Human three-dimensional reconstruction method and system based on DQN algorithm and attention mechanism

Abstract

The present application relates to a kind of human three-dimensional reconstruction method and system based on DQN algorithm and attention mechanism.The reconstruction method first obtains the DQN network with multi-view selection function, and a group of human RGB image containing multiple view information is input as external environment;Then randomly select a view RGB image for initialization, extract the human two-dimensional joint information in the view RGB image, and obtain the three-dimensional human grid under the initial view.Q network in DQN network is used to determine the next view and obtain the three-dimensional human grid of the next view.Attention mechanism is used to fuse the three-dimensional human grid of current view and next view, and the fusion result is used to update the three-dimensional human grid of next view, and the above process is continuously iterated until the preset iteration number is met, and the final fusion estimated three-dimensional human grid is output.The reconstruction method and system have high accuracy, and more complete information is extracted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, in particular to a human three-dimensional reconstruction method and system based on a DQN algorithm and an attention mechanism. BACKGROUND

[0002] There is a serious depth ambiguity problem in three-dimensional human pose and shape estimation from a single image based on a fixed view, and multiple different three-dimensional poses can point to the same two-dimensional pose. In addition, high degrees of freedom of limbs can easily cause self-occlusion, and complex external environments can also easily cause human foreground occlusion. Therefore, the accuracy obtained by estimating 3D human pose and shape from a single image in these cases is far from enough.

[0003] Due to the limited depth information contained in a single view and the possible occlusion problem, using multiple view images can exploit more abundant depth information and reduce the impact of occlusion on human estimation. However, the effect of this way of reconstruction is heavily dependent on the quality of the selected pictures, and if only multiple fixed-view images are artificially selected, the accuracy of the reconstruction cannot be maximized.

[0004] Currently, some scholars have turned their attention to three-dimensional human pose reconstruction based on deep reinforcement learning. Compared with selecting fixed-view pictures, a deep reinforcement learning algorithm can adaptively select the most beneficial view for reconstruction according to the reconstruction effect. However, the current method of selecting a view using deep reinforcement learning only estimates the pose information of the human body, ignoring the shape information of the human body, while the shape of the human body is also an important aspect of reconstructing a complete human body. In addition, in the fusion of multi-view information, the fusion of all view information uses a relatively simple median fusion, which does not maximize the role of fusion. Therefore, the existing human three-dimensional reconstruction method has the problems of low accuracy and insufficiently rich and complete information. SUMMARY

[0005] Therefore, it is necessary to provide a human three-dimensional reconstruction method and system based on a DQN algorithm and an attention mechanism to solve the technical problems of low accuracy and insufficiently rich and complete information in the prior art.

[0006] The present application discloses a human three-dimensional reconstruction method based on a DQN algorithm and an attention mechanism, comprising the steps of:

[0007] S1. Obtain a DQN network with a multi-view selection function, and input a group of human RGB images containing multiple view information into the DQN network as the input of the external environment of the DQN network.

[0008] S2. Randomly select one of the RGB images of the view angle to initialize the DQN network, and extract the human two-dimensional joint information in the RGB image of the view angle. Among them, the human two-dimensional joint information includes the joint information of the body, hand and facial expression.

[0009] S3. Estimate the human two-dimensional joint information and the corresponding original RGB image information to obtain the three-dimensional human mesh in the initial view angle, which is the initial state of the environment in the DQN network.

[0010] S4. Determine the action to be taken at present by using the Q network in the DQN network, and determine the next view angle by executing the action and obtain the three-dimensional human mesh in the next view angle according to the mode of S2-S3.

[0011] S5. Use the attention mechanism to perform attention fusion on the three-dimensional human mesh in the current view angle and the three-dimensional human mesh in the next view angle, and update the three-dimensional human mesh in the next view angle by using the fusion result.

[0012] S6. Loop execution of steps S4-S5 for continuous iteration until the final fused estimated three-dimensional human mesh is output after meeting the preset iteration number.

