chain based on the first probability diffusion model includes: The process of transforming the hidden space random vectors in the first probability diffusion model into key points in the model space is taken as the first process; The first process is modeled to obtain an inverse Markov random process, and an inverse Markov chain is formed. The conditional probability and transition function of each link in the inverse Markov chain are determined based on the parameters in the first probability diffusion model.

5. The progressive head pose estimation method based on the probability diffusion model according to claim 3, characterized in that, The step of detecting the head pose of the human body based on the first probability diffusion model and the second color image to obtain the human head pose detection result at the second time moment, and using it as the second head pose detection result, includes: Determine whether the second color image includes human facial information; If the human face information is included, the posture of the human face is detected to obtain a first detection result, and the first detection result is used as the second head posture detection result. If the human face information is not included, the posture of the human head is detected to obtain a second detection result, and the second detection result is used as the second head posture detection result.

6. The progressive head pose estimation method based on the probability diffusion model according to claim 5, characterized in that, If the human face information is included, the posture of the human face is detected to obtain a first detection result, and the first detection result is used as the second head posture detection result, including: Hopenet was used to perform key point detection on the human face to obtain facial key point information; The first detection result is obtained based on the facial key point information; The first detection result is output as the second head pose detection result.

7. The progressive head pose estimation method based on the probability diffusion model according to claim 5, characterized in that, If the human face information is not included, the posture of the human head is detected to obtain a second detection result, and the second detection result is used as the second head posture detection result, including: The top normal vector of the human head and the three-dimensional point data of the human head are obtained from the second color image, and the top normal vector is used as the preliminary posture information. The pose of the human head is detected according to the first probability diffusion model to obtain the point cloud voxel data of the human head. Based on the preliminary attitude information, the projection error between the point cloud voxel data and the three-dimensional point data is minimized to obtain the second detection result; The second detection result is output as the second head pose detection result.

8. The progressive head pose estimation method based on the probability diffusion model according to claim 7, characterized in that, The step of detecting the pose of the human head based on the first probability diffusion model to obtain point cloud voxel data of the human head includes: By combining a denoising diffusion model, Gaussian noise is removed from the point cloud voxel data; and / or The first probability diffusion model is trained using a conditional likelihood function based on the L2 norm to obtain the point cloud voxel data after noise removal.

Images