Method for measuring the three-dimensional profile of a moving object using structured light

Abstract

The application discloses a kind of structured light measurement methods of three-dimensional profile of moving object, including constructing a convolutional neural network, generating training data and using convolutional neural network to train to obtain neural network model, by the phase shift diagram of training to the three-dimensional reconstruction of the moving object to be measured etc. The measurement method of the present application combines neural network and structured light three-dimensional projection, eliminates motion error by neural network training, simplifies the iterative calculation of complex motion error, can realize the measurement of three-dimensional profile in real time when object is in uniform motion, variable speed motion, rotation and other various states, quickly and efficiently obtain accurate three-dimensional information of moving object;The measurement method of the present application uses each image shot, collects the information of each motion position and its before and after picture, and the information utilization rate is high;The three-dimensional profile of moving object restored by the measurement method of the present application is more accurate, more clear in detail, and high-frequency information is more completely retained.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of structured light three-dimensional measurement, and particularly relates to a structured light measurement method for three-dimensional profile of a moving object. BACKGROUND

[0002] In recent years, with the continuous progress of measurement technology and the continuous development of industrialization, the traditional two-dimensional measurement technology cannot meet the needs of people, and more and more fields begin to have precise and rapid measurement needs for three-dimensional information of objects. For example, the fields of cultural relic digitization, medical diagnosis, machine vision, workpiece online monitoring, natural disaster investigation, and map reconstruction. The three-dimensional measurement technology acquires the point coordinates of the measured object in the three-dimensional space by accurately scanning the shape structure of the measured object, and then obtains the three-dimensional model of the measured object. Compared with two-dimensional measurement, three-dimensional measurement has the advantages of more intuitiveness and perfection, and has the advantages of high precision, no contact with the surface of the object, fast scanning speed, etc., and shows great application value. Nowadays, it is increasingly valued by people.

[0003] Optical three-dimensional measurement mainly includes stereo vision measurement, laser scanning method, time flight method, and structured light projection measurement method. Among them, the structured light projection measurement method has always been a research hotspot in three-dimensional measurement due to its fast measurement speed, high precision, easy control, and low cost. The structured light projection measurement method mainly includes phase shift method and Fourier transform method. The measurement precision of the phase shift method is higher, but it needs to project and shoot multiple images to realize three-dimensional reconstruction, which is time-consuming. The Fourier transform method only needs one image, but the measurement precision of the Fourier transform method is not high and the calculation is complex. With the increase of the frame rate of the projector and the camera, the requirement for measurement precision is increasing, and the phase shift method has gradually become the main research direction of structured light three-dimensional measurement.

[0004] Although the current structured light three-dimensional measurement method can achieve high-speed and high-precision measurement, it can only measure static objects or objects moving very slowly. When three-dimensional measurement is performed on a high-speed moving object, a huge error will be generated. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a structured light measurement method for three-dimensional profile of a moving object, which combines a neural network and structured light three-dimensional projection, eliminates motion errors through neural network training, and realizes three-dimensional profile measurement of a high-speed moving object.

[0006] The technical solution adopted by the present application to solve the above technical problem is as follows: a structured light measurement method for three-dimensional profile of a moving object, comprising the following steps:

[0007] Step 1: constructing a convolutional neural network;

[0008] Step 2: generate training data and train a neural network model using a convolutional neural network

[0009] Step 2-1, first build an experimental device, the experimental device includes a computer controlled translation stepper motor, a rotary stepper motor, a projector and a camera, the rotary stepper motor is arranged above the translation stepper motor, the translation stepper motor is placed on a horizontal table top, the translation stepper motor is used to control the rotary stepper motor to move in front and back, up and down, left and right three directions, the output end of the rotary stepper motor is provided with a load platform, the camera and the projector are respectively arranged above the load platform, the upper surface of the load platform in the initial position is taken as a reference plane, the connecting line of the optical center of the camera and the optical center of the projector is parallel to the reference plane, the optical axis of the camera and the optical axis of the projector are in the same plane and the plane is perpendicular to the reference plane;

