CN112115953A - Optimized ORB algorithm based on RGB-D camera combined with plane detection and random sampling consistency algorithm - Google Patents

Abstract

The invention provides an optimized ORB algorithm based on an RGB-D camera combined with a plane detection and random sampling consistency algorithm, which comprises the following steps: s1: acquiring image data by using an RGB-D camera, wherein the image data comprises a color image and a depth image; s2: using an ORB algorithm to extract the feature points of the image data, and using a feature point uniformity evaluation method to judge the distribution uniformity of the feature points; s3: generating point clouds and performing down-sampling on the point clouds for the image data part with uniformly distributed characteristic points; s4: performing plane detection extraction on the point cloud subjected to down-sampling, and eliminating mismatching by using a random sampling consistency algorithm; s5: and for the image data part with non-uniform distribution of the feature points, using a set threshold value to extract the feature points and removing overlapped feature points by a non-maximum value inhibition method. The method can reduce the calculation amount, improve the accuracy of feature point extraction and reduce the mismatching, thereby realizing the requirements of the mobile robot on the accuracy and the real-time performance.

Description

An optimal algorithm based on RGB-D camera combining plane detection and random sampling consensus algorithm ORB algorithm

technical field

The invention is an optimized ORB algorithm based on RGB-D camera combined with plane detection and random sampling consistent algorithm, and belongs to the technical field of indoor mobile robot path planning and navigation.

Background technique

In recent years, intelligent mobile robot technology has developed rapidly and has been widely used in industries such as industry, military, logistics, office and home services. With the advent of RGB-D sensors, the research on mobile robot localization or SLAM using RGB-D sensors has developed rapidly. The advancement of related technologies such as image processing and point cloud processing, as well as the advantages of RGB-D sensors with rich information acquisition, non-contact measurement, easy installation and use, and low cost, make RGB-D sensors widely used in target recognition, tracking, etc. field. The first problem in the robot navigation problem is how to determine the scene model. In the past ten years, many solutions have relied on two-dimensional sensors such as laser and radar for map construction and robot pose estimation. With the advent of RGB-D cameras, more and more researchers have begun to focus on using RGB-D cameras to solve the problem of robotic indoor environment model building, and many influential research results have been produced.

At present, visual SLAM technology has gradually become a mainstream positioning scheme. However, monocular SLAM cannot obtain the depth information of pixels from an image, and it is necessary to estimate the depth of pixels through triangulation or inverse depth methods. Moreover, the depth information estimated by monocular SLAM has scale uncertainty, and with the accumulation of positioning errors, the phenomenon of "scale drift" is prone to occur. The binocular SLAM obtains the matching feature points by matching the images of the left and right cameras, and then estimates the depth information of the feature points according to the parallax method. Binocular SLAM has the advantages of large measurement range, etc., but it requires a large amount of calculation and requires high accuracy of the camera, and usually requires GPU acceleration to meet the real-time requirements. RGB-D camera is a new type of camera emerging in recent years, which can actively obtain the depth information of pixels in the image through physical hardware. Compared with monocular and binocular cameras, RGB-D cameras do not need to consume a lot of computing resources to calculate the depth of pixel points, and can directly measure the surrounding environment and obstacles in 3D, and generate dense points through RGB-D SLAM technology. Cloud maps provide convenience for subsequent navigation planning.

The commonly used RGB-D camera-based feature point extraction and matching methods include ORB algorithm and ICP algorithm. ORB (Oriented FAST and Rotated BRIEF) is a fast feature point extraction and description algorithm. The ORB algorithm is divided into two parts, namely feature point extraction and feature point description. Feature extraction is developed by the FAST (Features from Accelerated Segment Test) algorithm, and the feature point description is improved according to the BRIEF (Binary RobustIndependent Elementary Features) feature description algorithm. The ORB feature combines the detection method of FAST feature points with the Brief feature descriptor, and improves and optimizes them on the basis of their original ones. The biggest feature of the ORB algorithm is the fast calculation speed. This first benefits from the use of FAST to detect feature points. Thirdly, the descriptor is calculated using the BRIEF algorithm. The representation of the descriptor's unique binary string not only saves storage space, but also greatly shortens the matching time. The ICP algorithm was proposed by Besl and McKay 1992, Method for registration of 3-D shapes. The basic principle of the ICP algorithm is: in the target point cloud P and source point cloud Q with matching, respectively, according to certain constraints, find the nearest point (pi, qi), and then calculate the optimal matching parameters R and t, minimize the error function. This method can improve the efficiency of generating the optimal solution, but due to the large amount of calculation, it is less practical for mobile robots, thus increasing the cost.

