A Feature Point Matching and Screening Method Based on Adaptive Region Motion Statistics

Abstract

The present invention discloses a method for feature point matching and screening based on adaptive region motion statistics, belonging to the technical field of computer vision. The method of the present invention uses the method of dividing the maximum stable extremal regions with the idea of watershed, supplemented by the step of region merging, to achieve the adaptive division of image regions; by statistically counting the number of corresponding feature point matches in each region of two images, and based on the core idea of grid motion statistics, a regional motion statistics constraint scheme is proposed, and corresponding thresholds are set for regions with different numbers of feature point matches in different cases. Through data comparison under different data sets, it can be proved that compared with the grid motion statistics (GMS) method, the present invention can more accurately screen out the correct feature point matches, which is more beneficial to the calculation of subsequent links after feature matching.

Description

Technical Field

[0001] The present invention relates to a method for matching and screening feature points based on adaptive regional motion statistics, belonging to the field of computer vision. Background Art

[0002] Adaptive regional motion statistics divides an image into non-overlapping regions adaptively through the gray value information of the image. After counting the number of feature point matches between regions, the screening of feature point matches is carried out by using the constraint of the motion consistency of feature points in each region. Feature points are pixel points that can represent an image or an object in an identical or at least very similar invariant form in other similar images containing the same scene or object. Feature point matching is to find the corresponding relationship for the obtained feature points in two images according to their descriptors. There are many incorrect matches in the initially obtained feature matches, and it is necessary to screen the feature point matches to improve their accuracy so as to achieve the accurate use of feature point matching in application scenarios such as image stitching, visual tracking, and three-dimensional reconstruction.

[0003] The commonly used feature point matching and screening scheme uses the RANSAC method, sets the constraint conditions of feature point matching as geometric relationships such as the fundamental matrix or the homography matrix, and iteratively calculates to screen the feature point matches. Such constraint conditions are computationally complex and commonly use strict geometric constraints, reducing the efficiency of feature point matching and screening.

[0004] In view of the disadvantages of the traditional scheme, Bian J W proposed a method of grid motion statistics [Bian J W, Lin W Y, Matsushita Y, et al. GMS: Grid-based motion statistics for fast, ultra-robust feature correspondence [C]. Honolulu: Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 4181-4190.], which uses grid motion statistics that is more concise than geometric constraints, greatly shortens the running time of the method, and the accuracy of the screening result is not weaker than that of other feature point matching and screening methods. However, the method of grid motion statistics based on grid division of the image has good universality, but lacks regional information, and the grid cannot reflect the image gray level, texture or geometric shape and other information in the region, greatly limiting the accuracy of screening out feature point matches. Summary of the Invention

[0005] To solve the problem of low accuracy in the current feature point matching and screening method, the present invention provides a feature point matching and screening method based on adaptive region motion statistics, which includes the following steps:

[0006] Step 1: Divide the image into several regions by using the watershed algorithm based on the relationship of pixel gray values in the image where the feature point matching is located;

[0007] Step 2: For the regions obtained in Step 1, use the constraint conditions of the maximally stable extremal regions to screen out the maximally stable extremal regions with relatively small changes in region area caused by changes in pixel gray thresholds;

[0008] Step 3: Remove the overlapping parts in the regions of the maximally stable extremal regions and mark the region numbers;

[0009] Step 4: Count the number of feature point matches and the corresponding relationships in each region;

[0010] Step 5: Screen out the correct region correspondences through the constraint conditions of region motion statistics and retain the feature point matches therein.

[0011] Optionally, the specific content of Step 1 is as follows:

[0012] Step 11: Initialize the data structures, including: a region stack, a history stack, and a boundary heap. The region stack is a data structure for storing pixel information of image regions under different gray thresholds. The history stack is a data structure for recording the process of increasing the threshold in the region stack. The boundary heap is a data structure for storing the set of boundary pixels of the regions;

[0013] A marker chunk is first placed in the region stack. When the marker chunk is popped, Step 1 ends, and the gray threshold corresponding to the marker chunk is set to 256;

[0014] Step 12: Set the starting pixel as the current pixel. The gray value of the starting pixel is the current gray threshold, and mark it as visited;

[0015] Step 13: Add an empty chunk to the region stack, and the value of the empty chunk is the current pixel gray value;

