CN107403451B - Adaptive binary feature monocular visual odometry method and computer and robot - Google Patents

Abstract

The invention belongs to the technical field of computer vision, and discloses an adaptive binary feature monocular visual odometry method, a computer and a robot. The feature detection threshold is adaptively adjusted according to the illumination changes in the environment, and is stable under different illumination conditions. A constant number of features is detected; at the same time, the detected features are organized by constructing a quad-tree, and the features are randomly selected by a combination of depth-first search and breadth-first search, so that the final output features are evenly distributed on the entire image. The final output feature of the invention is evenly distributed on the whole image, which reduces the influence of the abnormal value of the feature on the dynamic target on the pose estimation accuracy.

Description

Adaptive binary feature monocular visual odometry method and computer and robot

technical field

The invention belongs to the technical field of computer vision, and in particular relates to an adaptive binary feature monocular visual odometry method, a computer and a robot.

Background technique

In computer vision, Visual Odometry (VO) refers to the use of input video frames from single or multiple cameras to estimate the position and orientation of a mobile robot. It is not only one of the core technologies of automatic navigation of mobile robots, but also has huge application prospects in the fields of emerging wearable device computing, 3D reconstruction, virtual reality, augmented reality and autonomous driving. The visual odometry method can be roughly divided into the following three categories according to the type and number of cameras used: the first category is the visual odometry method based on the binocular camera, the outstanding advantage is that a certain The depth information of the camera at the moment can be directly calculated from the two images of different viewpoints of the two images through the known disparity relationship between the two images, so that the actual location of the map point in the scene can be easily known; there are some The disadvantage is that the installation and configuration of the binocular cameras are inconvenient, and the switching between the large-scale outdoor environment and the small-scale indoor environment caused by the fixed parallax between the binocular cameras is inconvenient. The second type is the visual odometry method based on RGB-D camera. In addition to the image of the scene, the RGB-D camera also integrates other sensors to directly obtain the dense depth map of the scene. The representative of this type of camera is the Kinect depth map. The disadvantage of the camera is that the hardware cost is high, and the effective depth distance detected by the sensor is too short, so the applicable environment is limited. The third category is the visual odometry method based on a monocular camera, which can perform localization and map construction using only a single camera. However, the monocular camera can only obtain one image of the scene at each moment, and the acquisition of depth information of map points in the scene needs to be calculated through the correlation between multiple images, so the acquisition of depth information is the monocular visual mileage. a difficulty in the calculation. Since the calculation of depth information in monocular vision is obtained by the correspondence between images of different viewpoints, there is a scale factor between the calculated depth value of the map point and the actual depth value, and the uncertainty of depth makes monocular vision The odometry method can freely switch between large-scale outdoor environments and small-scale indoor environments. Therefore, the monocular visual odometry method has the advantages of low hardware cost and wide range of applicable environments. When the environmental conditions in the camera motion scene are good, the existing visual odometry methods estimate the camera motion pose trajectory, which can achieve better real-time performance and robust performance. However, the existing visual odometry methods still have the following shortcomings: when the ambient light in the camera motion scene is too bright or too dark, the existing visual odometry methods are often unable to capture video frames under such lighting conditions. Sufficient features are extracted, so that the camera pose trajectory estimation is interrupted, or the accuracy of pose estimation is greatly reduced due to the excessive proportion of outliers; when there are dynamic moving objects in the camera motion scene, the detected features on the dynamic objects will be Causes pose estimation errors; when the frame rate of the captured video frame of the camera and the frame rate of the video processed by the visual odometry method are not synchronized, frame skipping will occur, which can easily lead to the interruption of the trajectory estimation of the visual odometry method.

To sum up, the problems existing in the prior art are: the existing visual odometry methods cannot extract enough features, and the camera pose trajectory estimation is interrupted or the accuracy of pose estimation is greatly reduced due to the excessive proportion of abnormal points; When there is a dynamic moving target in the camera motion scene, the detected features on the dynamic target will cause pose estimation errors; when the frame rate of the captured video frame of the camera and the frame rate of the video processed by the visual odometry method are not synchronized, frame skipping will occur. phenomenon, leading to the interruption of trajectory estimation by the visual odometry method.

SUMMARY OF THE INVENTION

Aiming at the problems existing in the prior art, the present invention provides an adaptive binary feature monocular visual odometry method, a computer and a robot.

