CN106683173A - A Method of Improving the Density of 3D Reconstruction Point Cloud Based on Neighborhood Block Matching - Google Patents

Abstract

The invention discloses a method for improving the density of a three-dimensional reconstruction point cloud based on neighborhood block matching, comprising: using a three-dimensional reconstruction algorithm based on an image sequence to obtain a rough and sparse object point cloud; transformation matrix in the space; process the original image again, and use the neighborhood-based block matching algorithm to perform dense feature matching in the image; Check the legality, and map the feature points that meet the requirements to the corresponding positions in the 3D point cloud; use the outlier deletion algorithm based on the object outline to perform an outlier filtering on the obtained point cloud, and perform a color remapping to obtain the quality A dense point cloud that is much better than the original point cloud. The invention can obtain a point cloud whose quality and density are much higher than those of the traditional algorithm, can greatly improve the effect of the original algorithm, and improve the reconstruction quality; it has high universality and strong robustness.

Description

A Method of Improving the Density of 3D Reconstruction Point Cloud Based on Neighborhood Block Matching

technical field

The invention belongs to the technical field of computer vision, and in particular relates to a method for increasing the density of a three-dimensional reconstruction point cloud based on neighborhood block matching.

Background technique

3D reconstruction based on image sequences involves many disciplines, and belongs to the reverse process of camera imaging principles. The research scope mainly includes: object target recognition, feature detection, feature matching and other fields. Since its birth, 3D reconstruction technology has become one of the hot spots and difficulties in the field of computer vision. Its input only requires general color images, and its high universality provides unparalleled convenience when modeling objects in the real world. However, since there are too many parameters that need to be calculated and measured in 3D reconstruction, how to obtain a high-quality reconstruction model has become a difficult problem to overcome. Usually, the core step of 3D reconstruction is how to obtain a high-quality point cloud that can sufficiently express the information of the object model. Generally, 3D reconstruction is divided into the following two categories: (1) 3D reconstruction methods based on image sequences. The 3D reconstruction method based on the image sequence obtains the input image by shooting around the object. Get the camera transformation between two frames based on the features detected in the image and the feature matches between the images. By minimizing the error of the camera position of all images, the rotation and translation of the camera relative to the world coordinate system are finally determined when each frame of image is taken. According to this information, the real position of each feature point in space can be determined, so as to obtain the object's position. Sparse point cloud. Then increase the density of the obtained point cloud through some point cloud growth algorithms. (2) 3D reconstruction method based on depth camera. The 3D reconstruction method based on the depth camera is to capture images and depth maps in real time by using the depth camera. That is, for each frame of image, the position of each pixel in the frame can be obtained. With the movement of the camera, the point cloud of the next frame is continuously matched with the point cloud of the frame, the movement of the camera position between the two frames is detected, and the point clouds of different frames are fused together to obtain the dense point cloud of the object. . Among the above methods, the 3D reconstruction based on the image sequence has the advantages of easy access to input data, low requirements on the shooting environment, and strong robustness. However, since the image can only provide information in two-dimensional space, there are many errors and noises in the calculation when solving the three-dimensional space coordinates in reverse. Not high enough, resulting in unsatisfactory model results. The method based on the depth camera can obtain sufficiently dense feature points because each pixel point in the acquired image can be regarded as a feature point. However, because it requires a high-quality depth camera, and it is difficult to cope with long-distance shooting. At the same time, the maximum range of civilian depth cameras is about 10m, so this method has too high requirements on the shooting environment and equipment. However, 3D reconstruction based on image sequences is not subject to this restriction, as long as the object presents enough features in the image.

To sum up, the quality of the point cloud obtained by 3D reconstruction based on image sequences is not high enough, and there are large errors; the method based on the depth camera has high requirements on the shooting environment, requires a high-quality depth camera, and is difficult to cope with the distance shooting.