[0010] Step 2-2: place the object on the load platform, project N stripes on the object using the projector, N≥3, the expression of the stripe is:

[0011]

[0012] Where f is the stripe frequency, x and y are the horizontal and vertical coordinates of the pixel points in the camera respectively;

[0013] When the picture is projected onto the object, the height of the object will cause the stripe to deform, and the height information is converted into phase information. The camera captures N stripes with object phase information, and the expression is:

[0014]

[0015] Where A(x,y) is the background noise, B(x,y) is the reflection noise, is the phase value converted from the height of the object;

[0016] When the object moves in a certain direction, the coordinates of the pixel points change from (x,y) to (x+ε x ,y+ε y ), where ε x ,ε y are the motion noises in the horizontal and vertical directions respectively. Since the background noise does not change with the movement of the object, the reflection noise changes with the movement of the object. The expression of the stripe of the moving object is:

[0017]

[0018] The computer controls the initial position and movement of the translational and rotary stepper motors to ensure that the object can move multiple times along the same trajectory; the projector projects N stripes and triggers the camera to capture images synchronously; the computer controls the projection to synchronize with the object's movement to ensure that the object's position in the multiple images captured by the camera during each movement is consistent with the object's position during the first capture.

[0019] Let the object move along the same trajectory M times. Take N images as a group for each movement, where M = N. The phase shift difference between each group of images is... The image captured is expressed as:

[0020]

[0021] Take N images with m = N / 2 and round down as data group 1. The objects in the images in data group 1 are in the same position and at the midpoint of the entire motion process. Take N images with n = 0 as data group 2. The phase of the objects in the images in data group 2 is the same as that in data group 1, and the positions of the objects change sequentially.

[0022] Steps 2-3: Repeat step 2-2, projecting different objects along different trajectories to obtain multiple data sets 1 and 2. Use multiple data sets 1 as the output of the convolutional neural network and multiple data sets 2 as the input to the convolutional neural network to obtain training data. Train the convolutional neural network and use the root mean square error as the loss function to calculate the difference between the input and output data to evaluate the quality of the training results and determine whether the training is complete. After training is completed, the neural network model is obtained.

[0023] Step 3: Perform 3D reconstruction of the moving object under test from the trained phase-shifted image.

[0024] The object to be tested is placed on a platform, and a computer is used to control the translational stepper motor and the rotary stepper motor to make the object to be tested move with the platform. t images are captured, where t>N. The kth to k+N-1th images are input into a neural network model to obtain an N-step phase-shifted image at the kth position without motion error.

[0025] The enclosed phase is calculated using the least squares method with an N-step phase shift diagram, and its expression is as follows:

[0026]

[0027] Unwrap the wrapped phase, remove the 2πfx term, and obtain the phase value caused by the object's height without motion error. By combining this with trigonometric methods, the height h(x,y) and phase value of the object under test can be obtained. Correspondence:

[0028]

[0029] wherein d is the distance between the optical center of the camera and the optical center of the projector, and l is the height from the optical center of the camera to the reference plane;

[0030] After the three-dimensional profile of the to-be-measured moving object at the kth position is restored, k is set to 1, 2,..., t-N+1, the three-dimensional profiles of the to-be-measured moving object at the continuous k different positions are obtained, and the three-dimensional profile measurement of the moving object is realized.

[0031] The measurement method of the present application trains the convolutional neural network by taking the moving items in the same phase of the plurality of groups of stripe pictures as the input of the convolutional neural network and taking the same position items as the output of the convolutional neural network. The stripe pictures of the moving object taken by the camera can obtain the static stripe pictures at the same position through the neural network, realize the elimination of the motion error, and obtain the three-dimensional profile of the moving object without the motion error by unwrapping the phase of the stripe pictures. Further, the three-dimensional video of a continuous moving object can be obtained from the continuous three-dimensional profile of the moving object.