[0013] As a further improvement of the above scheme, in S3, the SMPLify-X model is used to estimate the two-dimensional joint information and the corresponding original RGB image information to obtain the SMPL-X model of the three-dimensional human mesh. The expression formula of the SMPL-X model is:

[0014]

[0015] In the formula, θ represents the joint of the limb, face and hand, β represents the shape of the person, and ψ represents the facial expression. The three-dimensional human mesh is composed of a plurality of triangular facets, and the three-dimensional vertex is represented as

[0016] As a further improvement of the above scheme, in S4, after obtaining the initial state of the environment in the DQN network, the Q network predicts the Q value of each action according to the input state, and randomly selects an action with a preset probability ε. If the small probability event does not occur, the action with the maximum current Q value is selected as the action to be taken at present.

[0017] As a further improvement of the above scheme, the environment of the DQN network changes from the current state s to the next state s' after executing the action, and gives the corresponding reward. The expression formula of the reward is:

[0018]

[0019] In the formula, r represents the reward, which consists of the L2 loss between the predicted value of the 3D joint and the true value, and the L2 loss between the predicted value of the three-dimensional human mesh and the pseudo true value. j 3d represents the 3D joint obtained by regression from the estimated three-dimensional human mesh according to the pre-trained regression matrix, j 3d is the true value of the 3D joint. v is the vertex of the estimated three-dimensional human mesh, v' is the pseudo true value of the three-dimensional human mesh vertex. ω1 and ω2 are scaling factors, λ1 and λ2 are weight coefficients.

[0020] As a further improvement of the above scheme, after the environment performs the action and gives the reward, a track interacting with the environment is stored in the experience pool of the DQN network. The track is represented as (s, a, r, s'), and a is the action.

[0021] Wherein, when the experience pool stores a preset number of experiences, the Q network randomly extracts a batch of tracks from the experience pool.

[0022] As a further improvement of the above scheme, 200 memories are stored in the experience pool. The size of each memory is 31425*2+2.

[0023] As a further improvement of the above scheme, in S5, the optimal weight of fusion is determined by the attention mechanism, and the specific process is as follows:

[0024] (1) Preserve the geometric feature information of the three-dimensional human mesh, and at the same time, downsample the three-dimensional human mesh to encode the three-dimensional human mesh into And a plurality of downsampled three-dimensional human meshes are cascaded to obtain the encoded vector. Wherein p is the encoded vector.

[0025] (2) Define a Query matrix and a Key matrix, the Query matrix and the Key matrix are respectively obtained by point multiplication of the corresponding weight matrix W q and W k and the encoded vector p.

[0026] (3) Multiply the Query matrix and the Key matrix and pass through the softmax function to obtain the attention of each three-dimensional human mesh, that is, the weight coefficient allocated, and then allocate the weight coefficient to the initial unsampled three-dimensional human mesh to obtain the final fused three-dimensional human mesh.

[0027] As a further improvement of the above scheme, in the whole fusion process, the weight matrix W q and W k are optimized by the back propagation algorithm, and are supervised by the true value of the 3D joint.

[0028] The application further discloses a human three-dimensional reconstruction system based on the DQN algorithm and the attention mechanism, which applies any one of the human three-dimensional reconstruction methods based on the DQN algorithm and the attention mechanism.

[0029] The image acquisition module is used for acquiring a set of human RGB images containing multiple visual angle information.

[0030] The two-dimensional human posture extraction module is used for extracting human two-dimensional joint information in the RGB image of each visual angle.

[0031] The three-dimensional human posture and shape extraction module is used for obtaining a three-dimensional human mesh under an initial visual angle according to the estimation of the human two-dimensional joint information and corresponding original RGB image information.

[0032] The DQN network module comprises a Q network unit and an environment unit.

[0033] The attention mechanism module is used for attention fusion of the three-dimensional human mesh under the current visual angle and the three-dimensional human mesh under the next visual angle, and the three-dimensional human mesh under the next visual angle is updated by using the fusion result.

[0034] As a further improvement of the above scheme, the two-dimensional human posture extraction module extracts human two-dimensional joint information in the image by using openpose.