The present invention is implemented in this way, an adaptive binary feature monocular visual odometry method, wherein the adaptive binary feature monocular visual odometry method adaptively adjusts the feature detection threshold according to the illumination changes in the environment, A constant number of features is stably detected under different lighting conditions; at the same time, the detected features are organized by constructing a quad-tree, and the features are randomly selected by a combination of depth-first search and breadth-first search, so that the final output features are evenly distributed. on the entire image.

Further, the adaptive binary feature monocular visual odometry method includes the following steps:

In step 1, two processes are created, denoted as Process I and Process II respectively; the process Process I is used for the video frame acquisition of the camera, and the process Process II is used for the visual odometry to estimate the camera pose;

Step 2, open up a shared memory area for data communication between Process I and Process II; record this shared memory area as a frame sequence pool, and maintain a first-in, first-out video frame queue on the frame sequence pool;

Step 3, open camera in process Process 1, collect video frame; If the resolution of video frame is too large, reduce resolution to video frame according to proportional relationship, video frame is pressed into frame sequence pool;

Step 4, the process Process II sequentially fetches the video frame images from the shared memory pool;

Step 5: Calculate the histogram statistics of the grayscale image. The width of each histogram block is 16 pixels; divide the number of pixels of each histogram block by the total number of images to obtain each histogram block. the statistical probability of the block;

Step 6: Set different corner detection thresholds according to different image entropy sizes, and detect different numbers of feature points;

In step 7, the Gaussian average operator is used to filter the image, filter out noise, divide the whole image into small image blocks, and use the threshold value adaptively set according to the environment to separate the small image blocks through the AGAST corner detection algorithm. corner feature detection;

Step 8: Use the non-maximum suppression algorithm to obtain the AGAST corner points with the maximum response in the local image block;

Step 9, if after the non-maximum value is suppressed, the number of feature points is also greater than the specified number of feature points to be selected, randomly and uniformly select the specified number of feature point sets;

Step 10: Calculate the direction of the AGAST corner feature;

Step 11: Generate binary feature descriptors through the rBRIEF algorithm to obtain adaptive uniform binary features detected on the image;

Step 12: Search for a matching point set corresponding to the adaptive uniform binary feature on the current frame and the previous frame through the Hamming distance and the directional distance difference between the adaptive uniform binary features;

Step 13: Calculate the pose matrix of the camera through the searched corresponding sets of matching feature points between the front and rear frames.

Further, the calculation formula of video frame resolution reduction in the step 3 is:

Among them, (w, h) is the size of the original image, and (w', h') represents the size of the image after downscaling.

Further, in the described step 4, the process Process II takes out the video frame image in sequence from the shared memory pool; the image is a color image, and is converted into a grayscale image, which is denoted as I _gray , and the adaptive uniform binary feature detection is performed to the image I _gray . Algorithm to detect adaptive uniform binary features on I _gray .

Further, in the described step 5, calculate the image entropy Entropy (I _gray ) of I _gray , calculate the histogram statistics of grayscale image, and the width of each histogram block is 16 pixels; Utilize the pixel of each histogram block Divide the number by the total number of images to obtain the statistical probability p _i of each histogram block, and the image entropy can be calculated according to the following formula:

Further, the method for randomly and uniformly selecting a specified number of feature points in the step 9 specifically includes:

All feature points on the image are organized into a quadtree, and the search algorithm combining breadth-first search and depth-first search is used to randomly and uniformly select feature points, and the image is recursively divided into upper left and The lower left, upper right and lower right four blocks, until there are no feature points in the image block; use the BFS algorithm to randomly search and select the nodes on the quadtree, if a node is a leaf node and the node has only one feature point , then select the feature point; otherwise, continue to search for other nodes randomly; if a node is not a leaf node, you need to select a feature point in the image area represented by the node, and randomly find a leaf containing the feature point that was not selected before. Nodes; iterate until the specified number of feature points is selected.

Further, the direction θ of the AGAST corner feature calculated in the tenth step, the calculation formula is as follows:

Among them, M _ij represents the geometric moment of the image block where the feature is located, which is calculated according to M _ij =∑ _{x, y} x ⁱ y ^j ·I(x, y); x _c and y _c represent the centroid of the image block where the feature is located, according to x _c = M ₁₀ /M ₀₀ and the formula y _c =M ₀₁ /M ₀₀ are calculated; u′ _ij represents the central moment of the image block where the feature is located, and is calculated from the geometric moment Mi _ij of the image block and the centroids x _c , y _c .

