CN102855470B - Estimation method of human posture based on depth image - Google Patents

Abstract

The invention discloses a human body pose estimation method. The method estimates the human body pose according to a depth image of the human body, so as to simulate the human body in the image through a virtual human body model. The depth image includes a plurality of frames. The method includes the following steps : establish a virtual human body model, the human body model is composed of a skeleton model and a skin model; initialize the parameters of the virtual human body model; filter the current frame of the depth image; correspond to the virtual human body model and the depth image point detection; for the current frame of the depth image, establish and optimize an objective function according to the result of the corresponding point detection, the objective function is used to describe the size of the posture difference between the virtual human body model and the depth image, The current pose of the virtual human body model is updated by minimizing the value of the objective function. The virtual human body model established by the invention has high degree of freedom, good skin deformation effect, fast convergence speed and small error of attitude estimation, and meanwhile, the depth image is obtained by using a depth camera, so that the device of the human body movement attitude estimation system is simple and convenient for popularization and application.

Description

Human Pose Estimation Method Based on Depth Image

technical field

The invention belongs to the fields of image processing, computer graphics, human kinematics, optimization theory and computer application, and in particular relates to a human body posture method based on a depth image.

Background technique

Human pose estimation is the core problem of human motion capture. The so-called human pose estimation refers to matching the abstract hierarchical features with the human body model, so as to obtain the poses of the target at different times. The posture expression of the human body includes two aspects, one is the position and direction of the whole human body in the world coordinates; the other is the angle of the joints of each part of the body and the skin deformation affected by the joint angles. The main application areas of human motion pose estimation can be divided into three directions: monitoring, control, and analysis.

In terms of monitoring applications, some traditional applications include automatic detection and positioning of pedestrians in airports or subways, people counting or crowd flow, congestion analysis, etc. With the improvement of security awareness, some new applications have emerged in recent years-analysis of the behavior and actions of individuals or groups of people. For example, in queuing and shopping, detecting abnormal behavior or performing identification.

In control applications, people use motion estimation results or attitude parameters to control the target. This is most widely used in human-computer interaction. In the entertainment industry, such as movies and game animations, the applications are becoming more and more extensive. People can use the captured people's shapes, appearances and movements to make 3D movies or reconstruct 3D models of people in games.

In terms of analysis applications, it includes automatic diagnosis of surgical patients, analysis and improvement of athletes' movements, etc. In terms of visual media, there are applications such as content-based video retrieval and video compression. In addition, related applications have also been obtained in the automotive industry, such as automatic control of airbags, sleep detection and pedestrian detection.

At present, the more mature human body motion capture systems on the market are based on electromechanical, electromagnetic and special optical signs. Magnetic or optical markers are attached to human limbs, and their three-dimensional trajectories are used to describe target motion. These systems are automatic, but the equipment is too bulky and expensive to be widely used. Therefore, human motion capture technology based on computer vision has become a research hotspot. It uses the basic principles of computer vision to directly extract 3D human motion sequences from videos. This method does not require any sensors attached to human joints, which ensures unrestricted human motion, and is low in cost and high in efficiency. Most of the current popular methods use matching techniques based on human body models. The goal of this method is to find a set of pose parameters in the state space, so that the pose of the human body corresponding to this parameter is most consistent with the underlying features extracted from the observed image.

In the field of motion tracking based on computer vision, the general research method is to determine the position of the human body in the first frame of the image sequence at the beginning of the tracking, and the determination of the human target in the subsequent sequence depends on the continuity and kinematic constraints of human motion. condition. There are two ways to determine the position of the human body in the first frame: one is to artificially specify the first pose of the target or set the human body model as the approximate pose of the first frame, which is not conducive to the automation of human body tracking. The second is to use the part detection method to determine each part of the body after removing the background other than the human body. This method can be partially automated, but requires strict guarantees for human scene segmentation. In the follow-up human tracking and 3D pose estimation, there are model-based and model-free methods. The general method based on the model is to establish a 3D model of the human body in advance, and match the model with the first frame of the motion sequence. Model parameters, resulting in a model motion sequence. The disadvantage of this method is that the tracking of subsequent frames has cumulative errors, and long-term tracking is prone to errors. The model-free method does not need to build a human body model, but uses learning or sample-based methods to estimate human motion posture based on the geometry, texture, color and other information presented by human motion. Its disadvantage is that the human motion posture is difficult to describe with a limited number of states, relies on prior knowledge, and can only track a specific set of actions. Both tracking methods can be implemented with monocular or multi-camera.