Contents of the invention

The purpose of the present invention is to provide a method for improving the density of 3D reconstruction point cloud based on neighborhood block matching, aiming at solving the problem that the quality of the point cloud obtained by 3D reconstruction based on image sequence is not high enough; the method based on depth camera is too demanding on the shooting environment. High, requires a high-quality depth camera, and struggles to cope with longer distance shots.

The present invention is achieved in this way, a method for improving the density of a three-dimensional reconstruction point cloud based on neighborhood block matching, the method for increasing the density of a three-dimensional reconstruction point cloud based on neighborhood block matching includes the following steps:

Step 1, using a 3D reconstruction algorithm based on image sequences to obtain rough and sparse object point clouds, and obtain the transformation matrix of the camera in 3D space when each frame of image is taken;

Step 2, process the original image again, and use the neighborhood-based block matching algorithm to perform dense feature matching in the image;

Step 3: Next, according to the obtained position of the camera in space, the legality of the obtained dense feature point pairs is checked, and the feature points that meet the requirements are mapped to the corresponding positions in the 3D point cloud;

Step 4: Use the outlier deletion algorithm based on the object outline to perform an outlier filtering on the obtained point cloud, and perform a color remapping to obtain a dense point cloud with a final quality much better than the original point cloud.

Further, the method for increasing the density of the 3D reconstruction point cloud based on neighborhood block matching specifically includes the following steps:

In the first step, traditional 3D reconstruction is performed using a sequence of images taken around the camera;

The second step is to select adjacent images to perform dense feature generation algorithm based on neighborhood block matching, and determine the neighborhood step size n, that is, each image will calculate dense features with two left and right images;

In the third step, after determining the step size n of adjacent image matching, traverse the entire image sequence, take each image as a reference image, and perform a neighborhood-based block matching algorithm with the n-frame images behind it; in the original image, along How many pixels in the row and column directions take a point as the feature point to be matched, which is used for feature calculation based on block matching;

The fourth step, after determining the sampling rate r, for each pair of images (P ₁ , P ₂ ) that need to be matched with features, the matching point set Pt in P ₁ that needs to be calculated is obtained; traversing all the feature points in Pt, Take each feature point as the center, select an image block of X*Y size around it, use the interpolation algorithm to expand its ranks and columns to 8 times the original, and record it as I _src ; in P ₂ , use the same coordinate point as In the center, select an image block of M*N size, and use an interpolation algorithm to expand its rows and columns to 8 times the original size, which is denoted as I _dist ;

The fifth step, when the size of the I _dist image block is large enough, the image block I _match corresponding to the image block I _src can be found _; at this time, the position of the center of the image block I _match in Figure P2 is to construct the image block I _{src The} matching points corresponding to the feature points used in the image P2;

In the sixth step, after the feature matching is completed, the dense feature pairs between the required image pairs are obtained; for each frame in the image sequence, the image that has undergone feature matching calculation with it is recorded as P ₂ ; the shooting P ₁ and P ₂ when the camera optical center positions are recorded as C ₁ and C ₂ respectively; traverse the positions of all feature points in the feature point set Pt of P ₁ , and find its position Pt ₁ on the phase plane in three-dimensional space ^; At the same time, find the position Pt ₂ ^' of the matching point of Pt ₁ on the image P ₂ on the phase plane in the three-dimensional space; and make rays C ₁ Pt ₁ ' and C ₂ Pt ₂ ^' in the three-dimensional space. Get rays L ₁ and L ₂ ;

The seventh step, for the obtained rays L ₁ and L ₂ , if the matching point pair Pt ₁ and Pt ₂ is the correct matching point, then L ₁ and L ₂ should intersect at the corresponding point of the object model in space; Set a threshold T, when the distance between the closest points between the two straight lines of different planes is less than T, this pair of matching points is a legal pair of matching points;

The eighth step, for the obtained legal matching point pair, if the made rays L ₁ and L ₂ intersect at one point, then add the intersection point as a new feature point into the initial point cloud; if L ₁ and L ₂ are the closest point distance is less than the threshold T, then take the midpoint position of the nearest point pair of L ₁ and L ₂ as a new three-dimensional point and add it to the initial point cloud;

The ninth step is to perform noise filtering based on the outline of the object. For each frame of image, extract the outline of the object, and back-project the point cloud to the position of the phase plane in the three-dimensional space through the obtained camera transformation matrix. Points are retained; after backprojection, in any frame, points that appear outside the outline of the object are deleted from the point cloud.