Further, the step 12 specifically includes:

First, set the size of the search window, each feature point will find a matching point in the neighborhood window where it is located;

Then, find all the test feature point sets {p′ _i } in the neighborhood window where the query feature point p _i is located, calculate the Hamming distance between p _i and all feature point descriptors in {p′ _i }, and write down The smallest Hamming distance and the next smallest Hamming distance, and their corresponding test feature points; when the smallest Hamming distance is less than the specified threshold t _m , and the smallest Hamming distance and the next smallest Hamming distance If the ratio is less than a certain ratio _rm , then the corresponding test feature point is the feature point that matches the query feature point;

Finally, use the Delaunay triangle mesh generation algorithm to generate a triangle mesh at the matching points of the current frame, and remove those triangles in the triangle mesh whose area is significantly larger than most meshes, as well as the vertices on the triangle that are far away from most meshes; calculate The displacement value of each triangle pixel vertex in the triangular mesh relative to the matching point on the previous frame, when the displacement value is greater than a certain threshold, the matching point is eliminated; the matching corresponding point set between the two frames before and after is finally found.

Another object of the present invention is to provide a computer applying the adaptive binary feature monocular visual odometry method.

Another object of the present invention is to provide a robot applying the adaptive binary feature monocular visual odometry method.

The advantages and positive effects of the present invention are: the feature detection threshold can be adaptively adjusted according to the illumination changes in the environment, and a constant number of features can be stably detected under different illumination conditions. It can be seen from (b) of Fig. 11 that in the case of severe illumination changes in the camera motion trajectory scene in (a), the adaptive uniform binary feature proposed by this method can be stably detected compared to the ORB feature. A specified number of features, and the real-time performance of both remains on the same order of magnitude. In addition, the detected features are organized by constructing a quad-tree, and the features are randomly selected by a combination of depth-first search and breadth-first search, so that the final output features are evenly distributed on the entire image, reducing the features on dynamic targets. The effect of outliers on the accuracy of pose estimation. It can be seen from Figure 10 that the fitting accuracy between the camera motion trajectory estimated by the monocular visual odometry method based on adaptive uniform binary features and the actual camera motion trajectory proposed by this method is better than that based on ORB features. The camera motion trajectory estimated by the monocular visual odometry method is high.

Description of drawings

FIG. 1 is a flowchart of a method for adaptive binary feature monocular visual odometry provided by an embodiment of the present invention.

FIG. 2 is a flowchart of the implementation of the method for self-adaptive binary feature monocular visual odometry provided by an embodiment of the present invention.

FIG. 3 is a schematic diagram of a memory organization form of a frame sequence pool provided by an embodiment of the present invention.

FIG. 4 is a flowchart of an adaptive uniform binary feature detection algorithm provided by an embodiment of the present invention.

FIG. 5 is a schematic diagram of the relationship between image contrast and image entropy provided by an embodiment of the present invention.

FIG. 6 is a schematic diagram of an AGAST corner detection result after the image is not segmented (a) and the image is segmented (b) according to an embodiment of the present invention.

FIG. 7 is a schematic diagram of a result of random and uniform selection of feature points according to an embodiment of the present invention.

FIG. 8 is a schematic diagram of an adaptive uniform binary feature detection result provided by an embodiment of the present invention.

FIG. 9 is a schematic diagram of a corresponding feature point set for searching and matching between two frames according to an embodiment of the present invention.

10 is a schematic diagram showing the comparison of the pose and trajectory calculation results of the monocular visual odometry method based on adaptive uniform binary features and ORB features provided by an embodiment of the present invention;

Among them, (a) represents the actual trajectory of the camera motion; (b) the solid line represents the monocular visual odometry method based on adaptive uniform binary features, and the dotted line represents the monocular visual odometry method based on ORB features. It can be seen from (b) that the camera motion trajectory calculated by the monocular visual odometry method based on adaptive uniform binary features is more accurate in fitting the actual camera motion trajectory in (a) than based on the ORB feature. The monocular visual odometry method.