[0016] Step 14: Sequentially access the four-neighborhood of the current pixel, mark the unvisited neighborhood pixels as visited. If the pixel gray value is not less than the current gray threshold, put it into the boundary heap; if the pixel gray value is lower than the current gray threshold, put the current pixel into the top of the region stack, set this neighborhood pixel as the current pixel, and return to Step 13;

[0017] Step 15: Accumulate the current pixel into the chunk at the top of the region stack, pop the boundary heap pixel. If the heap is empty, end;

[0018] Step 16: If the popped boundary gray value is equal to the gray value of the top chunk in the current region stack, take the popped pixel as the current pixel and return to Step 14; if the popped boundary gray value is higher than the gray value of the top chunk in the current region stack, process all components in the region stack until the gray values of all chunks in the region stack are higher than the current popped boundary gray value;

[0019] Step 17: If the popped boundary gray value is less than the gray value of the second top chunk in the region stack, record the top chunk, set the gray value of the top chunk to the gray value of this boundary, and set the boundary pixel as the current pixel and return to Step 14; conversely, if the popped boundary gray value is not less than the gray value of the second top chunk in the region stack, add the top chunk in the region stack to the historical stack, merge the top chunk in the region stack and the second top chunk in the region stack, and return to Step 16.

[0020] Optionally, the specific content of Step 2 is as follows:

[0021] After Step 1 is completed, regions under each threshold are obtained. Use the following formula to determine whether a region is a maximally stable extremal region:

[0022]

[0023] where i is the gray value threshold, Q i is a certain region when the threshold is i, delta is the change of the gray threshold, and q(i) is the change rate of region Q i when the threshold is i. When it is less than the set maximum change rate, it is considered that this region is a maximally stable extremal region.

[0024] Optionally, the change delta of the gray threshold is 2.

[0025] Optionally, the specific content of Step 3 is as follows:

[0026] Based on Step 1 and Step 2, it is known that a region with a certain threshold is gradually expanded from regions within it with thresholds smaller than it. The operation of retaining the largest region is as follows:

[0027] Initialize a matrix with the image size, mark all values in the matrix as -1, traverse the region set in reverse order, starting from label 0. By traversing whether the pixel coordinates in the region are already marked in the matrix, determine whether this region is a smaller region. If the matrix coordinate is already marked, it means that there is already a larger region than this region marked, then remove this region and proceed to traverse the next region.

[0028] Optionally, the specific content of Step 4 is as follows:

[0029] Two images I a , I bAfter performing the above steps 1, 2, and 3, the image matrix M after region division has been obtained. a , M b , and the set of feature point matches to be filtered is matches. The i-th match in matches i = {f ia , f ib}, where f ia , f ib correspond to the coordinates of the corresponding feature points of the i-th feature point match in images I a , I b respectively.

[0030] Initialize the mapping relationships Map ab and Map ba . Denote the number of feature point matches in the region labeled α in I ab corresponding to the region labeled β in I a by Map b [α][β].

[0031] Traverse the set of feature point matches matches. For f ia , f ib in the set of feature point matches matches, check their corresponding labels in the image matrices M a , M b respectively. If the corresponding value is -1, discard this pair of matches. If the corresponding labels in M a , M b are α and β, then increment the values of both Map ab [α][β] and Map ba [β][α] by 1. If there is no such corresponding relationship, new mappings need to be added to Map ab and Map ba with an initial value of 1. After traversal, the feature statistics corresponding to the regions are obtained.

[0032] Optionally, step 5 is specifically as follows:

[0033] After step 4, the feature point match mapping relationships Map ab and Map ba for each region in the two images have been obtained. At this stage, constraints for regional motion statistics are imposed on them to filter out the correct corresponding regions and retain the feature point matches within the regions.

[0034] Traverse the mapping Map ab to find the mapping with the most feature point matches in the region in image I a corresponding to the region in image I b . Then verify whether this mapping relationship is also the one with the most feature point matches in the region in image I ba in Map bThe middle region corresponds to Figure I a The mapping with the most matching feature points in the middle region

[0035] Optionally, step 5 uses a double associative container, including:

[0036] Map ab During traversal, if the key is α, the corresponding mapping of the value is Map ab [α], find the mapping with the largest value among them, denoted as Map ab [α][β], and at the same time calculate Map ab The sum of the values of [α], that is, the number of matching feature points in region α is denoted as sum α ;

[0037] Verify Map ba Whether the mapping group [β][α] is also the largest in Map ba [β], if satisfied, judge whether the value of Map ab [α][β] is greater than the following set threshold τ:

[0038]

[0039] where average is the average number of matching feature points in the region in Figure I a in the middle region

[0040] Optionally, the four-neighborhood of the current pixel is the right, bottom, left, and upper neighborhoods

[0041] The beneficial effects of the present invention are:

[0042] The present invention uses the method of dividing the maximum stable extreme value region with the watershed idea, supplemented by the step of region merging, to realize the adaptive division of the image region; by counting the number of matching feature points corresponding to each region in two images, and based on the core idea of grid motion statistics, a region motion statistics constraint scheme is proposed, and corresponding thresholds are set for regions with different numbers of matching feature points. The simulation results under different data sets prove that the present invention can more accurately screen out the correct feature point matches and is more conducive to the calculation of the subsequent links after feature matching Description of the Drawings

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings

[0044] Figure 1This is an example diagram of adaptive region division for the feature point matching and screening method of adaptive region motion statistics in the second embodiment of the present invention.

[0045] Figure 2 This is the motion consistency analysis of the feature point matching and screening method of adaptive region motion statistics in the second embodiment of the present invention.

[0046] Figure 3 This is the correct / wrong region corresponding probability mass function diagram of the feature point matching and screening method of adaptive region motion statistics in the second embodiment of the present invention.

[0047] Figure 4 This is the result diagram of the feature point matching and screening without using the method in the second embodiment of the present invention.

[0048] Figure 5 This is the result diagram of the feature point matching and screening method of adaptive region motion statistics in the second embodiment of the present invention.

[0049] Figure 6 This is the algorithm comparison line chart of the perspective change data set of the feature point matching and screening method of adaptive region motion statistics in the second embodiment of the present invention.

[0050] Figure 7 This is the algorithm comparison line chart of the blur change data set of the feature point matching and screening method of adaptive region motion statistics in the second embodiment of the present invention.

[0051] Figure 8 This is the algorithm comparison line chart of the brightness change data set of the feature point matching and screening method of adaptive region motion statistics in the second embodiment of the present invention. Detailed implementation method

[0052] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will further describe the embodiments of the present invention in detail with reference to the accompanying drawings.

[0053] Embodiment 1:

[0054] This embodiment provides a feature point matching and screening method for adaptive region motion statistics, including the following steps:

[0055] Step 1: Divide the image into several regions by using the watershed algorithm based on the relationship of pixel gray values in the image where the feature points are matched;

[0056] Step 2: For the regions obtained in Step 1, use the constraint conditions of the maximally stable extremal regions to screen out the regions with relatively small changes in region area caused by changes in pixel gray value thresholds;

[0057] Step 3: Remove the overlapping parts in the regions of the maximally stable extremal regions and mark the region numbers;

[0058] Step 4: Count the number of feature point matches and their corresponding relationships in each region;

[0059] Step 5: Through the constraint conditions of regional motion statistics, filter out the correct region correspondences and retain the feature point matches among them.

[0060] Embodiment 2:

[0061] This embodiment provides a method for screening feature point matches based on adaptive regional motion statistics, including the following steps:

[0062] Step 1: Based on the relationship of pixel gray values in the image where the feature point matches are located, use the watershed method [Vincent L, Soille P. Watersheds in digital spaces: an efficient algorithm based on immersion simulations [J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 1991, 13(06): 583 - 598.] to divide the image into regions under different thresholds. The specific steps are as follows:

[0063] Step 11: Initialize the data structures, including: a region stack, a history stack, and a boundary heap. The region stack is a data structure for storing pixel information of image regions under different gray thresholds. The history stack is a data structure for recording the process of increasing the threshold in the region stack. The boundary heap is a data structure for storing the set of boundary pixels of the regions;

[0064] First, a marker chunk is placed in the region stack. When the marker chunk is popped, Step 1 ends, and the gray threshold corresponding to the marker chunk is set to 256;

[0065] Step 12: Set the starting pixel as the current pixel. The gray value of the starting pixel is the current gray threshold, and mark it as visited;

[0066] Step 13: Add an empty chunk to the region stack. The value of the empty chunk is the gray value of the current pixel;

[0067] Step 14: Sequentially access the four - neighborhood of the current pixel. Mark the unvisited neighborhood pixels as visited. If the pixel gray value is not less than the current gray threshold, put it into the boundary heap; if the pixel gray value is lower than the current gray threshold, put the current pixel on the top of the region stack, set this neighborhood pixel as the current pixel, and return to Step 13;