The tenth step is to back-project the point cloud to the phase plane of each frame in space, record the point in the point cloud of each pixel on the phase plane that is closest to it during back-projection, and convert the pixel of the pixel point on the image to The value is assigned to the point in the point cloud closest to him, so as to complete the remapping of the point cloud color.

Further, the camera transformation matrix includes the rotation and translation of the optical center position of the camera relative to the origin of the world coordinate system, ie [R|T], where R is a rotation matrix and T is a translation vector.

Further, in the second step, a shooting method of rotating 5° between every two frames is used, and the step size used is 2.

Further, in the step 3, a sampling rate of one pixel every two rows and two columns is used; the number of selected pixels is directly proportional to the time overhead of the algorithm.

Further, the values of M and N in step 4 need to be greater than X and Y respectively; use the parameter settings where X and Y are 40, and M and N are 80.

Further, in the step five, the formula for calculating the position (x, y) correlation coefficient R is as follows:

in:

Among them, T(x, y) represents the pixel value at the coordinate (x, y) position in the image I _src , and I(x, y) represents the pixel value at the coordinate (x, y) position in the image I _dist . w and h represent the width and height of the image, respectively.

Further, in the step six, when searching for feature points on the phase plane of each frame, the transformation matrix [R|T] of the camera needs to be used, the focal length of the photo is f, and in a certain frame F, the position of the feature points is at The image is (x, y), and its rotation and translation matrix is [R|T], then its three-dimensional coordinates (X, Y, Z) in the world coordinate system are calculated as follows:

Further, in the step seven, the set threshold for considering the two rays L ₁ and L ₂ to be legal is 1e ⁻² ;

In the eighth step, the parameter k is 5;

In said step nine, the method of back-projecting the point cloud to the phase plane of the i-th frame needs to use the transformation matrix [R|T] of the camera; a point (X, Y, Z) in the point cloud, its projection in the i-th frame The formula for calculating the coordinates (u, v) in is:

Among them, f is the focal length of the i-th frame;

In the tenth step, if multiple phase planes map the pixel of a certain point to a point Pt in the point cloud, then the RGB of Pt takes the average value of the pixels of the corresponding points of these planes.

Another object of the present invention is to provide a computer vision system that applies the method for increasing the density of a 3D reconstruction point cloud based on neighborhood block matching.

The method for improving the density of 3D reconstruction point cloud based on neighborhood block matching provided by the present invention has no special requirements for image sequences and does not rely too much on parameter adjustment, which can greatly improve the quality and density of point cloud. On the basis of the three-dimensional reconstruction based on the image sequence, the present invention further expands the obtained point cloud to increase its completeness and density. On the basis of the existing algorithm for obtaining sparse point cloud by 3D reconstruction, the feature of point cloud is greatly increased, thereby improving the algorithm of the final point cloud result.

Compared with the three-dimensional reconstruction algorithm based on SIFT features, the present invention can obtain point cloud density, completeness, and color restoration degree far better than common algorithms. At the same time, it is difficult for SIFT to extract correct feature points for areas with relatively smooth features, but the present invention is based on a fast matching method, which can maximize the use of information in the neighborhood of each pixel and increase correct feature points as much as possible. At the same time, based on the calculated camera transformation moment, it eliminates the inaccurate shortcomings of the traditional fast matching algorithm, and can accurately supplement missing or too sparse point clouds into better quality point clouds.