FIG. 11 is a performance comparison diagram between an adaptive uniform binary feature and an ORB feature provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

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

As shown in FIG. 1 , the adaptive binary feature monocular visual odometry method provided by the embodiment of the present invention includes the following steps:

S101: Create two processes, denoted as Process I and Process II respectively; wherein, the process Process I is used for video frame acquisition of the camera, and the process Process II is used for visual odometry to estimate the camera pose;

S102: Open up a shared memory area for data communication between Process I and Process II. Remember this shared memory area as the frame sequence pool, and maintain a first-in, first-out video frame queue on the frame sequence pool;

S103: Turn on the camera in the process Process I to capture video frames. If the resolution of the video frame is too large, reduce the resolution of the video frame according to a certain proportional relationship, and then push the video frame into the frame sequence pool;

S104: The process Process II sequentially fetches video frame images from the shared memory pool;

S105: Calculate the histogram statistics of the grayscale image, the width of each histogram block is 16 pixels; divide the number of pixels of each histogram block by the total number of images to obtain each histogram block The statistical probability of ;

S106: Set different corner detection thresholds according to different image entropy sizes to detect different numbers of feature points;

S107: Use the Gaussian average operator to filter the image, filter out noise, divide the whole image into small image blocks, and use the threshold value adaptively set according to the environment to separate the small image blocks through the AGAST corner detection algorithm. corner feature detection;

S108: Using the non-maximum value suppression algorithm, obtain the AGAST corner point of the maximum value maximum response in the local image block;

S109: If the number of feature points is also greater than the specified number of feature points to be selected after the non-maximum value is suppressed, then randomly and uniformly select the specified number of feature point sets;

S110: the direction of the calculated AGAST corner feature;

S111: Generate a binary feature descriptor through the rBRIEF algorithm to obtain an adaptive uniform binary feature detected on the image;

S112: Search for a matching point set corresponding to the adaptive uniform binary feature on the current frame and the previous frame by using the Hamming distance and the directional distance difference between the adaptive uniform binary features;

S113: Calculate the pose matrix of the camera through the searched corresponding sets of matching feature points between the front and rear frames.

The application principle of the present invention will be further described below with reference to the accompanying drawings.

As shown in FIG. 2 , the adaptive binary feature monocular visual odometry method provided by the embodiment of the present invention specifically includes the following steps:

Step 1: Create two processes, denoted as Process I and Process II. Among them, Process I is used for camera video frame acquisition, and Process II is used for visual odometry to estimate camera pose.

Step 2: Open up a shared memory area for data communication between Process I and Process II. Remember this shared memory area as the frame sequence pool. A first-in, first-out video frame queue is maintained on the frame sequence pool. The memory organization of the frame sequence pool is shown in Figure 3.

Among them, some information such as flag bits, queue head and tail positions, and image size are reserved in the frame sequence pool information header. b indicates whether there is an image in the queue, s indicates the position of the queue head, e indicates the position of the queue tail, w indicates the width of the image, h indicates the height of the image, and c indicates the number of channels of the image.

Step 3: Open the camera in Process I to capture video frames. If the resolution of the video frame is too large, reduce the resolution of the video frame according to a certain proportional relationship, and then push the video frame into the frame sequence pool. The formula for calculating the video frame resolution reduction is as follows:

Among them, (w, h) is the size of the original image, and (w', h') represents the size of the image after downscaling.

Step 4: Process Process II sequentially fetches video frame images from the shared memory pool. If the image is a color image, convert it to a grayscale image, denoted as I _gray . First, the adaptive uniform binary feature detection algorithm is performed on the image I _gray to detect the adaptive uniform binary feature on the I _gray , as shown in step 5 to step 11. The flowchart of the adaptive uniform binary feature detection algorithm is shown in FIG. 4 . Then, search for the matching corresponding feature point set between the current frame and the previous frame, as in step 12. Finally, the pose of the camera is calculated by matching the corresponding point sets between the front and rear frames, as in step 13.