SUMMARY OF THE INVENTION

In view of this, the purpose of the present invention is to provide an optimized ORB algorithm based on RGB-D camera combined with plane detection and random sampling consensus algorithm, which is used to reduce the amount of calculation, improve the accuracy of feature point extraction, and reduce false matching, so as to achieve The requirements for accuracy and real-time performance of mobile robots.

In order to solve the above-mentioned technical problems, the technical scheme adopted in the present invention is:

An optimized ORB algorithm based on RGB-D camera combined with plane detection and random sampling consensus algorithm, the method includes the following steps:

S1: Use an RGB-D camera to acquire image data, including color images and depth images;

S2: Use the ORB algorithm to extract feature points from the image data, and use the feature point uniformity evaluation method to judge the feature point distribution uniformity;

S3: For the part of the image data where the feature points are evenly distributed, generate a point cloud and downsample it;

S4: Perform plane detection and extraction on the down-sampled point cloud, and use the random sampling consensus algorithm to eliminate false matches;

S5: For the part of the image data with uneven distribution of feature points, use the set threshold for feature point extraction and the non-maximum value suppression method to eliminate overlapping feature points;

S6: For the point cloud after eliminating the mismatch in S4 and the feature point after eliminating the overlapping feature point in S5, project it back to the two-dimensional image plane, reconstruct and equalize the grayscale image.

Preferably, in S1, the RGB-D camera includes a Kinect camera.

Preferably, S2 specifically includes the following steps:

S21: Use the feature point extraction Oriented FAST algorithm to determine whether the feature point x is a feature point, when it is judged that the feature point x is a feature point, calculate the main direction of the feature point, and name the feature point as a key point, so that the key point directional;

The method of judging whether the feature point x is a feature point is: draw a circle with the feature point x as the center. The distance between the feature points x is larger than l _x +t, or is smaller than l _x -t, if such a requirement is met, it is judged that the feature point x is a feature point;

Among them, l _x represents the distance between the feature point x and the pixel point on the circle; l represents the distance; t represents the threshold, which is the adjustment amount of the range; n=16; 9≤m≤12;

S22: Use the feature point to describe the rBRIEF algorithm, perform Gaussian smoothing on the image data, select the pixel point y in the neighborhood with the key point as the center, and form n point pairs (xi, yi), I (x, y) mutual Compare the gray value, I represents the gray value, x>y takes 1, otherwise takes 0, generates an n-dimensional feature descriptor, and defines n point pairs (xi, yi) as a 2*n matrix S,

Use θ to rotate S,

S _θ = R _θ S (2)

In formula (2), S _θ represents the matrix whose rotation angle is θ, and θ is the rotation angle θ along the main direction of the feature point;

Neighborhood is the selection of pixel points y in the circle with the key point as the center passing through k pixels; among them, 0<k<n;

S23 : a method for evaluating the uniformity of feature point distribution. The image data is segmented by different division methods, and after segmentation, the image data part with uniform feature point distribution and the image data part with uneven feature point distribution are obtained.

Further preferably, the method for evaluating the uniformity of feature point distribution is as follows: first, the image data is initially divided into several sub-regions S _i , and each sub-region S _i is divided into several second-level sub-regions S _ij , and the second-level sub-region S _ij includes: S _i1 to S _ij regions, according to the number of feature points in the secondary sub-region S _ij to evaluate whether the feature points in the region are evenly distributed; if the number of feature points from S _i1 to S _ij are all similar, the approximate calculation method is: If the variance value is less than 15, it can be _judged that the distribution of feature points in the sub-region is uniform, otherwise it is judged to be uneven.