[0035] Compared with the prior art, the technical scheme disclosed by the application has the following beneficial effects:

[0036] 1. The human body three-dimensional reconstruction method based on the DQN algorithm and the attention mechanism solves the incompleteness of the body posture caused by the depth blur and the occlusion by designing a deep reinforcement learning DQN algorithm to enable an intelligent agent to adaptively select a plurality of view angles that are most beneficial to the reconstruction of the human body, and simultaneously introduces an attention mechanism for fusion, that is, different attention weight coefficients are assigned to the estimated three-dimensional human body grid, and the three-dimensional human body information under each view angle is effectively fused according to the weight coefficients, so that more accurate three-dimensional human body estimation is obtained. In addition, the reconstruction method of the present application not only estimates the posture and shape of the human body, but also estimates the posture and shape of the hand and face expression, so as to solve the problem of incomplete three-dimensional human body information estimation.

[0037] 2. The human body three-dimensional reconstruction system has the same beneficial effects as the reconstruction method, and will not be described here. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flow chart of the human body three-dimensional reconstruction method based on the DQN algorithm and the attention mechanism in embodiment 1 of the present application;

[0039] Figure 2 is a general framework diagram of the human body three-dimensional reconstruction method based on the DQN algorithm and the attention mechanism in embodiment 1 of the present application;

[0040] Figure 3 is an architecture schematic diagram of the deep reinforcement learning DQN network used in embodiment 1 of the present application;

[0041] Figure 4 is a process schematic diagram of iterative fusion in embodiment 1 of the present application;

[0042] Figure 5 is a qualitative result diagram on the panoptic dataset in embodiment 1 of the present application;

[0043] Figure 6 is a visualization of the fusion result and a visualization of the estimation of each view angle by the underlying pose estimator in the reinforcement learning intelligent agent view angle selection process in embodiment 1 of the present application;

[0044] Figure 7 is a qualitative result diagram on the CMU Panoptic dataset in embodiment 1 of the present application; in the diagram, from left to right, are the original image, VIBE, TCMR, Metro and the result of the method of the present embodiment. DETAILED DESCRIPTION

[0045] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing the specific embodiments only and is not intended to be limiting of the application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.

[0047] Embodiment 1

[0048] Please refer to Figure 1 and Figure 2 , the embodiment provides a human three-dimensional reconstruction method based on DQN algorithm and attention mechanism, comprising the steps of:

[0049] S1. Obtain a DQN network with multi-view selection function, and input a set of human RGB images containing multiple view information into the DQN network as the external environment of the DQN network. As shown in Figure 3 , the DQN network includes environment, Q network, target network, loss function and experience pool.

[0050] S2. Randomly select one of the RGB images of the view to initialize the DQN network, and extract the human two-dimensional joint information in the RGB image of the view. Among them, the human two-dimensional joint information includes the joint information of the body, the hand and the facial expression.

[0051] In step S2, in the DQN network with multi-view selection function, the input of the external environment is the RGB image containing multiple view information. First, an RGB image of a view is randomly initialized and selected. The 2D joint extractor openpose extracts the two-dimensional joint information in the image, including the information of the body, the hand and the facial expression.

[0052] S3. Estimate the human two-dimensional joint information and the corresponding original RGB image information to obtain a three-dimensional human mesh under the initialization view, and the three-dimensional human mesh is the initial state of the environment in the DQN network.

[0053] In this embodiment, the SMPLify-X model is used to estimate the two-dimensional joint information and the corresponding original RGB image information to obtain the SMPL-X model of the three-dimensional human body mesh. The expression formula of the SMPL-X model is:

[0054]

[0055] In the formula, θ represents the joints of the limbs, face, and hands, β represents the shape of the person, and ψ represents facial expression. The 3D human body mesh consists of several triangular patches, and its 3D vertices are represented as... This refers to a state s of the environment.

[0056] S4. Use the Q-network in the DQN network to determine the action to be taken, execute the action to determine the next viewpoint, and obtain the 3D human body mesh of the next viewpoint in the manner of S2 to S3.