Due to the ambiguity of mapping from 3D to 2D in the reconstruction of ordinary images without depth information, and it is very difficult to estimate complex motion poses, in the past ten years of research, most human motion tracking technologies are Based on the realization of multi-camera conditions, depth information is obtained in this way. However, the condition of multi-camera needs to be calibrated and it is inconvenient to arrange in ordinary households, which is not conducive to the application of motion capture technology to thousands of households. In the past two years, with the emergence of depth cameras, people can use a single camera to obtain depth images, and the human pose estimation technology based on a single depth camera has become a research hotspot.

Contents of the invention

(1) Technical problems to be solved

The technical problem to be solved by the present invention is to propose a human body pose estimation method based on depth images, so as to solve the shortcomings of the existing human body pose estimation that requires multiple cameras, complicated equipment, and difficult to realize.

(2) Technical solution

In order to solve the above-mentioned technical problems, the present invention proposes a method for estimating human body posture. The method estimates the human body posture according to the depth image of the human body, so as to simulate the human body in the image through a virtual human body model. The depth image includes a plurality of frames, The method comprises the steps of:

S1. Establish a virtual human body model, which is composed of a skeleton model and a skin model;

S2. Initializing the parameters of the virtual human body model;

S3. Filter the current frame of the depth image;

S4. Perform corresponding point detection on the virtual human body model and the depth image;

S5. For the current frame of the depth image, establish and optimize an objective function according to the result of the corresponding point detection, the objective function is used to describe the size of the posture difference between the virtual human body model and the depth image, by Minimizing the value of the objective function updates the current pose of the virtual human model.

(3) Beneficial effects

The virtual human body model established by the present invention has high degree of freedom, good skin deformation effect, fast convergence speed and small error of attitude estimation, and at the same time, the depth image is obtained by using a depth camera, so that the device of the human body movement attitude estimation system is simple and convenient for popularization and application. The invention has broad application prospects in the field of human motion capture, and also provides application trends for human-computer interaction, game development, and film production.

Description of drawings

Fig. 1 shows the overall flow chart of each module of the present invention;

Fig. 2 shows the flowchart of the parameter initialization module of the present invention;

Fig. 3 shows the flowchart of the depth image filtering module of the present invention;

Fig. 4 shows the main flowchart of the layered pose estimation module of the present invention;

Fig. 5 shows the flowchart of translation and rotation parameter estimation in the layered attitude estimation module of the present invention;

Fig. 6 shows the flowchart of the joint angle parameter estimation in the layered pose estimation module of the present invention;

Fig. 7 shows the flowchart of the golden section method in the translation and rotation parameter estimation and the joint angle parameter estimation in the layered posture estimation module of the present invention to obtain the iterative step size;

Fig. 8 shows the human skeleton model of the present invention.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

The present invention is a method for estimating the posture of a human body using a depth image acquired by a depth camera, and presenting the result of the posture estimation as a virtual human body model. The depth image based on the present invention is obtained by a single depth camera. The depth image has depth information and can obtain three-dimensional coordinates in the scene, so it plays a very important role in motion estimation.

As shown in Figure 1, the method of the present invention carries out the estimation of human body posture according to the depth image of human body, to simulate the human body in this image by virtual human body model, and this depth image comprises a plurality of frames, and this method comprises the following steps:

1) Virtual human body model establishment step S1. The purpose of the step S1 of establishing the virtual human body model is to establish a non-rigid and deformable human body motion model for simulating the actual skin deformation of the human body and displaying the result of human body pose estimation.

2) Parameter initialization step S2. The parameter initialization step S2 includes reading random samples of three-dimensional coordinates of the virtual human body model, initializing posture parameters, obtaining coordinate transformation matrix and joint angle transformation matrix, etc.

3) Depth image filtering module step S3. The depth image filtering step S3 is a process of sampling, quantizing and smoothing and filtering a frame of the read-in depth image.

4) Corresponding point detection step S4. The corresponding point detection step S4 is used to find the corresponding relationship between the model point set and the depth image point set obtained by random sampling. The corresponding point means that each point in the model point set has a point of the depth image point set obtained by random sampling. correspond. In the present invention, a corresponding point detection method based on a multi-dimensional search tree is used.