[0032] As preferred, the convolutional neural network constructed in step 1 comprises a feature extraction network and a feature fusion network, the feature extraction network comprises a first feature extraction module, a second feature extraction module, a third feature extraction module, a fourth feature extraction module and a fifth feature extraction module connected in sequence, the first feature extraction module, the second feature extraction module, the third feature extraction module and the fourth feature extraction module each comprise a first convolutional layer, a first linear rectifier layer, a second convolutional layer, a second linear rectifier layer and a down-sampling layer connected in sequence, the fifth feature extraction module comprises a third convolutional layer, a third linear rectifier layer, a fourth convolutional layer and a fourth linear rectifier layer connected in sequence; the feature fusion network comprises a first feature fusion module, a second feature fusion module, a third feature fusion module, a fourth feature fusion module and an image output module connected in sequence, the first feature fusion module, the second feature fusion module, the third feature fusion module and the fourth feature fusion module each comprise an up-sampling layer, a fifth convolutional layer, a fifth linear rectifier layer, a sixth convolutional layer and a sixth linear rectifier layer connected in sequence, and the image output module is a seventh convolutional layer; the image obtained after the second linear rectifier layer in the first feature extraction module is superimposed with the image after the up-sampling layer in the fourth feature fusion module, the image obtained after the second linear rectifier layer in the second feature extraction module is superimposed with the image after the up-sampling layer in the third feature fusion module, the image obtained after the second linear rectifier layer in the third feature extraction module is superimposed with the image after the up-sampling layer in the second feature fusion module, and the image obtained after the second linear rectifier layer in the fourth feature extraction module is superimposed with the image after the up-sampling layer in the first feature fusion module.

[0033] As preferred, the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer and the sixth convolutional layer each adopt a convolution kernel with a size of 3*3 and a padding value of 2, the seventh convolutional layer adopts a convolution kernel with a size of 1*1 and a padding value of 0, the down-sampling layer adopts a maximum pooling layer with a kernel size of 2*2, and the up-sampling layer has a kernel size of 2*2. The first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer and the sixth convolutional layer each adopt a convolution kernel with a size of 3*3 and a padding value of 2, so that the picture size before and after the convolutional layers is consistent. The linear rectifier layer is used to avoid the problem of negative image values and constant stacking of exponential explosion. The down-sampling layer is used to reduce the picture resolution to obtain more abundant image features, and the down-sampling layer adopts a maximum pooling layer with a kernel size of 2*2, so that the image height and width are halved. The up-sampling layer is used to restore the image to the initial definition, so as to fuse the features of the image with the initial image.

[0034] Compared with the prior art, the present application has the following advantages:

[0035] (1) The measuring method of the present application combines neural network and structured light three-dimensional projection, eliminates motion error through neural network training, simplifies the complicated iterative calculation of motion error, and can realize real-time motion three-dimensional profile measurement of objects in uniform motion, variable speed motion, rotation motion and other states, and quickly and efficiently obtain accurate three-dimensional information of the moving object;

[0036] (2) The measuring method of the present application uses each image taken to collect information of each motion position and the pictures before and after it, and has high information utilization rate;

[0037] (3) The three-dimensional profile of the moving object restored by the measuring method of the present application has higher accuracy, clearer details and more complete high-frequency information retention. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is the main flow chart of the measuring method in the embodiment;

[0039] Figure 2 is the framework diagram of the convolutional neural network constructed in step 1 of the embodiment;

[0040] Figure 3 is a schematic diagram of the experimental device built in step 2-1 of the embodiment. DETAILED DESCRIPTION

[0041] The present application will be further described in detail below in combination with the embodiment and the accompanying drawings.