[0068] Step 15: Accumulate the current pixel into the chunk at the top of the region stack, pop the boundary - heap pixel. If the heap is empty, end;

[0069] Step 16: If the popped boundary gray value is equal to the gray value of the top chunk in the current region stack, take the popped pixel as the current pixel and return to Step 14; if the popped boundary gray value is higher than the gray value of the top chunk in the current region stack, process all components in the region stack until the gray values of all chunks in the region stack are higher than the current popped boundary gray value;

[0070] Step 17: If the popped boundary gray value is less than the gray value of the second chunk from the top of the region stack, record the top chunk, set the gray value of the top chunk to the gray value of this boundary, and set the boundary pixel as the current pixel and return to Step 14; conversely, if the popped boundary gray value is not less than the gray value of the second chunk from the top of the region stack, add the top chunk of the region stack to the historical stack, merge the top chunk of the region stack and the second chunk from the top of the region stack, and return to Step 16.

[0071] Step 2: After Step 1 ends, regions under each threshold are obtained. Use the maximally stable extremal regions constraint [Nistér D, Stewénius H. Linear time maximally stable extremal regions [C]. Berlin: Proceedings of European conference on computer vision. 2008: 183 - 196.] to screen out regions with relatively small changes in region area caused by changes in pixel gray value thresholds. The formula is as follows, where i is the gray value threshold, Q i is a certain region when the threshold is i, delta is a small change in the gray value threshold, and q(i) is the change rate of region Q i when the threshold is i. When it is less than the set maximum change rate, this region is considered a maximally stable extremal region. Here, to adapt to the overall algorithm, delta is generally set to 2.

[0072]

[0073] Step 3: Remove the overlapping parts in the maximally stable extremal regions and mark the region numbers. Based on the processes of Steps 1 and 2, it can be known that the region of a certain threshold is gradually expanded from the regions within it that are smaller than its threshold. The operation of retaining the largest region is as follows.

[0074] Initialize a matrix of the image size, mark all the values in the matrix as -1, traverse the region set in reverse order, starting from label 0. By traversing whether the pixel coordinates in the region are already marked in the matrix, determine whether this region is a smaller region. If the matrix coordinate is already marked, it means that there is already a larger region than this region marked, so remove this region and proceed to traverse the next region. Finally, for example, Figure 1The effect diagram of image adaptive partitioning shown

[0075] Step 4: Count the number of feature point matches and their corresponding relationships in each region. The two images I a , I b After steps 1, 2, and 3, the image matrix M a , M b has been obtained. The set of feature point matches to be screened is matches, where the i-th match matches i ={f ia , f ib}, f ia and f ib correspond to the coordinates of the corresponding feature points of the i-th feature point match on the images I a , I b respectively. Initialize the mapping relationships Map ab and Map ba . For convenience of expression here, Map ab [α][β] represents the number of feature point matches in the region numbered α in I a corresponding to the region numbered β in I b . However, to implement this mapping relationship here, it is recommended to use a double associative container instead of a two-dimensional array-like implementation.

[0076] Traverse matches, and for f i , f ia in matches ib , check the corresponding labels in M a , M b respectively. If the corresponding value is -1, discard this pair of matches. If the corresponding labels in M a , M b are α, β, then increment the values corresponding to Map ab [α][β] and Map ba [β][α] by 1. If there is no such corresponding relationship, new mappings need to be added to Map ab and Map ba , and the initial value is set to 1. After traversing, the feature statistics corresponding to the regions are obtained.

[0077] Step 5: After step 4, the mapping relationships Map ab and Map ba of the feature points in each region of the two images have been obtained. At this stage, perform regional motion statistics constraints on them, screen out the correct corresponding regions, and retain the feature point matches within the regions.

[0078] The regional motion statistics constraint believes that motion consistency will cause other matches in the neighborhood of the correct match to have similar motions. For example Figure 2After the feature point matching of the left and right images shown, a certain area is selected for display. In the area outlined by the black wireframe in the figure, the number of incorrect matches is 10, and the number of correct matches is 68. Among them, the correct matches are aggregated, and the incorrect matches are scattered in various areas of the image. Only retaining the feature matches within the aggregated area can achieve the purpose of eliminating incorrect matches.