Further preferably, the image data segmentation method includes segmentation into center and peripheral directions, segmentation from upper left to lower right, or segmentation from lower left to upper right.

Preferably, S3 specifically includes the following steps:

S31: According to the color image and the depth image with uniform distribution of feature points, use the following formula (3) to obtain the point cloud,

Formula (3) uses the pixel coordinate system ouv; c _X , c _Y , f _X , f _Y , s are the camera internal parameters; u, v are the pixel coordinates of the feature points; d is the depth of the feature points; the coordinates of the feature points are ( X, Y, Z), a number of points defined by coordinates constitute a point cloud;

S32: Process with grid filter to extract plane from point cloud.

Preferably, S4 specifically includes the following steps:

S41: The formula for plane extraction is as follows:

aX+bY+cZ+d=0 (4)

In formula (4), a, b, c represent constants;

S42: Extract planes from the point cloud data with noise using the random sampling consensus method. The extraction judgment condition is: when the number of remaining points in the point cloud is greater than the threshold value g of the total number of point clouds or the extraction plane is smaller than the threshold value h, the feature points are extracted. After these feature points are extracted, the remaining points are extracted again until the threshold h for the number of extraction planes is reached or the number of remaining points is less than the threshold g;

Among them, 20%≤g≤40%; 3≤h≤5.

Preferably, S5 specifically includes the following steps:

S51: Change the threshold t according to the oFAST algorithm of S2, the changed threshold t is t', and reduce the range between l _x +t' and l _x -t', and adjust the range of t' according to the feature point extraction result , determine the changed t' value to make it optimal;

S52: If there are multiple key points in the neighborhood of a certain key point, compare the value J of these feature points, keep the largest value J, and delete the rest, where the value J is defined as follows:

In formula (5), l _xy -l _x and l _x -l _xy both represent the distance between the key point and the known feature point.

Preferably, in S6, the projection formula is:

In formula (6), s is the scale factor. After projection, the grayscale image of each plane is reconstructed. After the grayscale image is equalized by the grayscale histogram, the image is clearer and clearer, and the depth band can be reduced. noise coming.

Beneficial effects of the present invention: whether it is the feature point extraction of a single object or the feature point extraction of multiple objects, the accuracy of the algorithm of the present invention is obviously higher than that of using the ORB algorithm and the ICP algorithm alone, and the parameters such as the calibration error of the camera are not The results are slightly affected, but the accuracy brought by the algorithm of the present invention is still much higher than other algorithms. In addition, because the algorithm of the present invention combines the ORB algorithm, and inserts the uniformity judgment of the distribution of regional feature points and combines the plane extraction and random sampling consistent methods, the running time is slightly longer than that of the ORB algorithm alone, but it does not affect Practicality of the algorithm of the present invention. However, the ICP algorithm has a long running time due to the excessive calculation amount, which weakens its practicability and increases the demand for mobile robot hardware. It can be said that the algorithm of the present invention can greatly improve the accuracy of image feature point extraction by sacrificing negligible extra running time. Therefore, the present invention can improve the accuracy of feature point extraction while keeping less running time, thereby improving the positioning accuracy and real-time performance of the robot.

Description of drawings

Fig. 1 is the flow chart of the present invention;

Fig. 2 is the picture before (left) and after extraction (right) of freiburg1_desk picture among the present invention;

Fig. 3 is the picture before (left) and after extraction (right) of freiburg1_room picture extraction among the present invention;

FIG. 4 is a picture of the freiburg1_teddy picture before (left) and after (right) extraction in the present invention.