[0057] After obtaining the state of the environment, the Q-network of DQN predicts the Q-value of each action based on the input state s, and randomly selects an action with a preset probability ε. If the low-probability event does not occur, a greedy strategy is used to select the action with the largest current Q-value as the action to be taken. After the environment executes action a, it immediately changes to the next state s' and receives a reward r from the environment. The formula for reward r is:

[0058] r=λ1ω1||j 3d -j 3d ′||2+λ2ω2||vv′||2

[0059] In the formula, the reward r consists of the L2 loss between the predicted and true values ​​of the 3D joints and the L2 loss between the predicted and pseudo-true values ​​of the 3D human body mesh. 3d j represents the 3D joints obtained from the estimated 3D human body mesh based on the pre-trained regression matrix. 3d v′ represents the true value of a 3D joint. v represents the estimated vertices of the 3D human body mesh, and v′ represents the pseudo-true value of a 3D human body mesh vertex. ω1 and ω2 are scaling factors, and λ1 and λ2 are weighting coefficients.

[0060] In this way, a trajectory (s, a, r, s') interacting with the environment is stored in the experience pool. After the experience pool has stored a certain amount of experience, the Q network randomly selects a batch of n trajectories from the experience pool. The purpose of random selection is to break the correlation between samples. In addition, the target network has the same network structure as the Q network. The loss function between the Q network and the target network is used to update the parameters of the Q network. During the update process, the parameters of the target network remain unchanged. Every N steps, the parameters of the Q network are synchronously copied to the target network.

[0061] The experience pool of the DQN stores 200 memories, each memory has a size of 31425*2+2, stores the current state state, the next state next state, the action action and the reward reward, 16 memories are randomly selected from the 200 memories as a batch for training, Adam is used as the optimizer to update the model parameters, and the learning rate is set to 0.01.

[0062] S5. The three-dimensional human body grid of the current perspective is fused with the three-dimensional human body grid of the next perspective by using the attention mechanism, and the three-dimensional human body grid of the next perspective is updated by using the fusion result.

[0063] In the application, in order to capture the detailed features under different perspectives, the three-dimensional human body grid reconstructed under the selected perspective is also fused. Fusing the three-dimensional human body grids under different perspectives requires assigning weight coefficients to them, and the fusion methods mentioned earlier all choose simple median fusion. Different weight coefficients have a certain influence on the final fusion result, so an attention mechanism is used to find the optimal weight for fusion, and the specific process is as follows:

[0064] First, in order to reduce the amount of calculation, the three-dimensional human body grid is down-sampled while retaining its geometric feature information, and the three-dimensional human body grid is encoded as And a plurality of down-sampled three-dimensional human body grids are concatenated to obtain the encoded vector. Wherein p is the encoded vector.

[0065] Then a Query matrix and a Key matrix are defined, the Query matrix and the Key matrix are obtained by point multiplication of the corresponding weight matrix W q and W k and the encoded vector p. Secondly, we multiply the Query matrix and the Key matrix and pass them through a softmax function to obtain the attention score of each three-dimensional human body grid, that is, the weight coefficient assigned to the initial three-dimensional human body grid without down-sampling to obtain the final fused three-dimensional human body grid.

[0066] In the whole process, the matrices W q and W k are trained by backpropagation algorithm to optimize, and are supervised by the true value of the 3D joint point.

[0067] S6. The steps S4-S5 are repeatedly executed for continuous iteration until the final fused estimated three-dimensional human body grid is output after meeting the preset number of iterations.

[0068] Please refer to Figure 4In the embodiment, according to the attention mechanism, the three-dimensional human body grid is fused, and the state obtained in each round of iteration update process is obtained by attention fusion of the three-dimensional human body grid in the current round and the state in the last round, except the initial state. After continuous iteration, the three-dimensional human body is output after the agent reaches the threshold condition set in advance. The present application solves the problem that the existing method is difficult to actively select the view angle according to multiple images to adaptively fuse and estimate the three-dimensional human body posture and shape model with more accurate.

[0069] The embodiment also provides experiments to verify the effect of the three-dimensional human body reconstruction method based on the DQN algorithm and the attention mechanism of the present application.