Step 5: Calculate the image entropy Entropy (I _gray ) of I _gray . First calculate the histogram statistics of the grayscale image, each histogram bin is 16 pixels wide. Then divide the number of pixels of each histogram block by the total number of images to obtain the statistical probability p _i of each histogram block. Then, the image entropy can be calculated according to the following formula:

Step 6: Set different corner detection thresholds according to different image entropy values to detect different numbers of feature points. An image with a larger entropy indicates that the contrast of the image is higher, which reflects the good lighting conditions of the environment. At this time, a higher corner detection threshold is set. An image with smaller entropy indicates that the contrast of the image is smaller, reflecting that the ambient light is too bright or too dark, or the image is too dark or too bright caused by the exposure of the camera, and a lower corner detection threshold is set at this time. Since 16 histogram bins are divided, the possible interval for the entropy value is [0, 4]. The relationship between image entropy and contrast is shown in Figure 5.

The image entropy of the video frame collected by the camera is calculated in the actual scene environment, and it is found that the contrast of the image is very small when the entropy value is [0, 2.7]. The present invention sets the corner detection threshold at this time at 10, when the entropy value is When the image contrast is good at (2.7, 3.0], the present invention sets the corner detection threshold at 30, and when the entropy value is at (3.0, 4], the corner detection threshold is set at 40. It is also possible to divide more There is a small image entropy value interval and the corresponding relationship between the threshold value, but it is not necessary to establish a continuous functional relationship between the image entropy value and the feature detection threshold value, because the difference between the threshold value by two or three pixels does not greatly affect the results of AGAST corner detection.

Step 7: Use a 3×3 Gaussian average operator to filter the image I _gray to filter out noise. Divide the whole image into small image blocks, and use the threshold value adaptively set according to the environment in step 6, and perform corner feature detection on these small image blocks through the AGAST corner detection algorithm.

When AGAST corner detection is performed on the entire image, the detected corners are sometimes clustered in the image. When the image is not divided into blocks, the detected corners are mainly concentrated on the left and right trees, as shown in Figure 6(a) . When the entire image is divided into small image blocks, and corner detection is performed on these image blocks, each area can have a suitable number of corner points, which can offset the corner points on the dynamic target in the image to a certain extent. and other outliers.

The number of AGAST feature points detected after the image is divided into blocks is more, mainly because the AGAST corner detection algorithm uses a binary decision tree to determine the final corner point. After the image is divided into blocks, there is a binary decision on each small block. Tree, the corresponding feature points detected on the entire image are more. At the same time, features are detected on each region after segmentation, as shown in Figure 6(b).

It is recommended that the size of the image block should not be too small, which will cause too many image blocks for image segmentation and increase the amount of calculation. The size of the image block selected by the present invention in the experimental test on the image is 50 pixels, because the feature calculation time is the closest to the feature calculation time for the entire image, and the feature points are basically distributed on the entire image area.

Step 8: Use the non-maximum suppression algorithm to obtain the AGAST corner points with the maximum response in the local image block.

Step 9: If the number of feature points is also greater than the specified number of feature points to be selected after the non-maximum value suppression in step 8, then randomly and uniformly select the specified number of feature point sets, and the random selection result is shown in Figure 7. The method of randomly and uniformly selecting a specified number of feature points is as follows:

All feature points on the image are organized into a quad-tree, and a search algorithm combining breadth first search (BFS) and depth first search (DFS) is used to randomly and uniformly select feature points. First, the image is recursively divided into upper left, lower left, upper right and lower right four blocks according to the central axis of length and width, until there are no feature points in the image block; then the BFS algorithm is used to randomly search and select the nodes on the quadtree , if a node is a leaf node and the node has only one feature point, then select the feature point. Otherwise, continue to use the BFS algorithm to randomly search for other nodes. If a node is not a leaf node, but a feature point needs to be selected in the image area represented by the node, then the DFS algorithm is used to randomly find a leaf node that contains a feature point that was not selected before. Iterate the above process until the specified number of feature points is selected.

Step 10: Calculate the direction θ of the AGAST corner feature obtained in Step 9, and the calculation formula is as follows:

Wherein, M _ij represents the geometric moment of the image block where the feature is located, and is calculated according to the formula M _ij =∑ _{x, y} x ⁱ y ^j ·I(x, y). x _c and y _c represent the centroid of the image block where the feature is located, and are calculated according to the formula x _c =M ₁₀ /M ₀₀ and the formula y _c =M ₀₁ /M ₀₀ .

Step 11: Generate a binary feature descriptor through the rBRIEF algorithm to obtain an adaptive uniform binary feature detected on I _gray , as shown in Figure 8.