Detailed ways

A method for optimizing a prediction model of an SCR denitration system based on machine learning of the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

Example 1

As shown in Figure 1, an optimized ORB algorithm based on RGB-D camera combined with plane detection and random sampling consensus algorithm, the method includes the following steps:

S1: Use an RGB-D camera to acquire image data, including color images and depth images;

S2: Use the ORB algorithm to extract feature points from the image data, and use the feature point uniformity evaluation method to judge the feature point distribution uniformity;

S3: For the part of the image data where the feature points are evenly distributed, generate a point cloud and downsample it;

S4: Perform plane detection and extraction on the down-sampled point cloud, and use the random sampling consensus algorithm to eliminate false matches;

S5: For the part of the image data with uneven distribution of feature points, use the set threshold for feature point extraction and the non-maximum value suppression method to eliminate overlapping feature points;

S6: For the point cloud after eliminating the mismatch in S4 and the feature point after eliminating the overlapping feature point in S5, project it back to the two-dimensional image plane, reconstruct and equalize the grayscale image.

Preferably, in S1, the RGB-D camera includes a Kinect camera.

Preferably, S2 specifically includes the following steps:

S21: Use the feature point extraction Oriented FAST algorithm (ie the oFAST algorithm) to determine whether the feature point x is a feature point, when it is judged that the feature point x is a feature point, calculate the main direction of the feature point, and name the feature point as the key point to make key points (ie, detectors) directional;

The method of judging whether the feature point x is a feature point is: draw a circle with the feature point x as the center, the circle passes through 16 pixels, and whether there are at least 12 consecutive pixels among the 16 pixels on the circumference. The distance between the feature points x is larger than l _x +t, or is smaller than l _x -t, if such a requirement is met, it is judged that the feature point x is a feature point;

Among them, l _x represents the distance between the feature point x and 12 consecutive pixels on the circle; l represents the distance; t represents the threshold, which is the adjustment amount of the range. Setting the t value can filter out the distance close to the x point. For pixel points, if the t value is not set or the t value is too small, there will be too many pixel points that meet the judgment condition, resulting in too many feature points.

S22: Use the feature points to describe the rBRIEF algorithm, perform Gaussian smoothing on the image data, select the pixel point y in the neighborhood with the key point as the center, and form 6 point pairs (xi, yi), the point pairs are randomly selected, I(x, y) compares the gray values with each other, I represents the gray value, x>y takes 1, otherwise takes 0, generates an n-dimensional feature descriptor, and defines n point pairs (xi, yi) as 2*6 matrix S,

Use θ to rotate S,

S _θ = R _θ S (2)

In formula (2), S _θ represents the matrix whose rotation angle is θ, and θ is the rotation angle θ along the main direction of the feature point;

Neighborhood is the selection of pixel points y in the circle with the key point as the center passing through k pixels; among them, 0<k<n;

S23 : a method for evaluating the uniformity of feature point distribution. The image data is segmented by different division methods, and after segmentation, the image data part with uniform feature point distribution and the image data part with uneven feature point distribution are obtained.

Further preferably, the method for evaluating the uniformity of feature point distribution is as follows: first, the image data is initially divided into several sub-regions S _i , and each sub-region S _i is divided into several second-level sub-regions S _ij , and the second-level sub-region S _ij includes: S _i1 to S _ij regions, according to the number of feature points in the secondary sub-region S _ij to _{evaluate whether} the feature points in the region are evenly distributed; If the variance value is less than 15, it can be _judged that the distribution of feature points in the sub-region is uniform, otherwise it is judged to be uneven.

Further preferably, the image data segmentation method includes segmentation into center and peripheral directions, segmentation from upper left to lower right, or segmentation from lower left to upper right.

In this embodiment, the distribution of feature points is judged for each region through image segmentation, so that different algorithms are used according to different distributions.

Preferably, S3 specifically includes the following steps:

S31: According to the color image and the depth image with uniform distribution of feature points, use the following formula (3) to obtain the point cloud,

Formula (3) uses the pixel coordinate system ouv; c _X , c _Y , f _X , f _Y , s are the camera internal parameters; u, v are the pixel coordinates of the feature points; d is the depth of the feature points; the coordinates of the feature points are ( X, Y, Z), a number of points defined by coordinates constitute a point cloud;

S32: Use a grid filter to process, extract a plane from the point cloud, and use a z-direction interval filter to filter out points that are far away. Points with a long distance are too far away from other points, so if they are not filtered out, there will be only one point in the extracted plane, which will increase the number of planes that do not meet the feature points.