Compared with the method based on the depth image, since this method requires a corresponding depth image for each frame, and the displacement between adjacent frames cannot be too large, it is difficult to have real use value in the scene where the depth image cannot be obtained. At present, the high-resolution depth cameras for civilian use can only detect the depth range within 10 meters. While laser detection has a sufficient distance, it cannot achieve sufficient accuracy. As long as the present invention can obtain simple images, high-quality model reconstruction can be realized, and the displacement between two frames is not as accurate as the method based on depth images, which increases the universality and ease of use; Perform traditional 3D reconstruction of images based on SIFT features; determine the neighborhood matching step size n; determine the sampling step size, that is, the feature points that need to be matched; traverse the image sequence, and perform feature extraction based on neighborhood block matching for the image pairs that need to be calculated; Based on the camera matrix obtained in 1 and the matching features obtained in 4, the features in 4 are filtered; based on the image contour, the point cloud in space is filtered and color corrected again. The present invention can obtain a point cloud whose quality and density are much higher than those of the traditional algorithm, can greatly improve the effect of the original algorithm, and improve the reconstruction quality; it is a highly universal and robust method for improving the final point cloud quality of 3D reconstruction Methods.

Description of drawings

Fig. 1 is a flowchart of a method for increasing the density of a 3D reconstruction point cloud based on neighborhood block matching provided by an embodiment of the present invention.

Fig. 2 is a flowchart of Embodiment 1 provided by the embodiment of the present invention.

detailed description

In order to make the object, technical solution and advantages of the present invention more clear, the present invention will be further described in detail below in conjunction with the examples. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

The application principle of the present invention will be described in detail below in conjunction with the accompanying drawings.

As shown in Figure 1, the method for improving the density of the 3D reconstruction point cloud based on neighborhood block matching provided by the embodiment of the present invention includes the following steps:

S101: Obtain a rough and sparse object point cloud using a traditional image sequence-based three-dimensional reconstruction algorithm, and obtain the transformation matrix of the camera in three-dimensional space when each frame of image is captured;

S102: Process the original image again, and use a neighborhood-based block matching algorithm to perform dense feature matching in the image. This feature matching method can obtain matching feature pairs that are sufficiently dense but have low precision;

S103: Next, according to the obtained position of the camera in space, perform a legality check on the obtained dense feature point pairs, and map the required feature points to corresponding positions in the 3D point cloud;

S104: Use an outlier deletion algorithm based on the object outline to perform outlier filtering on the obtained point cloud, and perform a color remapping to obtain a dense point cloud with a final quality much better than the original point cloud.

The application principle of the present invention will be further described below in combination with specific embodiments.

Example 1:

The method for increasing the density of the 3D reconstruction point cloud based on neighborhood block matching provided by the embodiment of the present invention includes the following steps:

Step 1: Shoot around the target object to obtain images from various angles.

Step 2: Use the traditional 3D reconstruction method based on the SIFT feature matching algorithm for reconstruction, and obtain the camera rotation matrix when shooting each frame.

Step 3: Determine the step size for image matching, and use the feature matching algorithm based on neighborhood block matching to extract dense matching feature points. The feature point extraction formula is described in detail in the description of S5.

Step 4: After obtaining the dense matching feature points, check their validity in the three-dimensional space, delete the feature points that do not meet the threshold condition from the point set, and calculate their real positions in the space that meet the threshold condition.

Step 5: Add all legal point clouds in step 4 to the original point cloud to improve the density and completeness of the original point cloud.

Step 6: Carry out the contour-based noise point filtering algorithm and color correction algorithm to filter out a small number of wrong points and correct the color of the expanded point cloud to complete the entire algorithm process.

According to the above steps, a more dense and complete high-quality point cloud is finally obtained.

The method for improving the density of the 3D reconstruction point cloud based on neighborhood block matching provided by the embodiment of the present invention specifically includes the following steps:

S1 uses a sequence of images to perform traditional 3D reconstruction. Because traditional 3D reconstruction is based on the SIFT feature algorithm, and SIFT has high accuracy and can perform feature extraction at the sub-pixel level, it can obtain accurate camera position transformation in 3D space. matrix.