[0042] The structured light measuring method of the three-dimensional profile of the moving object of the embodiment, the main flow chart of the measuring method is shown in Figure 1 , and includes the following steps:

[0043] Step 1: Construct a convolutional neural network, as Figure 2As shown, the convolutional neural network includes a feature extraction network and a feature fusion network, the feature extraction network includes a first feature extraction module, a second feature extraction module, a third feature extraction module, a fourth feature extraction module and a fifth feature extraction module connected in sequence, the first feature extraction module, the second feature extraction module, the third feature extraction module and the fourth feature extraction module respectively include a first convolutional layer, a first linear rectifier layer, a second convolutional layer, a second linear rectifier layer and a down-sampling layer connected in sequence, the fifth feature extraction module includes a third convolutional layer, a third linear rectifier layer, a fourth convolutional layer and a fourth linear rectifier layer connected in sequence; the feature fusion network includes a first feature fusion module, a second feature fusion module, a third feature fusion module, a fourth feature fusion module and an image output module connected in sequence, the first feature fusion module, the second feature fusion module, the third feature fusion module and the fourth feature fusion module respectively include an up-sampling layer, a fifth convolutional layer, a fifth linear rectifier layer, a sixth convolutional layer and a sixth linear rectifier layer connected in sequence, and the image output module is a seventh convolutional layer; the image obtained after the second linear rectifier layer in the first feature extraction module is superimposed with the image after the up-sampling layer in the fourth feature fusion module, the image obtained after the second linear rectifier layer in the second feature extraction module is superimposed with the image after the up-sampling layer in the third feature fusion module, the image obtained after the second linear rectifier layer in the third feature extraction module is superimposed with the image after the up-sampling layer in the second feature fusion module, and the image obtained after the second linear rectifier layer in the fourth feature extraction module is superimposed with the image after the up-sampling layer in the first feature fusion module; specifically, the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer and the sixth convolutional layer respectively adopt a convolutional kernel with a size of 3*3 and a padding value of 2, the seventh convolutional layer adopts a convolutional kernel with a size of 1*1 and a padding value of 0, the down-sampling layer adopts a maximum pooling layer with a kernel size of 2*2, and the up-sampling layer has a kernel size of 2*2;

[0044] Step 2: generate training data and train the convolutional neural network to obtain a neural network model

[0045] Step 2-1, first build an experimental device, such as Figure 3 As shown, the experimental device includes a translation stepping motor, a rotation stepping motor, a projector and a camera controlled by a computer, the rotation stepping motor is arranged above the translation stepping motor, the translation stepping motor is placed on a horizontal table top, the translation stepping motor is used to control the rotation stepping motor to move in front and back, up and down, left and right three directions, the output end of the rotation stepping motor is provided with a load platform, the camera and the projector are respectively arranged above the load platform, the upper surface of the load platform in the initial position is taken as a reference plane, the connecting line of the optical center of the camera and the optical center of the projector is parallel to the reference plane, and the optical axis of the camera and the optical axis of the projector are in the same plane and the plane is perpendicular to the reference plane;

[0046] Step 2-2: Place the object on the object platform, use the projector to project N fringes on the object, N≥3, the expression of the fringe is:

[0047]

[0048] Where f is the fringe frequency, x and y are the horizontal and vertical coordinates of the pixel point in the camera respectively;

[0049] When the picture is projected onto the object, the height of the object will cause the fringe to deform, and the height information is converted into phase information. The camera captures N fringes with object phase information, and the expression is:

[0050]

[0051] Where A(x, y) is the background noise, B(x, y) is the reflection noise, is the phase value converted from the height of the object;

[0052] When the object moves in a certain direction, the coordinates of its pixel points change from (x, y) to (x+ε x ,y+ε y ), where ε x ,ε y are the motion noises in the horizontal and vertical directions respectively. Since the background noise does not change with the motion of the object, the reflection noise changes with the motion of the object. The expression of the fringe of the moving object is:

[0053]

[0054] Use the computer to control the initial position and motion mode of the translation stepper motor and the rotation stepper motor to ensure that the object can move along the same motion trajectory multiple times. Use the projector to project N fringes and use the projector to trigger the camera to take pictures synchronously. Use the computer to control the projection and the motion of the object to ensure that the position of the object in the multiple pictures taken by the camera is consistent with the position of the object when the first picture is taken during each motion of the object.