5) Hierarchical pose estimation step S5. The layered pose estimation step S5 combines the method of layered estimation and the steepest descent method to iteratively optimize the pose parameters of each frame to obtain the estimated pose of the human body in the current frame.

6) Step S6 judges whether there is an image stream input, if there is still image data input, then go to the depth image filtering step S3, and continue to process the next frame of image to ensure the continuity of pose estimation; if not, the process ends.

Each main step is described in detail below.

S1. Establish a virtual human body model, the human body model is composed of a skeleton model and a skin model.

As shown in Figure 8, the skeleton model of the human body model is composed of 21 joints and 20 joint connections, of which 7 joints have three degrees of freedom, 6 joints have two degrees of freedom, 2 joints have one degree of freedom, and 6 joints have three degrees of freedom. A joint has no degrees of freedom. Therefore, the joint angle has a total of 35 degrees of freedom. Each joint has a unique ID. The joint ID in Figure 8, the degree of freedom of the joint, and the ID of the parent joint of the joint are shown in the following table:

In this table, for example, the second row represents the joint with ID 0, the joint degree of freedom is 3, and the parent joint ID is 0, that is, 0 is its own parent joint; the third row represents the joint with ID 1, and the joint degree of freedom is 3. The parent joint ID is 0, that is, joint 0 is the parent joint of joint 1, and so on.

In addition, the skeleton model has three degrees of freedom in translation and three degrees of freedom in rotation.

The skin model adopts a grid structure, and the nodes of the grid are skin nodes. The principle of non-rigid skin deformation is: the coordinate transformation of each skin node is not only affected by the joint closest to the skin node, but also affected by other joints, and the transformation of the skin node in the local coordinate system of each joint is determined by a certain The weights are added to get new coordinates.

S2. Initialize the parameters of the virtual human body model.

Parameter initialization is to prepare for pose estimation. As shown in Figure 2, the specific implementation process is as follows:

S21. Read the point set of the virtual human body model. The so-called point set refers to the three-dimensional coordinate information of the human body model, the normal vector information, the three-dimensional coordinate information of the skin node, and the weight distribution information of the bone node to the skin node. In one embodiment of the present invention, the point set of the human body model is obtained by reading a grid file with a suffix of .mesh, which contains the above-mentioned point set information. In this step, the three-dimensional coordinate information in the grid file is read in binary mode, and the coordinates are stored as a three-dimensional vector chain structure.

S22. Randomly sample the skin model of the human body model to obtain a model point set for pose estimation. After obtaining the 3D coordinate information of the skin model from step S21, since there are many skin node coordinates in the grid file, we do not need so many points for pose estimation, and in order to ensure the speed of the estimation algorithm, the size of the point set should be moderate . Therefore, in this step, the skin model is randomly sampled to obtain the model point set for pose estimation. In one embodiment, the number of sampling points is 500.

S23. Initialize the posture parameters of the virtual human body model. This step is to initialize the posture parameters of the human body model. The posture parameters refer to the parameters of the skeleton model and the parameters of the randomly sampled skin model, including the space transformation parameters of the human body model and the translation, rotation and joint angle of each skin node. parameter. We set the initial value of translation and rotation to 0, and the initial value of each joint angle is also 0.

S24. Obtain the joint transformation matrix of the skeleton model of the virtual human body model. The joint transformation matrix refers to the transformation matrix from the local coordinate system of the child joint itself to the local coordinate system of its parent joint. The transformation matrix from the local coordinate system of the child joint to the local coordinate system of the parent joint consists of a translation matrix and a rotation matrix. The transformation matrix from the local coordinate system of each joint to the world coordinate system can be obtained by concatenating multiple child-parent coordinate system transformation matrices.

According to an embodiment of the present invention, this step needs to read out the three-dimensional translation vector and the rotation quadruple from the .skeleton bone information file (.skeleton file is derived by the Maya model, contains the coordinates and transformation vectors of human skeleton joints) and converts them It is converted into a fourth-order matrix form, so that translation transformation and rotation transformation can be realized in the form of matrix multiplication.

Step S25, obtaining the joint angle transformation matrix of the skeleton model. The joint angle transformation matrix is the rotation matrix from the child joint to the parent joint with the joint angle as a variable. In this step, first define the rotation matrix from each child joint to the parent joint, and then calculate the derivative of each rotation matrix with respect to the joint angle, in preparation for calculating the gradient of the objective function in pose estimation.