In this embodiment, color information (ie, grayscale) is obtained through a color image, and distance information is obtained through a depth image, so that the 3D camera coordinates of the pixels can be calculated, and a point cloud can be generated. This embodiment generates a point cloud from an RGB image and a depth image.

Preferably, S4 specifically includes the following steps:

S41: The formula for plane extraction is as follows:

aX+bY+cZ+d=0 (4)

In formula (4), a, b, c represent constants;

S42: Extract planes from the point cloud data with noise by using the random sampling and consensus method. The extraction judgment condition is: the remaining points of the point cloud are greater than 30% of the total number or the extraction planes are less than 3, and feature points are extracted. When these feature points are extracted , and perform surface extraction on the remaining points again until the number of extracted planes reaches 3 or the number of remaining points is less than 30% of the total number.

Preferably, S5 specifically includes the following steps:

S51: Change the threshold t according to the oFAST algorithm of S2, the changed threshold t is t', and reduce the range between l _x +t' and l _x -t', and adjust the range of t' according to the feature point extraction result , determine the changed t' value to make it optimal;

S52: If there are multiple key points in the neighborhood of a certain key point, compare the value J of these feature points, keep the largest value J, and delete the rest, where the value J is defined as follows:

In formula (5), l _xy -l _x and l _x -l _xy both represent the distance between the key point and the known feature point.

Preferably, in S6, the projection formula is:

In formula (6), s is the scale factor, which can be selected according to the actual situation. After projection, the grayscale image of each plane is reconstructed, and the grayscale image is equalized by the grayscale histogram to make the image more clear and distinct. , and can reduce the noise brought by the depth.

According to the values in Table 1 and Table 2, different algorithms and the feature point extraction accuracy and running time of this algorithm are compared.

As shown in FIGS. 2 to 4 , the present invention extracts feature points for desks, rooms and teddy bears respectively. After extraction, feature points are distributed evenly, representative and accurate, so that subsequent reconstructed images are clearer.

Table 1 is a comparison table of the accuracy of feature point extraction of three pictures of the algorithm of the present invention, the ORB algorithm and the ICP algorithm

Image category freiburg1_desk freiburg1_room freiburg1_teddy ORB algorithm 83.45% 76.09% 85.64% ICP algorithm 83.18% 83.56% 85.08% The algorithm of the invention 93.44% 90.28% 97.21%

Table 2 is a comparison table of the running time of three pictures of the algorithm of the present invention, ORB algorithm and ICP algorithm

Image category freiburg1_desk freiburg1_room freiburg1_teddy ORB algorithm 0.7523 0.9735 0.6285 ICP algorithm 1.0028 1.5236 0.9658 The algorithm of the invention 0.7886 1.0032 0.6422

As can be seen from Table 1, whether it is the feature point extraction of a single object or the feature point extraction of multiple objects, the accuracy of the algorithm of the present invention is significantly higher than that of the ORB algorithm and the ICP algorithm alone, and the parameters such as the calibration error of the camera are There is a slight impact on the results, but the accuracy brought by the algorithm of the present invention is still much higher than that of other algorithms.

As can be seen from Table 2, since the algorithm of the present invention combines the ORB algorithm, and inserts the uniformity judgment of the distribution of regional feature points and combines the consistent method of plane extraction and random sampling, the running time is slightly shorter than that of the ORB algorithm alone. long, but does not affect the practicability of the algorithm of the present invention. However, the ICP algorithm has a long running time due to the excessive calculation amount, which weakens its practicability and increases the demand for mobile robot hardware. It can be said that the algorithm of the present invention can greatly improve the accuracy of image feature point extraction by sacrificing negligible extra running time.

Example 2

The difference between this embodiment and Embodiment 1 is only that: in S21, m=9; in S42, if the number of remaining points of the point cloud is greater than 40% of the total number or the extraction plane is less than 5, feature points are extracted.