[0055] Let the object move along the same motion trajectory M times, and take N pictures each time to form a group, where M=N. The phase shift difference between each group of pictures taken is The expression of the taken images is:

[0056]

[0057] Take the N images with m=N / 2 and take the integer part as data group 1. The positions of the objects in the pictures in data group 1 are the same and are at the middle moment of the entire motion process. Take the N images with n=0 as data group 2. The phases of the objects in the pictures in data group 2 are consistent with those in data group 1, and the positions of the objects change in order.

[0058] Step 2-3: Repeat step 2-2, projecting different objects along different trajectories to obtain multiple data sets 1 and 2. Use multiple data sets 1 as the output of the convolutional neural network and multiple data sets 2 as the input of the convolutional neural network to obtain training data. Train the convolutional neural network and use the root mean square error as the loss function to calculate the difference between the input data and the output data, thereby evaluating the quality of the training results and determining whether the training is complete.

[0059] When training the neural network, the learning rate is set to 0.0001, and then reduced to 0.2 every 50 iterations. A good learning rate strategy is to encourage the neural network to converge quickly and ensure that it converges to the global optimum. After training, the neural network model is obtained.

[0060] Step 3: Perform 3D reconstruction of the moving object under test from the trained phase-shifted image.

[0061] The object to be tested is placed on a platform, and a computer is used to control the translational stepper motor and the rotary stepper motor to make the object to be tested move with the platform. t images are captured, where t>N. The kth to k+N-1th images are input into a neural network model to obtain an N-step phase-shifted image at the kth position without motion error.

[0062] The enclosed phase is calculated using the least squares method with an N-step phase shift diagram, and its expression is as follows:

[0063]

[0064] Unwrap the wrapped phase, remove the 2πfx term, and obtain the phase value caused by the object's height without motion error. By combining this with trigonometric methods, the height h(x,y) and phase value of the object under test can be obtained. Correspondence:

[0065]

[0066] Where d is the distance between the optical center of the camera and the optical center of the projector, and l is the height of the optical center of the camera from the reference plane;

[0067] After reconstructing the three-dimensional contour of the moving object at the k-th position, let k = 1, 2, ..., t-N+1 to obtain the three-dimensional contours of the moving object at k consecutive different positions, thus realizing the three-dimensional contour measurement of the moving object.

[0068] In a specific application case, the camera adopts Hikvision MV-CA004-10um, the projector adopts LightCrafter 3010 of Texas Instruments, the translation stepping motor adopts a three-axis stepping screw sliding table, and the rotation stepping motor adopts a rotation sliding table with a radius of 50 cm. In the experimental device built, the optical axis of the camera is perpendicular to the reference plane and is 600 mm away from the reference plane, and the optical center of the camera is 200 mm away from the optical center of the projector.

[0069] A group of different translation stepping motor directions and speeds v (v <100 mm / s) and a rotation stepping motor speed r (r <10 rad / s) are randomly set, an object (length, width and height are less than 100 mm) is placed at different positions of the object platform, ensuring that the object rotates around different points, and the initial positions of the translation stepping motor and the rotation stepping motor are adjusted to make the object in the field of view of the camera and saved. Four-step phase shift fringe patterns with a frequency of 64 and N = 4 are projected by the projector and captured by the camera, and the capture frequency is 200 frames / s, thereby obtaining four images. The translation stepping motor and the rotation stepping motor are restored to the initial position to repeat the projection and shooting three times, and the initial phase difference of the projected fringe is 0.5π each time. Thus, 16 pictures are obtained, and 512*512 images containing the object are cut out at the same position of the pictures, and 1, 2, 3 and 4 of them are set as data group 1, the object phases of which are 0, 0.5π, π and 1.5π respectively, and the object position is the middle time of this movement; 2, 6, 10 and 14 of them are set as data group 2, the object phases of which are 0, 0.5π, π and 1.5π respectively, and the object position changes in turn.