[0070] (1) Dataset

[0071] The CMU Panoptic Studio dataset with a large multi-view system is selected in the embodiment, which contains a large number of single-person and multi-person scenes, has 480 VGA videos, 31 high-definition videos, camera calibration data and the like. The panoptic dataset provides data at a speed of 30 frames per second, uses high-definition videos with 30 view angles and extracts them into picture form as the input of the model.

[0072] The data used is the single-person Pose scene in the panoptic dataset, which contains various ranges of motions of single persons. The scene is randomly divided into a training set and a test set, the model learns how to select a suitable view angle to maximize the overall reward on the training set, and the test set is used for final evaluation.

[0073] (2) Experimental details

[0074] SMPLlify-X is used as the underlying human body posture and shape estimator, and the agent continuously selects the next view angle for optimization through the feedback of the underlying estimator. The true value of the three-dimensional human body joint is used as supervision. The experience pool of the DQN stores 200 memories, the size of which is 31425*2+2, stores the current state state, the next state next_state, the action action and the reward reward, and 16 of the 200 memories are randomly selected as a batch for training, Adam is used as the optimizer to update the model parameters, and the learning rate is set to 0.01.

[0075] (3) Experimental results

[0076] The model of the embodiment adaptively selects the view angle for three-dimensional human body posture and shape reconstruction through the deep reinforcement learning algorithm, including the body, the hand and the facial expression. The visualization result is as shown in Figure 5 To more comprehensively show the reconstruction effect, Figure 5RGB pictures and the corresponding reconstruction results are given from four different views. It can be seen that in the case of certain occlusion

[0077] , our model can still reconstruct a relatively accurate 3D human body by choosing the view that is beneficial to the reconstruction and supplementing information.

[0078] Please refer to Figure 6 , Figure 6 to show the process of view selection and fusion of the agent and the effect of the underlying estimator on the reconstruction of a single view. It can be seen that in the estimation of a single view without fusion, due to the existence of occlusion, the leg and hand pose estimation is not accurate enough, and the human body shape also has certain differences. But after our model selects the view and fuses the estimated human body model, the leg pose gradually changes and gradually gets more accurate results.

[0079] (4) Qualitative experiments

[0080] Please refer to Figure 7 , Figure 7 to show the results of VIBE, TCMR, Metro and the method of the embodiment on the CMU Panoptic dataset. Metro is trained based on single-view pictures, but VIBE and TCMR are trained based on videos. We can see from Figure 7 that for human bodies with different poses and shapes, VIBE has the phenomenon of body interpenetration; the leg and foot estimation of TCMR is not accurate enough; due to single view, Metro estimates the invisible body part not well enough. Therefore, the results show that compared with the other three methods, the method of the embodiment can estimate and recover the 3D human body in better pose and shape in terms of details, in addition, the method of the embodiment also optimizes the fusion of facial expression and hand pose, so the hand and face pose of the 3D human body model obtained is no longer in the form of average pose, and looks more complete and realistic.

[0081] In summary, the reconstruction method uses the DQN algorithm to adaptively select the view angle provided by the agent to the underlying three-dimensional pose and shape estimator, and encourages the agent to select the appropriate view angle by giving the correct reward. Meanwhile, we also introduce an attention fusion module to effectively fuse the three-dimensional human mesh information under different view angles. Existing three-dimensional human estimation methods based on deep reinforcement learning are trained in the policy gradient framework, and only the mean value of the human 3D joint is fused, without training and fusing the three-dimensional shape information of the human body, and without paying attention to the detailed information of the hand and face expressions of the human body. The hand and face expressions of the obtained model are in the average posture, and the information-rich complete human model information is not constructed. Compared with the prior art, the present application takes the three-dimensional face vertex of the three-dimensional human mesh as the state of the deep reinforcement learning algorithm environment, because the three-dimensional human mesh not only contains the position information of the human joints, but also contains the shape information of the human body, in addition, the hand and face expressions of the human body are trained by attention fusion, so as to obtain a three-dimensional human model with more complete and rich information. From the results, our method not only has a good estimation of the human posture, but also has a more accurate estimation of the shape of the human body and the details of the hand and face expressions.