S2 selects adjacent images to perform a dense feature generation algorithm based on neighborhood block matching. First, determine the neighborhood step size n, that is, each image will perform dense feature calculation with two left and right images.

S3 After determining the step size n of adjacent image matching, traverse the entire image sequence, take each image as a reference image, and perform a neighborhood-based block matching algorithm with the n frames of images behind it. The algorithm first needs to determine the sampling rate r, that is, in the original image, how many pixels along the row and column directions take a point as a feature point to be matched, which is used for feature calculation based on block matching.

S4 After determining the sampling rate r, for each pair of images (P ₁ , P ₂ ) that need to be matched with features, the matching point set Pt in P ₁ that needs to be calculated is obtained. Traverse all the feature points in Pt, take each feature point as the center, select an image block of size X*Y around it, use the interpolation algorithm to expand its row and column to 8 times the original size, and record it as I _src . In P ₂ , with the same coordinate point as the center, select an image block of M*N size, and also use the interpolation algorithm to expand its ranks and columns to 8 times the original size, which is recorded as I _dist .

S5 is based on the assumption that when the camera is shooting around the image, the motion range between two adjacent frames is very small. Therefore, in S4, when the size of the image block I _dist is large enough, the image block I _match corresponding to the image block I _src can be found therein. At this time, the position of the center of the image block I _match in Figure P ₂ is the corresponding matching point on the image P ₂ of the feature points used when constructing the image block I _src . Due to the previous interpolation work, the final block matching The precision is 1/8 pixel.

After S6 completes the feature matching in S5, the dense feature pairs between the required graphs are obtained. However, due to the change of the camera angle of the object between the two images, it is only possible to find an image block as close as possible to I _src in I _dist . Therefore, the feature matching pairs extracted in this way have relatively large noise and need further processing. In S1, the position of the camera in three-dimensional space when each frame of the image is taken is obtained. In S5, the matching feature point pairs between image pairs are obtained. For each frame in the image sequence, take P ₁ as an example, and denote the image that has undergone feature matching calculation with it as P ₂ . The optical center positions of the camera when shooting P ₁ and P ₂ are respectively denoted as C ₁ and C ₂ . Traversing the positions of all the feature points in the feature point set Pt of P ₁ , taking the point Pt ₁ as an example, find its position Pt ₁ ^' on the phase plane in the three-dimensional space. At the same time, find the position Pt ₂ ^' of the matching point of Pt ₁ on the image P ₂ on the phase plane in the three-dimensional space. And make rays C ₁ Pt ₁ ^' and C ₂ Pt ₂ ^' in three-dimensional space. Rays L ₁ and L ₂ are obtained.

S7 For the rays L ₁ and L ₂ obtained in S6, if the matching point pair Pt ₁ and Pt ₂ are correct matching points, then L ₁ and L ₂ should intersect at the corresponding point of the object model in space. However, due to the various errors introduced in the calculation and the inherent errors of floating-point calculations, it is difficult to find a situation where it really intersects at a point. Therefore, in most cases, L ₁ and L ₂ are two straight lines with different planes, and a threshold T is set in the algorithm. When the distance between the closest points between the two straight lines with different planes is less than T, the It is considered that this pair of matching points is a legal pair of matching points.

S8 For the legal matching point pair obtained in S7, if the made rays L ₁ and L ₂ intersect at one point, add the intersection point into the initial point cloud as a new feature point. If L ₁ and L ₂ are straight lines on different planes whose closest point distance is less than the threshold T. Then take the midpoint position of the nearest point pair of L ₁ and L ₂ as a new three-dimensional point and add it to the initial point cloud. For the newly added point in the 3D point cloud, the color of the new point is the mean value of the colors of the nearest k points around it. The color of the point cloud obtained in this way has errors, and the specific correction method is shown in S10.

Although the results obtained by S9 in S8 have filtered out most of the wrong matching points, there are still errors in a small number of matching points, so it is filtered out based on the object outline. For each frame of image, object contours are extracted. Through the camera transformation matrix obtained in S1, the point cloud can be back-projected to the position of the phase plane in the three-dimensional space. After this, the points within the contour area are kept. After the back projection, in any frame, the points that appear outside the outline of the object are deleted from the point cloud.