[0070] 1000 groups of different motor parameters are set and 100 different objects are shot, a total of 1000 data groups 1 and 1000 data groups 2 are obtained, all the data groups 1 are used as the output of the neural network, and all the data groups 2 are used as the input of the neural network for neural network training. In the training process, any 800 groups are set as the training set, and the remaining 200 groups are set as the validation set to avoid over-saturation of the neural network, and finally a neural network that can train the fringe pattern of the moving object to the fringe pattern of the object at the same position is obtained.

[0071] A group of stepping motor parameters different from the above setting is randomly set, 20 consecutive four-step phase shift pictures are projected and shot, and 17 groups of four-step phase shift pictures without motion error are obtained by removing the motion error of 1-4, 2-5, 3-6, etc. 17 groups through the neural network, and the three-dimensional profile of the object position in the second to eighteenth pictures is obtained by phase unwrapping. The average error caused by movement can be reduced from 2.274 mm to 0.416 mm by the method of the present application.

Claims

1. A structured light method of three-dimensional profilometry of a moving object, characterized in that, The method comprises the following steps: Step 1: constructing a convolutional neural network; Step 2: generating training data and training the convolutional neural network to obtain a neural network model Step 2-1, first, an experimental device is built, which comprises a computer-controlled translation stepper motor, a rotation stepper motor, a projector and a camera, the rotation stepper motor is arranged above the translation stepper motor, the translation stepper motor is placed on a horizontal table top, the translation stepper motor is used to control the rotation stepper motor to move in the front-back, up-down and left-right directions, an object carrying platform is installed at the output end of the rotation stepper motor, the camera and the projector are arranged above the object carrying platform respectively, the upper surface of the object carrying platform in an initial position is taken as a reference plane, the line connecting the optical center of the camera and the optical center of the projector is parallel to the reference plane, the optical axis of the camera and the optical axis of the projector are in the same plane and the plane is perpendicular to the reference plane; Step 2-2: placing an object on the object carrying platform, projecting N stripes on the object using the projector, N≥3, the expression of the stripes is: Wherein f is the stripe frequency, x and y are the horizontal coordinate and vertical coordinate of the pixel point in the camera respectively; When the picture is projected onto the object, the height of the object will cause the stripes to deform, and the height information is converted into phase information, the camera captures N stripes with the phase information of the object, and the expression is: where A(x,y) is the background noise, B(x,y) is the reflection noise, is the phase value of the object height transformation; When the object moves in a certain direction, the coordinates of the pixel points change from (x, y) to (x+ε x ,y+ε y ), where ε x ,ε y are the motion noises in the horizontal coordinate direction and the vertical coordinate direction, respectively. Since the background noise does not change with the movement of the object, the reflection noise changes with the movement of the object, and the expression of the stripe of the moving object is: The initial position and movement mode of the computer-controlled translation stepper motor and rotation stepper motor are used to ensure that the object can move along the same movement trajectory multiple times; The projector is used to project N stripes and the camera is triggered synchronously by the projector; The computer is used to control the projection and the movement of the object to be synchronous, so that the position of the object in the multiple pictures captured by the camera in each movement is consistent with the position of the object in the first shooting; Let the object move along the same movement trajectory M times, and each time N images are taken as a group, wherein M=N, each group of images The phase shift difference of the picture is The expression of the picture taken is: n=0,1,2,…,N-1,m=0,1,2,…,M-1 Take the N images of m=N / 2 and take the integer part downward as data group 1, the object positions in the pictures in data group 1 are the same and are in the middle moment of the whole movement process; take the N images of n=0 as data group 2, the object phases in data group 2 are consistent with data group 1, and the object positions change in order; Step 2-3: repeat step 2-2, project different objects along different trajectories, and take multiple data groups 1 and multiple data groups 2, use multiple data groups 1 as the output of the convolutional neural network, use multiple data groups 2 as the input of the convolutional neural network, and obtain training data; the convolutional neural network is trained, and the root mean square error is used as the loss function to calculate the difference between the input data and the output data, so as to evaluate the advantages and disadvantages of the training result and judge whether the training is completed; after the training is completed, a neural network model is obtained; Step 3: three-dimensional reconstruction of the measured moving object is performed by using the trained phase shift map Place the object to be measured on the object platform, use the computer to control the translation stepper motor and the rotation stepper motor, make the object to be measured move with the object platform, and take t images, t>N, input the kth to k+N-1 images taken into the neural network model to obtain N step phase shift images of the object to be measured at the kth position without motion error; The wrapped phase is obtained by least square calculation using the N step phase shift images, and the expression is: The package phase is unwrapped, and the 2πfx term is removed, to obtain the phase value caused by the object height without motion error In combination with the triangulation method, the height h(x, y) of the object to be measured and the phase value are related as follows: ​ Wherein d is the distance between the optical center of the camera and the optical center of the projector, and l is the height from the optical center of the camera to the reference plane; After the three-dimensional profile of the object to be measured at the kth position is restored, k is set to 1, 2,..., t-N+1, the three-dimensional profiles of the object to be measured at the k different positions are obtained, and the three-dimensional profile measurement of the moving object is realized.