[0082] Embodiment 2

[0083] The embodiment provides a human three-dimensional reconstruction system based on a DQN algorithm and an attention mechanism, which applies the human three-dimensional reconstruction method based on the DQN algorithm and the attention mechanism in embodiment 1. The human three-dimensional reconstruction system comprises an image acquisition module, a two-dimensional human pose extraction module, a three-dimensional human pose and shape extraction module, a DQN network module, and an attention mechanism module.

[0084] The image acquisition module is used to acquire a set of human RGB images containing multiple view angle information.

[0085] The two-dimensional human pose extraction module is used to extract human two-dimensional joint information in the RGB image of each view angle. The human two-dimensional joint information includes joint information of the body, hand and face expressions.

[0086] The three-dimensional human pose and shape extraction module is used to estimate the three-dimensional human mesh under the initial view angle according to the human two-dimensional joint information and the corresponding original RGB image information.

[0087] The DQN network module comprises a Q network unit and an environment unit. The Q network unit is used to determine the action to be taken. The environment changes from the current state to the next state by performing the action, and gives the corresponding reward, so as to realize that the DQN network module determines the next view angle and obtains the three-dimensional human mesh of the next view angle.

[0088] The attention mechanism module is configured to perform attention fusion between the three-dimensional human body mesh of the current view and the three-dimensional human body mesh of the next view, and update the three-dimensional human body mesh of the next view by using a fusion result. The three-dimensional human body mesh of the final fusion estimation is output after a preset number of iterations.

[0089] The technical features of the above-described embodiments can be combined in any manner. To make the description concise, all possible combinations of the technical features in the above-described embodiments are not described, but as long as the combinations of the technical features do not exist in contradiction, they should be considered as falling within the scope of the present disclosure.

[0090] The above-described embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but should not be understood as a limitation on the scope of the present application. It should be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for human three-dimensional reconstruction based on a DQN algorithm and an attention mechanism, characterized in that, The method comprises the steps of: S1. obtaining a DQN network with multi-view selection function, and inputting a set of human body RGB images containing multiple view information into the DQN network as an external environment of the DQN network; S2. initializing the DQN network by randomly selecting an RGB image of one view, and extracting human body two-dimensional joint information in the RGB image of the view; wherein the human body two-dimensional joint information comprises joint information of body, hand and facial expression; S3. estimating the human body two-dimensional joint information and the corresponding original RGB image information to obtain a three-dimensional human body mesh in the initial view, which is an initial state of the environment in the DQN network; S4. determining an action to be taken in the current DQN network by using a Q network in the DQN network, and determining a next view by executing the action, and obtaining a three-dimensional human body mesh in the next view according to the method of S2-S3; S5. performing attention fusion on the three-dimensional human body mesh in the current view and the three-dimensional human body mesh in the next view by using an attention mechanism, and updating the three-dimensional human body mesh in the next view by using the fusion result; S6. performing continuous iteration by cyclically executing S4-S5 until the final fused and estimated three-dimensional human body mesh is output after a preset number of iterations is met.

2. The method of claim 1, wherein the DQN algorithm and attention mechanism-based human three-dimensional reconstruction method is characterized in that, In S3, the SMPLify-X model is used to estimate the two-dimensional joint information and the corresponding original RGB image information to obtain an SMPL-X model of the three-dimensional human body mesh; and an expression formula of the SMPL-X model is as follows: where θ represents the joints of the limbs, face and hands, β represents the shape of the person, and ψ represents the facial expression; the three-dimensional human mesh is composed of a plurality of triangular facets, and the three-dimensional vertices are represented as 3. The method of claim 1, wherein the DQN algorithm and attention mechanism-based human three-dimensional reconstruction method is characterized in that, In S4, after obtaining the initial state of the environment in the DQN network, the Q network predicts the Q value of each action according to the input state, and randomly selects an action with a preset probability ε; if the small probability event does not occur, the action with the maximum current Q value is selected as the action to be taken by using a greedy strategy.