The color of the newly added point cloud of S10 still has errors, and the method in S9 is used to back-project the point cloud to the phase plane of each frame in space. Record the point in the point cloud closest to each pixel on the phase plane during back projection. Assign the pixel value of the pixel point on the image to the point in the point cloud closest to it, so as to complete the remapping of the point cloud color.

The camera transformation matrix obtained in the step S1 includes the rotation and translation of the optical center position of the camera relative to the origin of the world coordinate system, that is, [R|T], where R is a rotation matrix and T is a translation vector.

The maximum value of the matching step selected in the step S2 depends on the magnitude of the motion during the camera shooting process. In this algorithm, the shooting method of rotating 5° between every two frames is used, and the step size used is 2. The larger the range of change between every two frames, the smaller the corresponding step size should be, and vice versa.

The sampling rate in step S3 will affect the number of matching points finally generated. In this method, a sampling rate of one pixel every 2 rows and 2 columns is used. The number of selected pixels is proportional to the time overhead of the algorithm.

The values of M and N in the step S4 need to be greater than X and Y respectively. In the algorithm experiment, the parameter settings of X, Y are 40, and M and N are 80 are used.

In the step S5, searching for the image block I _src in I _dist uses an image matching algorithm based on a correlation coefficient. That is, the image block I _src moves from left to right and from top to bottom on the image block I _dist , calculates the correlation coefficient R of each position I _src and the image block in the area of I _dist covered by it, and selects the position with the largest correlation coefficient The area covered by I _src in is used as the image block I _match . In the process of sliding the I _src , the coordinates (x, y) of the upper left corner of the I _src in the I _dist are recorded as the coordinates of the current position. The formula for calculating the position (x, y) correlation coefficient R is as follows:

in:

Among them, T(x, y) represents the pixel value at the coordinate (x, y) position in the image I _src , and I(x, y) represents the pixel value at the coordinate (x, y) position in the image I _dist . w and h represent the width and height of the image, respectively.

In step S6, when looking for feature points on the phase plane of each frame, the transformation matrix [R|T] of the camera needs to be used. The specific formula is as follows, assuming that the focal length of the photo is f, in a certain frame F, the position of the feature point is (x, y) on the image, and its rotation and translation matrix is [R|T], then its position in the world coordinate system The three-dimensional coordinate calculation formula is as follows:

In step S7, the threshold set in the experiment to consider the two rays L ₁ and L ₂ legal is 1e ⁻² .

In step S8, for the selection of parameter k, 5 is used in the experiment.

In step S9, the method of back-projecting the point cloud to the phase plane of the i-th frame needs to use the transformation matrix [R|T] of the camera. Assuming a point (X, Y, Z) in the point cloud, the calculation formula for the coordinates (u, v) of its projection in the i-th frame is:

Among them, f is the focal length of the i-th frame.

In step S10, if multiple phase planes map the pixel of a certain point to a point Pt in the point cloud, then the RGB of Pt takes the average value of the pixels of the corresponding points of these planes.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention should be included in the protection of the present invention. within range.

Claims (10)