2. A structured light method of measuring the three-dimensional profile of a moving object according to claim 1, wherein, The convolutional neural network constructed in step 1 includes a feature extraction network and a feature fusion network, the feature extraction network includes a first feature extraction module, a second feature extraction module, a third feature extraction module, a fourth feature extraction module and a fifth feature extraction module connected in sequence, the first feature extraction module, the second feature extraction module, the third feature extraction module and the fourth feature extraction module each include a first convolutional layer, a first linear rectifier layer, a second convolutional layer, a second linear rectifier layer and a down-sampling layer connected in sequence, the fifth feature extraction module includes a third convolutional layer, a third linear rectifier layer, a fourth convolutional layer and a fourth linear rectifier layer connected in sequence, the feature fusion network includes a first feature fusion module, a second feature fusion module, a third feature fusion module, a fourth feature fusion module and an image output module connected in sequence, the first feature fusion module, the second feature fusion module, the third feature fusion module and the fourth feature fusion module each include an up-sampling layer, a fifth convolutional layer, a fifth linear rectifier layer, a sixth convolutional layer and a sixth linear rectifier layer connected in sequence, and the image output module is a seventh convolutional layer, the image obtained after the second linear rectifier layer in the first feature extraction module is superimposed with the image after the up-sampling layer in the fourth feature fusion module, the image obtained after the second linear rectifier layer in the second feature extraction module is superimposed with the image after the up-sampling layer in the third feature fusion module, the image obtained after the second linear rectifier layer in the third feature extraction module is superimposed with the image after the up-sampling layer in the second feature fusion module, and the image obtained after the second linear rectifier layer in the fourth feature extraction module is superimposed with the image after the up-sampling layer in the first feature fusion module.

3. A structured light method of measuring the three-dimensional profile of a moving object according to claim 2, wherein, The first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer and the sixth convolutional layer each use a convolutional kernel with a size of 3x3 and a padding value of 2, the seventh convolutional layer uses a convolutional kernel with a size of 1x1 and a padding value of 0, the down-sampling layer uses a maximum pooling layer with a kernel size of 2x2, and the up-sampling layer has a kernel size of 2x2.

Images