Example 3

The difference between this embodiment and Embodiment 1 is only that: in S21, m=10; in S42, feature points are extracted when the number of remaining points of the point cloud is greater than 20% of the total number or the extraction plane is less than 4.

In the description provided herein, numerous specific details are set forth. It will be understood, however, that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "upper", "lower", "front", "rear", "left", "right", etc. are based on those shown in the accompanying drawings The orientation or positional relationship is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the indicated device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention.

Similarly, it is to be understood that in the above description of exemplary embodiments of the invention, various features of the invention are sometimes grouped together into a single embodiment, figure, or its description. This disclosure, however, should not be interpreted as reflecting an intention that the invention as claimed requires more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following the Detailed Description are hereby expressly incorporated into this Detailed Description, with each claim standing on its own as a separate embodiment of this invention.

As used herein, unless otherwise specified, the use of the ordinal numbers "first," "second," "third," etc. to describe common objects merely refers to different instances of similar objects, and is not intended to imply such The objects being described must have a given order in time, space, ordinal, or in any other way.

While the invention has been described in terms of a limited number of embodiments, those skilled in the art will appreciate, having the benefit of the above description, that other embodiments are conceivable within the scope of the invention thus described. Furthermore, it should be noted that the language used in this specification has been principally selected for readability and teaching purposes, rather than to explain or define the subject matter of the invention. Accordingly, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. This disclosure is intended to be illustrative, not restrictive, as to the scope of the present invention, which is defined by the appended claims.

The above is only the preferred embodiment of the present invention, it should be pointed out: for those skilled in the art, under the premise of not departing from the principle of the present invention, several improvements and modifications can also be made, and these improvements and modifications are also It should be regarded as the protection scope of the present invention.

Claims (8)

1. an optimized ORB algorithm based on RGB-D camera in conjunction with plane detection and random sampling consistent algorithm, is characterized in that, the method comprises the following steps: S1: Use an RGB-D camera to acquire image data, including color images and depth images; S2: Use the ORB algorithm to extract feature points from the image data, and use the feature point uniformity evaluation method to judge the feature point distribution uniformity; S3: For the part of the image data with uniform distribution of feature points, generate a point cloud and downsample it; S4: Perform plane detection and extraction on the down-sampled point cloud, and use a random sampling consensus algorithm to eliminate false matches; S5: For the part of the image data with uneven distribution of feature points, use the set threshold for feature point extraction and the non-maximum value suppression method to eliminate overlapping feature points; S6: For the point cloud after eliminating the mismatch in S4 and the feature point after eliminating the overlapping feature point in S5, project it back to the two-dimensional image plane, reconstruct and equalize the grayscale image. 2 . An optimized ORB algorithm based on RGB-D camera combining plane detection and random sampling consensus algorithm according to claim 1 , wherein, in S1 , the RGB-D camera includes a Kinect camera. 3 . 3. a kind of optimized ORB algorithm based on RGB-D camera combining plane detection and random sampling consistent algorithm according to claim 1, is characterized in that, S2 specifically comprises the following steps: S21: Use the feature point extraction Oriented FAST algorithm to determine whether the feature point x is a feature point, when it is judged that the feature point x is a feature point, calculate the main direction of the feature point, and name the feature point as a key point, so that the key point directional; The method of judging whether the feature point x is a feature point is: draw a circle with the feature point x as the center. The distance between the feature points x is larger than l _x +t, or smaller than l _x -t. If this requirement is met, it is judged that the feature point x is a feature point; among them, l _x represents the feature point x and the The distance between the pixels on the circle; l represents the distance; t represents the threshold, which is the adjustment amount of the range; n=16; 9≤m≤12; S22: Use the feature point to describe the rBRIEF algorithm, perform Gaussian smoothing on the image data, select the pixel point y in the neighborhood with the key point as the center, and form n point pairs (xi, yi), I (x, y) mutual Compare the gray value, I represents the gray value, x>y takes 1, otherwise takes 0, generates an n-dimensional feature descriptor, and defines n point pairs (xi, yi) as a 2*n matrix S,