S3. Filter the current frame of the depth image, the depth image is a depth image captured by a depth camera, and is an image that needs to be estimated for a human body pose. After obtaining the estimated posture of the present invention, the virtual human body model in the image can be simulated.

The depth image filtering step is to process the current frame of the depth image for pose estimation. As shown in Figure 3, the specific implementation process is:

S31. Acquire the current frame of the depth image. According to an embodiment of the present invention, the data source of the depth image is a dat file and a segmented human body contour image, and the dat file stores the data of the depth image. First read the gray value in the dat file in binary form to get a complete depth image. Then, the pixel-level AND operation is performed on the depth image and the human body contour image to obtain the depth information of the human body part. Specifically, the pixel value of the human body area in the human body contour image is 1, and the pixel value of the non-human body area is 0. After the depth image and the human body contour image are combined, the depth value of the human body part is obtained, and the non-human body part is 0.

S32. Randomly sample the depth information of the human body part in the depth image to obtain a point set of the depth image. What is obtained in step S31 is pixel points, but pose estimation does not require all the pixel points, so random sampling is required. The depth information of the human body part is randomly sampled to obtain the depth image point set. According to an embodiment of the present invention, the number of samples is 500.

S33. Perform smoothing processing on the depth image point set obtained after random sampling. Since the actual captured image may have some unreal depth information due to light emission and other reasons, we call it depth noise. Therefore, for the depth image point set obtained after random sampling, we use a template with a size of 5 pixels × 5 pixels to perform Gaussian smoothing filtering on it.

S34. Quantize the gray value of the depth image point set within the depth range of the human body model. The initial depth information is the gray value of 0-10000, for example. In order to express the depth information with the gray value, we need to quantize the gray value of the depth image point set to within a reasonable human body depth range. Specifically The range matches the depth range of the mannequin.

S4. Perform corresponding point detection on the human body model and the depth image.

Correspondence point detection is the key process of preprocessing. Mainly divided into two steps:

S41. Establish a multi-dimensional search tree for the depth image point set obtained in step S3. The multidimensional search tree is a binary tree structure, each node is a point of the depth image point set, and each layer divides the point set according to a dimension determined by the discriminator. The division rule is: for each layer, take the median of the selected dimension components of the layer as the node of the layer, the data smaller than the median is divided into the left subtree, and the data larger than the median is divided into It is the right subtree, so that each division ensures that the data volume of the left and right subtrees is almost equal. The discriminator used is: n mod k (n represents the nth layer of the tree, k represents the dimension).

S42. According to the multi-dimensional search tree, search for the corresponding relationship between the point set of the human body model and the point set of the depth image by using the principle of the nearest point search and the principle of minimizing the difference between the normal vectors. Given a point on the human body model, the corresponding point in the depth image must satisfy: 1) the ratio of the distance to the closest distance is less than a given threshold; 2) the square of the angle between the normal vectors of two points is less than a given threshold .

S5. For the current frame of the depth image, establish and optimize an objective function according to the result of the corresponding point detection, the objective function is used to describe the size of the posture difference between the virtual human body model and the depth image, by Minimizing the value of the objective function updates the current pose of the virtual human model.

The present invention uses a layered pose estimation method for pose estimation, and the hierarchical pose estimation is the core of the entire method of the present invention. The traditional pose estimation method is to estimate all parameters at the same time, which may lead to the problem that the objective function falls into a local minimum. To avoid this problem, we adopt a hierarchical estimation method to optimize the objective function, that is, in each iteration step, the translation and rotation parameters are estimated first, and then the joint angle parameters are estimated. As shown in Figure 4, the specific implementation process is:

S51. For each frame of the depth image, acquire a pose estimation result of the last frame of the virtual human body model. For the first frame of image, the initial pose parameters are obtained, and for the images after the second frame, the pose parameter estimation results of the previous frame are obtained.

S52. Establish an objective function for attitude estimation, and calculate an objective function value under the current attitude of the virtual human body model. This step records the initial value of the objective function at this iteration.

S53. For each frame of the depth image, estimate translation and rotation parameters of the virtual human body model. This step is to use the steepest descent method to estimate the translation and rotation parameters, then update the attitude parameters and the objective function, and then continue the subsequent parameter estimation. As shown in Figure 5, the specific implementation process is:

S531. Calculate translation gradients. First calculate the gradient of the mean square error to each translation component, and use the gradient formula based on Lorentz distribution modeling to calculate the gradient vector of the objective function to the translation parameter and normalize it.