4. The method of claim 3, wherein the DQN algorithm and attention mechanism-based human three-dimensional reconstruction method is characterized in that, After the action is executed, the environment in the DQN network changes from the current state s to the next state s', and a corresponding reward is given; and an expression formula of the reward is as follows: r = λ1ω1||j 3d -j 3d ||2+ λ2ω2||v-v' ||2 In the formula, r represents the reward, which is composed of the L2 loss between the predicted value and the true value of the 3D joint and the L2 loss between the predicted value and the pseudo true value of the three-dimensional human body mesh; j 3d represents the 3D joint obtained by regression from the estimated three-dimensional human body mesh according to the pre-trained regression matrix, j 3d is the true value of the 3D joint; v is the vertex of the estimated three-dimensional human body mesh, v′ is the pseudo true value of the three-dimensional human body mesh vertex; ω1 and ω2 are scaling factors, and λ1 and λ2 are weight coefficients.

5. The method of claim 4, wherein the DQN algorithm and attention mechanism-based human three-dimensional reconstruction method is characterized in that, After the action is executed and the reward is given, a trajectory interacting with the environment is stored in an experience pool of the DQN network; and the trajectory is represented as (s, a, r, s'), wherein a is the action. When the experience pool stores a preset number of experiences, the Q network randomly extracts a batch of trajectories from the experience pool.

6. The method of claim 5, wherein the DQN algorithm and attention mechanism-based human three-dimensional reconstruction method is characterized in that, The experience pool stores 200 memories; and the size of each memory is 31425*2+2.

7. The method of claim 4, wherein the DQN algorithm and attention mechanism-based human three-dimensional reconstruction method is characterized in that, In S5, the optimal weight of fusion is determined by using an attention mechanism, and the specific process is as follows: (1) retaining the geometric feature information of the three-dimensional human body mesh, while down-sampling the three-dimensional human body mesh, encoding the three-dimensional human body mesh into and concatenating a plurality of down-sampled three-dimensional human body meshes to obtain an encoded vector; wherein p is the encoded vector. (2) define a Query matrix and Key matrix, said Query matrix and Key matrix are respectively corresponding to the weight matrix W q and W k and the encoding vector p dot product is obtained; (3) multiplying the Query matrix and the Key matrix and obtaining the attention of each three-dimensional human body mesh by using a softmax function, i.e., the weight coefficient of the assignment, and then assigning the weight coefficient to the initial three-dimensional human body mesh without downsampling to obtain the final fused three-dimensional human body mesh.

8. The method of claim 7, wherein the DQN algorithm and attention mechanism-based human three-dimensional reconstruction method is characterized in that, The weight matrix W is trained by backpropagation algorithm throughout the fusion process q and W k are optimized and supervised by the ground truth of 3D joint points.

9. A human three-dimensional reconstruction system based on a DQN algorithm and an attention mechanism, characterized in that, The application of the human three-dimensional reconstruction method based on the DQN algorithm and the attention mechanism as claimed in any one of claims 1 to 8; The human three-dimensional reconstruction system comprises: An image acquisition module for acquiring a set of human RGB images containing multiple view information; A two-dimensional human pose extraction module for extracting human two-dimensional joint information in the RGB image of each view; the human two-dimensional joint information includes joint information of body, hand and facial expression; A three-dimensional human pose and shape extraction module for obtaining a three-dimensional human mesh in the initial view according to the estimation of the human two-dimensional joint information and the corresponding original RGB image information; A DQN network module including a Q network unit and an environment unit; the Q network unit is used to determine the action to be taken at present; the environment changes from the current state to the next state by performing the action and gives the corresponding reward, thereby realizing the DQN network module to determine the next view and obtain the three-dimensional human mesh in the next view; and An attention mechanism module for attention fusion of the three-dimensional human mesh in the current view and the three-dimensional human mesh in the next view, and updating the three-dimensional human mesh in the next view by using the fusion result; wherein the final fused estimated three-dimensional human mesh is output after a preset number of iterations.

10. The human body three-dimensional reconstruction system based on the DQN algorithm and the attention mechanism according to claim 9, characterized in that, The two-dimensional human pose extraction module extracts human two-dimensional joint information in the image by using openpose.

Images