Use θ to rotate S, S _θ = R _θ S (2) In formula (2), S _θ represents the matrix whose rotation angle is θ, and θ is the rotation angle θ along the main direction of the feature point; Neighborhood is the selection of pixel points y in the circle with the key point as the center passing through k pixels; among them, 0<k<n; S23 : a method for evaluating the uniformity of feature point distribution. The image data is segmented by different division methods, and after segmentation, the image data part with uniform feature point distribution and the image data part with uneven feature point distribution are obtained. 4. a kind of optimized ORB algorithm based on RGB-D camera combined with plane detection and random sampling consistent algorithm according to claim 3, it is characterized in that, feature point distribution uniformity evaluation method is: at first the image data is preliminarily divided into several Sub-region S _i , each sub-region S _i is divided into several secondary sub-regions S _ij again, the second-level sub-region S _ij includes S _i1 to S _ij regions, and the evaluation is based on the number of feature points in the second-level sub-region S _ij Whether the feature points in this area are evenly distributed; if the number of feature points from S _i1 to S _ij is similar, the similar calculation method is: by calculating the variance value of the feature points of the statistical distribution of the secondary sub-region, and judging according to the variance value, when the variance value When it is less than 15, it can be judged that the distribution of feature points in the _Si sub-region is uniform; otherwise, it can be judged that it is not uniform. 5. a kind of optimized ORB algorithm based on RGB-D camera combined with plane detection and random sampling consistent algorithm according to claim 3, is characterized in that, image data segmentation method comprises dividing into center and surrounding directions, from upper left to lower right Split or split from bottom left to top right. 6. a kind of optimized ORB algorithm based on RGB-D camera combining plane detection and random sampling consistent algorithm according to claim 5, is characterized in that, S3 specifically comprises the following steps: S31: According to the color image and the depth image with uniform distribution of feature points, use the following formula (3) to obtain the point cloud,

Formula (3) uses the pixel coordinate system ouv; c _X , c _Y , f _X , f _Y , s are the camera internal parameters; u, v are the pixel coordinates of the feature points; d is the depth of the feature points; the coordinates of the feature points are ( X, Y, Z), a number of points defined by coordinates constitute a point cloud; S32: Process with grid filter to extract plane from point cloud. 7. a kind of optimized ORB algorithm based on RGB-D camera combining plane detection and random sampling consistent algorithm according to claim 6, is characterized in that, S4 specifically comprises the following steps: S41: The formula for plane extraction is as follows: aX+bY+cZ+d=0 (4) In formula (4), a, b, c represent constants; S42: Extract planes from the point cloud data with noise using the random sampling consensus method. The extraction judgment condition is: when the number of remaining points in the point cloud is greater than the threshold value g of the total number of point clouds or the extraction plane is smaller than the threshold value h, the feature points are extracted. After these feature points are obtained, the remaining points are extracted again until the threshold h of the number of extracted planes is reached or the number of remaining points is less than the threshold g; Among them, 20%≤g≤40%; 3≤h≤5. 8. a kind of optimized ORB algorithm based on RGB-D camera combining plane detection and random sampling consistent algorithm according to claim 7, is characterized in that, S5 specifically comprises the following steps: S51: Change the threshold t according to the oFAST algorithm of S2, the changed threshold t is t', and reduce the range between l _x +t' and l _x -t', and adjust the range of t' according to the feature point extraction result , determine the changed t' value to make it optimal; S52: If there are multiple key points in the neighborhood of a certain key point, compare the value J of these feature points, keep the largest value J, and delete the rest, where the value J is defined as follows:

In formula (5), l _xy -l _x and l _x -l _xy both represent the distance between the key point and the known feature point. An optimized ORB algorithm based on RGB-D camera combined with plane detection and random sampling consistent algorithm according to claim 1, is characterized in that, in S6, the projection formula is:

In formula (6), s is the scale factor. After projection, the grayscale image of each plane is reconstructed. After the grayscale image is equalized by the grayscale histogram, the image is clearer and clearer, and the depth band can be reduced. noise coming.

Images