S532. Calculate the rotation gradient. First calculate the gradient of the mean square error to each rotation component, and use the gradient formula based on Lorentz distribution modeling to calculate the gradient vector of the objective function to the rotation parameter and normalize it.

S533. Calculating the iteration step size by the golden section method. Take the negative gradient direction of the translation and rotation gradient obtained in step S531 and step S532 as the iteration direction, and then perform a one-dimensional search along the iteration direction with the current attitude parameter as the starting point to obtain the iteration step size. The steps of calculating the iterative step size by the golden section method will be described in detail later.

S534. Update translation and rotation parameters. Multiply the iteration step size calculated in the previous step and the iteration direction to obtain the increment of the translation and rotation parameters, and then add this increment to the original attitude parameters.

S535. Calculate the objective function value under the current attitude to update the objective function, and prepare for the next attitude parameter estimation.

S54. Estimating joint angle parameters of the virtual human body model for each frame of the depth image. After estimating the overall translation and rotation parameters of the model, the joint angle parameters are estimated in sequence from the trunk to the limbs. As shown in Figure 6, the specific implementation process is:

S541. Select the first joint angle. The first joint angle takes the root joint of the bone model, that is, the joint with ID 0.

S542. Calculate the gradient of the current joint angle. The gradient calculation of the joint angle needs to use the derivative of the joint rotation matrix to the joint angle component in step S25. Since the transformation from the local coordinate system of each joint to the world coordinate system requires the cascading of the coordinate transformation matrices of multiple child and parent joints, the calculation of the gradient of the joint angle can be achieved using a recursive algorithm to calculate the impact of the objective function on the joint The gradient of each degree of freedom component of the angle, so as to obtain the gradient vector union, and then normalize.

S543. Calculate the iteration step size by the golden section method. The negative direction of the joint angle gradient vector in step S542 is taken as the iteration direction, and one-dimensional search is performed along the iteration direction with the current attitude parameter as the starting point to obtain the iteration step size. The steps of calculating the iterative step size by the golden section method will be described in detail later.

S544. Update the current joint angle. Multiply the step length calculated in the previous step and the iteration direction to obtain the increment of the current joint angle parameter, and then add this increment to the original attitude parameter.

S545. Calculate the objective function value under the current posture. Update the objective function in preparation for the next joint angle parameter estimation.

S546. Determine whether all joint angles have been estimated. If there are still joint angles that have not been estimated, select the next joint angle, and continue to cycle until all joint angles are estimated; if all joint angles are estimated, end the joint angle parameter estimation step.

S547. Select the next joint angle. The principle of selecting joint angles is to select them in the order from the trunk to the limbs from the inside to the outside. This ensures the best effect of stratified estimation.

S548. Determine whether the difference of the objective function satisfies the error requirement. This step determines whether to continue the optimization iteration. If the difference between the latest objective function value and the initial objective function value of this iteration, that is, the objective function value recorded in step S52, is less than a given threshold, the posture parameters obtained in this iteration are used as the result, and the human body motion model is updated, End the hierarchical pose estimation step; if it is greater than a given threshold, go to step S52 and proceed to the next iteration.

S549. Output the virtual human body model. This step is to update the final model pose of the current frame and output it for display after the pose estimation of a frame of image is completed.

Calculating the iterative step size of the golden section method Step S533 and Step S543

The step of calculating the iterative step size by the golden section method is used in both translation and rotation parameter estimation step S53 and joint angle parameter estimation step S54. The golden section method is a one-dimensional search algorithm with fast convergence speed and high precision. Each search takes the golden section point as the interval breakpoint, and gradually shortens the search interval, so as to find the numerical approximate solution of the minimum point. As shown in Figure 7, the specific implementation process is:

S5331. Input the iteration direction. In the translation and rotation parameter estimation step S53, the iteration direction is the negative direction of the normalized translation and rotation gradient vector. In the joint angle parameter estimation step S54, the iteration direction is the negative direction of the normalized joint angle gradient vector.

S5332. Determine an initial search interval. The left boundary of the initial division interval is 0, and the right boundary is the maximum allowed iteration step size, that is to say, the size of each component of the vector obtained by multiplying the iteration step size by the iteration direction cannot exceed the allowed value range.