Step 12: Search for a matching point set corresponding to the adaptive uniform binary feature on the current frame and the previous frame by using the Hamming distance and the directional distance difference between the adaptive uniform binary features, as shown in Figure 9.

First, set the size of the search window, each feature point will find matching points within its neighborhood window.

Then, find all the test feature point sets {p′ _i } in the neighborhood window where the query feature point p _i is located. Calculate the Hamming distance between p _i and all feature point descriptors in {p' _i }, note the smallest Hamming distance and the next smallest Hamming distance, and their corresponding test feature points. When the smallest Hamming distance is less than the specified threshold _{t m} , and the ratio between the smallest Hamming distance and the next smallest Hamming distance is less than a certain ratio rm , then the corresponding test feature point is the same as the query feature point. matching feature points. If the matching point of _pi is not found through the Hamming distance, then use the direction of the feature to find it again in the same way. Among them, the threshold _{t m} and the ratio rm are set according to the image entropy Entropy (I _gray ). When the image entropy is small, the threshold _{t m} and the ratio rm are set larger. When the image entropy is larger, the threshold _{t m} and the ratio rm are set smaller.

Finally, the Delaunay triangle mesh generation algorithm is used to generate a triangle mesh at the matching points of the current frame, and those triangles in the triangle mesh whose area is significantly larger than most meshes are culled, as well as the vertices on the triangle that are far away from most meshes. Then, the displacement value of each triangle pixel vertex in the triangular mesh relative to the matching point on the previous frame is calculated, and when the displacement value is greater than a certain threshold, the matching point is eliminated. The matching corresponding point set between the two frames before and after is finally found.

Step 13: Calculate the pose matrix of the camera through the corresponding sets of matching feature points between the front and rear frames searched in Step 12. The calculation of the camera pose is divided into two cases: when the system is just initialized, the camera pose is calculated by the eight-point algorithm based on RANSAC outlier rejection based on the corresponding set of matching feature points between the front and rear frames, and the triangulation reconstruction is performed. The 3D scene structure points corresponding to the feature points are obtained; when the system is initialized, the relationship between the reconstructed 3D scene structure points and the feature points on the current frame is found, and the PnP algorithm based on algebraic outlier rejection is used to calculate the pose of the camera. And the 3D scene structure points corresponding to the feature points are reconstructed by triangulation.

The comparison results of the pose calculation results of the monocular visual odometry method based on the adaptive uniform binary feature and the monocular visual odometry method based on the ORB feature are shown in Figure 10.

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 shall be included in the protection of the present invention. within the range.

Claims (9)

1. an adaptive binary feature monocular visual odometry method, is characterized in that, described adaptive binary feature monocular visual odometry method adaptively adjusts the feature detection threshold according to the illumination change situation in the environment, in different A constant number of features is stably detected under the illumination conditions of 10000000; meanwhile, the detected features are organized by constructing a quad-tree, and the features are randomly selected by a combination of depth-first search and breadth-first search, so that the final output features are evenly distributed in the on the entire image; The self-adaptive binary feature monocular visual odometry method includes the following steps: In step 1, two processes are created, denoted as Process I and Process II respectively; the process Process I is used for the video frame acquisition of the camera, and the process Process II is used for the visual odometry to estimate the camera pose; Step 2, open up a shared memory area for data communication between Process I and Process II; record this shared memory area as a frame sequence pool, and maintain a first-in, first-out video frame queue on the frame sequence pool; Step 3, open camera in process Process 1, collect video frame; If the resolution of video frame is too large, reduce resolution to video frame according to proportional relationship, video frame is pressed into frame sequence pool; Step 4, the process Process II sequentially takes out the video frame images from the shared memory pool; Step 5: Calculate the histogram statistics of the grayscale image. The width of each histogram block is 16 pixels; divide the number of pixels of each histogram block by the total number of images to obtain each histogram block. the statistical probability of the block; Step 6: Set different corner detection thresholds according to different image entropy sizes, and detect different numbers of feature points; In step 7, the Gaussian average operator is used to filter the image, filter out noise, divide the whole image into small image blocks, and use the threshold value adaptively set according to the environment to separate the small image blocks through the AGAST corner detection algorithm. corner feature detection; Step 8: Use the non-maximum suppression algorithm to obtain the AGAST corner points with the maximum response in the local image block; Step 9, if after the non-maximum value is suppressed, the number of feature points is also greater than the specified number of feature points to be selected, randomly and uniformly select the specified number of feature point sets; Step 10: Calculate the direction of the AGAST corner feature; Step 11: Generate binary feature descriptors through the rBRIEF algorithm to obtain adaptive uniform binary features detected on the image; Step 12: Search for a matching point set corresponding to the adaptive uniform binary feature on the current frame and the previous frame through the Hamming distance and the directional distance difference between the adaptive uniform binary features; Step 13: Calculate the pose matrix of the camera through the searched corresponding sets of matching feature points between the front and rear frames. 2. adaptive binary feature monocular visual odometry method as claimed in claim 1, is characterized in that, in described step 3, the calculation formula of video frame reduction resolution is:

Among them, (w, h) is the size of the original image, and (w', h') represents the size of the image after downscaling. 3. adaptive binary feature monocular visual odometry method as claimed in claim 1, is characterized in that, in described step 4, process ProcessII takes out video frame image sequentially from shared memory pool; Image is color image, converts is a grayscale image, denoted as I _gray , and an adaptive uniform binary feature detection algorithm is performed on the image I _gray to detect the adaptive uniform binary features on I _gray . 4. adaptive binary feature monocular visual odometry method as claimed in claim 1, is characterized in that, in described step 5, calculate the image entropy Entropy (I _gray ) of 1 _gray , calculate the histogram statistics of grayscale image , the width of each histogram block is 16 pixels; divide the number of pixels of each histogram block by the total number of images to obtain the statistical probability p _i of each histogram block, and the image entropy can be Calculate according to the following formula:

5. The self-adaptive binary feature monocular visual odometry method as claimed in claim 1, wherein the method for randomly and uniformly selecting a specified number of feature points in the step 9 specifically comprises: All feature points on the image are organized into a quadtree, and the search algorithm combining breadth-first search and depth-first search is used to randomly and uniformly select feature points, and the image is recursively divided into upper left and The lower left, upper right and lower right four blocks, until there are no feature points in the image block; use the BFS algorithm to randomly search and select the nodes on the quadtree, if a node is a leaf node and the node has only one feature point , then select the feature point; otherwise, continue to search for other nodes randomly; if a node is not a leaf node, you need to select a feature point in the image area represented by the node, and randomly find a leaf containing the feature point that was not selected before. Nodes; iterate until the specified number of feature points is selected. 6. adaptive binary feature monocular visual odometry method as claimed in claim 1, is characterized in that, the direction θ of the AGAST corner feature that calculates in described step ten, calculation formula is as follows:

Among them, M _ij represents the geometric moment of the image block where the feature is located, which is calculated according to M _ij =∑ _{x, y} x ⁱ y ^j ·I(x, y); x _c and y _c represent the centroid of the image block where the feature is located, according to x _c = M ₁₀ /M ₀₀ and the formula y _c =M ₀₁ /M ₀₀ is calculated. 7. The adaptive binary feature monocular visual odometry method as claimed in claim 1, wherein the step 12 specifically comprises: First, set the size of the search window, each feature point will find a matching point in the neighborhood window where it is located; Then, find all the test feature point sets {p′ _i } in the neighborhood window where the query feature point p _i is located, calculate the Hamming distance between px and all feature point descriptors in {p′ _i }, and record the minimum The Hamming distance and the next smallest Hamming distance of , and their corresponding test feature points; when the smallest Hamming distance is less than the specified threshold t _m , and the ratio between the smallest Hamming distance and the next smallest Hamming distance is less than a certain ratio r _m , then the corresponding test feature point is the feature point that matches the query feature point; Finally, use the Delaunay triangle mesh generation algorithm to generate a triangle mesh at the matching points of the current frame, and remove those triangles in the triangle mesh whose area is significantly larger than most meshes, as well as the vertices on the triangle that are far away from most meshes; calculate The displacement value of each triangle pixel vertex in the triangular mesh relative to the matching point on the previous frame, when the displacement value is greater than a certain threshold, the matching point is eliminated; the matching corresponding point set between the two frames before and after is finally found. 8. A computer applying the self-adaptive binary feature monocular visual odometry method according to any one of claims 1 to 7. 9. A robot applying the self-adaptive binary feature monocular visual odometry method according to any one of claims 1 to 7.

Images