1. A method for improving the dense degree of three-dimensional reconstruction point cloud based on neighborhood block matching, characterized in that, the method for improving the dense degree of three-dimensional reconstruction point cloud based on neighborhood block matching may further comprise the steps: Step 1, using a 3D reconstruction algorithm based on image sequences to obtain rough and sparse object point clouds, and obtain the transformation matrix of the camera in 3D space when each frame of image is taken; Step 2, process the original image again, and use the neighborhood-based block matching algorithm to perform dense feature matching in the image; Step 3: Next, according to the obtained position of the camera in space, the legality of the obtained dense feature point pairs is checked, and the feature points that meet the requirements are mapped to the corresponding positions in the 3D point cloud; Step 4: Use the outlier deletion algorithm based on the object outline to perform an outlier filtering on the obtained point cloud, and perform a color remapping to obtain a dense point cloud with a final quality much better than the original point cloud. 2. the method for improving the dense degree of three-dimensional reconstruction point cloud based on neighborhood block matching as claimed in claim 1, is characterized in that, the method for improving the dense degree of three-dimensional reconstruction point cloud based on neighborhood block matching specifically comprises the following steps: In the first step, traditional 3D reconstruction is performed using a sequence of images taken around the camera; The second step is to select adjacent images to perform dense feature generation algorithm based on neighborhood block matching, and determine the neighborhood step size n, that is, each image will calculate dense features with two left and right images; In the third step, after determining the step size n of adjacent image matching, traverse the entire image sequence, take each image as a reference image, and perform a neighborhood-based block matching algorithm with the n-frame images behind it; in the original image, along How many pixels in the row and column directions take a point as the feature point to be matched, which is used for feature calculation based on block matching; The fourth step, after determining the sampling rate r, for each pair of images (P ₁ , P ₂ ) that need to be matched with features, the matching point set Pt in P ₁ that needs to be calculated is obtained; traversing all the feature points in Pt, Take each feature point as the center, select an image block of X*Y size around it, use the interpolation algorithm to expand its ranks and columns to 8 times the original, and record it as I _src ; in P ₂ , use the same coordinate point as In the center, select an image block of M*N size, and use an interpolation algorithm to expand its rows and columns to 8 times the original size, which is denoted as I _dist ; The fifth step, when the size of the I _dist image block is large enough, the image block I _match corresponding to the image block I _src can be found _; at this time, the position of the center of the image block I _match in Figure P2 is to construct the image block I _{src The} matching points corresponding to the feature points used in the image P2; In the sixth step, after the feature matching is completed, the dense feature pairs between the required image pairs are obtained; for each frame in the image sequence, the image that has undergone feature matching calculation with it is recorded as P ₂ ; the shooting P ₁ and P ₂ when the camera optical center positions are recorded as C ₁ and C ₂ respectively; traverse the positions of all feature points in the feature point set Pt of P ₁ , and find its position Pt ₁ on the phase plane in three-dimensional space; At the same time, find the position Pt ₂ ' of the matching point of Pt ₁ on the image P ₂ on the phase plane in three-dimensional space; and make rays C ₁ Pt ₁ ' and C ₂ Pt ₂ ' in three-dimensional space; get rays L ₁ and L ₂ ; The seventh step, for the obtained rays L ₁ and L ₂ , if the matching point pair Pt ₁ and Pt ₂ is the correct matching point, then L ₁ and L ₂ should intersect at the corresponding point of the object model in space; Set a threshold T, when the distance between the closest points between the two straight lines of different planes is less than T, this pair of matching points is a legal pair of matching points; The eighth step, for the obtained legal matching point pair, if the made rays L ₁ and L ₂ intersect at one point, then add the intersection point as a new feature point into the initial point cloud; if L ₁ and L ₂ are the closest point distance is less than the threshold T, then take the midpoint position of the nearest point pair of L ₁ and L ₂ as a new three-dimensional point and add it to the initial point cloud; The ninth step is to perform noise filtering based on the outline of the object. For each frame of image, extract the outline of the object, and back-project the point cloud to the position of the phase plane in the three-dimensional space through the obtained camera transformation matrix. The points are retained; after the back projection, in any frame, the points that appear outside the outline of the object are deleted from the point cloud; The tenth step is to back-project the point cloud to the phase plane of each frame in space, record the point in the point cloud of each pixel on the phase plane that is closest to it during back-projection, and convert the pixel of the pixel point on the image to The value is assigned to the point in the point cloud closest to him, so as to complete the remapping of the point cloud color. 