S5333. Take two segmentation points. This step is to determine the initial segmentation point. Assuming that the initial search interval is [a, b], the initial segmentation points are r1=a+0.382(b-a) and r2=a+0.618(b-a).

S5334. Determine whether the difference between the two segmentation points satisfies the accuracy. If the accuracy is satisfied, go to step S5335; if not, go to step S5336 to continue the iterative search. The difference between the two segmentation points is the length of the segmentation interval, and the precision is preset.

S5335. Take the midpoint of the two split points as the step size. After the one-dimensional search converges to the error range, take the midpoint of the final two segmentation points as the iterative step size, and end the golden section method to find the iterative step size module.

S5336. Calculate the objective function value with the two segmentation points as the step size, that is, calculate the values of E(r1) and E(r2), and prepare for the next step of judgment.

S5337. Update the search interval and the two segmentation points. Assuming that the current search interval is [a, b], the split points are r1 and r2, and the two split points calculated in step S5336 are the objective function values E(r1) and E(r2) of the step size, update the split interval and the two The specific implementation method of the split point is as follows:

S53371. Determine whether E(r1) is smaller than E(r2). If it is smaller, set b=r2, r2=r1, r1=a+0.382(b-a), and turn to the fourth step. If not less than, go to the second step.

S53372. Determine whether E(r1) is greater than E(r2). If so, set a=r1, r1=r2, r2=a+0.618(b-a), and turn to the fourth step. If not, go to step 3.

S53373. This is the case where E(r1) is equal to E(r2), let a=r1, b=r2, r1=a+0.382(b-a), r2=a+0.618(b-a), go to the fourth step.

S53374. After updating the search interval and segmentation points, go to step S5334 to continue the segmentation search and continuously shorten the search interval until it converges within the error range.

The specific embodiments described above have further described the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above descriptions are only specific embodiments of the present invention, and are not intended to limit the present invention. Within the spirit and principles of the present invention, any modifications, equivalent replacements, improvements, etc., shall be included in the protection scope of the present invention.

Claims (5)

1. an estimation method of human posture, the method carries out the estimation of human body attitude according to the depth image of human body, and to be simulated the human body in this image by virtual human model, this depth image comprises multiple frame, it is characterized in that, the method comprises the steps:

S1, set up virtual human model, this manikin is made up of skeleton model and skin model;

S2, initialization is carried out to the parameter of described virtual human model;

S3, filtering is carried out to the present frame of described depth image;

S4, corresponding point detection is carried out to described virtual human model and depth image;

S5, present frame for described depth image, set up and optimization object function according to the result that described corresponding point detect, this objective function, for describing the size of the posture difference between described virtual human model and described depth image, upgrades the current pose of described virtual human model by the value minimizing objective function; Wherein

Described step S2 comprises:

S21, read the point set of virtual human model, so-called point set refers to that the three-dimensional coordinate information of the three-dimensional coordinate information of this virtual human model, normal information, skin node and bone node are to the weight allocation information of skin node;

S22, the skin model of described virtual human model carried out to the model point set that stochastic sampling obtains for Attitude estimation;

The attitude parameter of virtual human model described in S23, initialization;

S24, obtain the joint transformation matrix of the skeleton model of described virtual human model;

Described step S3 comprises:

The present frame of S31, acquisition depth image;

S32, stochastic sampling is carried out to the depth information of the human body parts in described depth image obtain depth image point set;

S33, to the smoothing process of depth image point set obtained after described stochastic sampling;

S34, the gray-scale value of the point set of described depth image is quantized to described virtual human model depth range within;

Described step S4 comprises:

S41, described depth image point set set up to multi-dimensional search tree;

S42, to set according to described multi-dimensional search, adopt closest approach search and the minimum principle of normal vector difference to find the corresponding relation of described virtual human model point set and described depth image point set;

Described step S5 comprises:

S51, each frame for described depth image, obtain the previous frame Attitude estimation result of described virtual human model;

S52, foundation, for the objective function of Attitude estimation, calculate the target function value under the current pose of described virtual human model;

S53, each frame for depth image, estimate translation and the rotation parameter of virtual human model;

S54, each frame for depth image, estimate the joint angle parameter of virtual human model;

Further, described step S53 comprises further:

S531, calculate the translation gradient of described virtual human model, namely first calculate square error to the gradient of each translational component, utilize the gradient formula based on Lorentz distribution modeling to calculate objective function to the gradient vector of translation parameters and normalization;

S532, calculate the rotation gradient of described virtual human model, namely first calculate square error to the gradient of each rotational component, utilize the gradient formula based on Lorentz distribution modeling to calculate objective function to the gradient vector of rotation parameter and normalization;

S533, the negative gradient direction of getting translation that described step S531 and step S532 draws and rotating gradient as iteration direction, then with current pose parameter for starting point carries out linear search along iteration direction, draw iteration step length;

S534, the iteration step length calculated by described step S533 are multiplied with iteration direction and obtain the recruitment of translation parameters and rotation parameter, then on the basis of former attitude parameter, add this recruitment;

Target function value under S535, calculating current pose is to upgrade described objective function.

2. estimation method of human posture as claimed in claim 1, it is characterized in that, step S54 comprises:

S541, selection first joint angle;

The gradient at S542, calculating current joint angle;

S543, the negative direction of getting the joint angle gradient vector of step S542 are iteration direction, with current pose parameter for starting point carries out linear search along iteration direction, obtain iteration step length;

S544, renewal current joint angle, the step-length calculated by step S543 is multiplied with iteration direction and obtains the recruitment of current joint angular dimensions, then on the basis of former attitude parameter, adds this recruitment;

Target function value under S545, calculating current pose is to upgrade objective function;

S546, judge whether that all joint angles are all estimated complete, if also have joint angle not estimate, then select next joint angle, circulation is always gone down, until all joint angles are estimated complete; If all joint angles are estimated complete, then terminate joint angle parametric estimation step;

S547, select next joint angle;

S548, difference as target function value and initial target function value are less than given threshold value, then by attitude parameter as a result, and upgrade described virtual human model, otherwise forward step S52 to;

S549, export described virtual human model.

3. estimation method of human posture as claimed in claim 2, it is characterized in that, step S533 comprises:

S5331, input iteration direction, described iteration direction is normalized translation and the negative direction rotating gradient vector;

S5332, determine that initial ranging is interval, iteration step length the is multiplied by each component size of vector that iteration direction obtains can not exceed the span of permission;

S5333, determine initial cut-point;

S5334, judge whether the difference of two cut-points meets precision, if meet precision, then goes to step S5335; If do not met, then go to step S5336;

S5335, the mid point getting two cut-points are step-length, and linear search converges to after in error range, gets the mid point of final two cut-points as iteration step length;

S5336, the target function value that to calculate with two cut-points be step-length;

S5337, the renewal region of search and two cut-points.

4. estimation method of human posture as claimed in claim 2, it is characterized in that, step S543 comprises:

S5431, input iteration direction, the negative direction of described iteration direction to be iteration direction be normalized joint angle gradient vector;

S5432, determine that initial ranging is interval, iteration step length the is multiplied by each component size of vector that iteration direction obtains can not exceed the span of permission;

S5433, determine initial cut-point;

S5434, judge whether the difference of two cut-points meets precision, if meet precision, then goes to step S5435; If do not met, then go to step S5436;

S5435, the mid point getting two cut-points are step-length, and linear search converges to after in error range, gets the mid point of final two cut-points as iteration step length;

S5436, the target function value that to calculate with two cut-points be step-length;

S5437, the renewal region of search and two cut-points.

5. estimation method of human posture as claimed in claim 3, it is characterized in that, step S5337 is: suppose that current search interval is for [a, b], cut-point is r1, r2, two cut-points that known steps S5336 calculates are target function value E (r1) and the E (r2) of step-length, upgrade between cut section and comprise with the step of two cut-points:

S53371, judge whether E (r1) is less than E (r2), if be less than, then makes b=r2, r2=r1, r1=a+0.382 (b-a), forward step S53374 to; If be not less than, then forward step S53372 to;

S53372, judge whether E (r1) is greater than E (r2), if be greater than, then makes a=r1, r1=r2, r2=a+0.618 (b-a), forward step S53374 to; If be not more than, then forward step S53373 to;

S53373, when E (r1) equals E (r2), make a=r1, b=r2, r1=a+0.382 (b-a), r2=a+0.618 (b-a), forwards step S53374 to;

S53374, the renewal region of search and cut-point are complete, forward step S5334 to.