3. The method for improving the denseness of a three-dimensional reconstruction point cloud based on neighborhood block matching as claimed in claim 2, wherein the camera transformation matrix includes the rotation and translation of the camera optical center position relative to the origin of the world coordinate system, i.e. [R|T], where R is the rotation matrix and T is the translation vector. 4. The method for improving the denseness of 3D reconstruction point cloud based on neighborhood block matching as claimed in claim 2, characterized in that, in said step 2, the shooting mode of rotating 5° between every two frames is used, and the step size used is for 2. 5. the method for improving the dense degree of three-dimensional reconstruction point cloud based on neighborhood block matching as claimed in claim 2, is characterized in that, in described step 3, use every 2 rows and 2 columns to get the sampling rate of a pixel point; Selected The number of pixels is directly proportional to the time overhead of the algorithm. 6. The method for improving the density of three-dimensional reconstruction point clouds based on neighborhood block matching as claimed in claim 2, wherein the values of M and N in the step 4 need to be greater than X and Y respectively; using X, Y is 40, M and N are parameter settings of 80. 7. the method for improving the density of three-dimensional reconstruction point cloud based on neighborhood block matching as claimed in claim 2, is characterized in that, in described step 5, the formula of calculating position (x, y) correlation coefficient R is as follows: $\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; I</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$ in: ${\mathrm{<span\; class="notranslate">\; T</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; *</span>{<span\; class="notranslate">\; \&Sigma;</span>}_{{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}}\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$ ${\mathrm{<span\; class="notranslate">\; I</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; *</span>{<span\; class="notranslate">\; \&Sigma;</span>}_{{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}}\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$ Wherein, T represents the image I _src , and I represents the image I _dist . 8. The method for improving the density of 3D reconstruction point cloud based on neighborhood block matching as claimed in claim 2, characterized in that, in said step 6, when looking for feature points on the phase plane of each frame, it is necessary to use a camera The transformation matrix [R|T], the focal length of the photo is f, in a certain frame F, the position of the feature point is (x, y) on the image, and its rotation and translation matrix is [R|T], then it is in the world The three-dimensional coordinate calculation formula in the coordinate system is as follows: $\left[\begin{array}{c}\mathrm{<span\; class="notranslate">\; x</span>}\\ \mathrm{<span\; class="notranslate">\; Y</span>}\\ \mathrm{<span\; class="notranslate">\; Z</span>}\end{array}\right]<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; R</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; (</span>\left[\begin{array}{c}\mathrm{<span\; class="notranslate">\; x</span>}\\ \mathrm{<span\; class="notranslate">\; the\; y</span>}\\ \mathrm{<span\; class="notranslate">\; f</span>}\end{array}\right]<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; .</span>$ 9. The method for improving the density of 3D reconstruction point cloud based on neighborhood block matching as claimed in claim 2, characterized in that, in said step 7, the set threshold for considering the two rays L ₁ and L ₂ to be legal is 1e ^{− 2} ; In the eighth step, the parameter k is 5; In said step nine, the method of back-projecting the point cloud to the phase plane of the i-th frame needs to use the transformation matrix [R|T] of the camera; a point (X, Y, Z) in the point cloud, its projection in the i-th frame The formula for calculating the coordinates (u, v) in is: $\left[\begin{array}{c}\mathrm{<span\; class="notranslate">\; u</span>}\\ \mathrm{<span\; class="notranslate">\; v</span>}\\ \mathrm{<span\; class="notranslate">\; 1</span>}\end{array}\right]<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; \&CenterDot;</span>\left[\begin{array}{c}\mathrm{<span\; class="notranslate">\; x</span>}\\ \mathrm{<span\; class="notranslate">\; Y</span>}\\ \mathrm{<span\; class="notranslate">\; Z</span>}\end{array}\right]<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; ;</span>$ Among them, f is the focal length of the i-th frame; In the tenth step, if multiple phase planes map the pixel of a certain point to a point Pt in the point cloud, then the RGB of Pt takes the average value of the pixels of the corresponding points of these planes. 10. A computer vision system applying the method for increasing the density of a three-dimensional reconstruction point cloud based on neighborhood block matching according to any one of claims 1 to 9.