[0079] Traverse the mapping Map ab to find the mapping in which the area in Figure I a has the most feature point matches corresponding to the area in Figure I b , and then verify whether this mapping relationship is also the mapping in Map ba in which the area in Figure I b has the most feature point matches corresponding to the area in Figure I a . Specifically, the method in the double-associated container: When traversing Map ab , if the key is α and the corresponding mapping of the value is Map ab [α], find the mapping with the largest value among them, denoted as Map ab [α][β]. At the same time, calculate the sum of the values of Map ab [α], that is, the number of feature point matches in area α is denoted as sum α . Verify whether the mapping group Map ba [β][α] is also the largest in Map ba [β]. If it is satisfied, judge whether the value of Map ab [α][β] is greater than the set threshold τ.

[0080] The threshold division is based on the conclusion: The more the number of matches contained in the area, the stronger the distinguishability between correct and incorrect matches. This conclusion can be supported by probability calculation. Since the matching of each feature point is independent, the binomial distribution can be used to approximate the number of correct feature point matches contained in two corresponding areas. Assume that the accuracy rate of the matching algorithm is 0.6. When a feature point in a certain area of Figure I a is matched incorrectly but still matches to the area of Figure I b , the probability is set to 0.1. At this time, according to the probability mass function of the binomial distribution, the results shown in Figure 3 can be obtained, where Figure 3 (a) The number of feature point matches in the area is set to 100, Figure 3 (b) The number of feature point matches in the area is set to 1000. The dotted line in the figure corresponds to the probability mass function curve of the error, and the solid line corresponds to the probability mass function curve of the correct. It can be seen that the overall probability curve is bimodal, there is a certain distance between the two curves, and only when the number of features in two corresponding areas reaches the threshold, will the event of correct area correspondence occur. Therefore, it can be seen that the more the number of matches contained in the area, the stronger the distinguishability between correct and incorrect matches.

[0081]

[0082] In the above threshold τ setting, where average is the average number of feature point matches contained in the area in Figure I a There are three types of thresholds: one is that the number of feature matches in this area is much less than the average. At this time, when the number of feature point matches is small, the distinguishability between the correct area corresponding and the wrong area corresponding is weak, so a more stringent threshold setting is required, and even if the corresponding judgment is wrong, the loss of feature point matches is less; another is that the number of feature matches in this area is much greater than the average, that is, greater than average * 4. At this time, when there are many feature points in the area, the distinguishability between the correct area corresponding and the wrong grid area is large, and a relatively loose threshold is set, which can not only ensure the reliability of the method, but also prevent the feature points in this area from being easily eliminated; in addition to the above two situations, the threshold is set within the range between the above two situations.

[0083] The specific usage effect of the feature point matching and screening method for adaptive area motion statistics in this embodiment is as follows:

[0084] The system environment is as follows. Table 1 is the system hardware configuration table, and Table 2 is the system software configuration table.

[0085] Table 1: System Hardware Configuration Table

[0086]

[0087] Table 2: System Software Configuration Table

[0088] Software Related Information Operating System Windows 10 64-bit Visual Studio 2019 OpenCV 4.5.5

[0089] Comparison of implementation effects:

[0090] Figure 4 The result graph without using the feature point matching and screening method is shown.

[0091] Figure 5 The effect graph of using the feature point matching and screening method for adaptive area motion statistics of the present invention is shown. It can be seen that there are many irregular and chaotic wrong matches in the feature point matching before using. After the method of the present invention, it can be seen that the feature point matching becomes more orderly and there is motion consistency in the matching.

[0092] Figure 6It shows that on a sequence of images with a changing viewing angle, in the case of an affine transformation where the image changes from small to large, without a screening method as a reference, the precision of screening and matching point features of the method of the present invention and the classical GMS method is compared, and the precision of the homography calculated after screening feature point matching is compared. It can be seen that in most image pairs, the present invention is slightly better than the classical GMS method, and is more conducive to the calculation of the homography matrix after feature matching.

[0093] Figure 7 It shows that on a sequence of blurred images obtained due to camera focus change, without a screening method as a reference, the precision of screening and matching point features of the method of the present invention and the classical GMS method is compared, and the precision of the homography calculated after screening feature point matching is compared. It can be seen that in most image pairs, the present invention is slightly better than the classical GMS method, and is more conducive to the calculation of the homography matrix after feature matching.

[0094] Figure 8 It shows that on a sequence of brightness images obtained by changing the camera aperture, without a screening method as a reference, the precision of screening and matching point features of the method of the present invention and the classical GMS method is compared, and the precision of the homography calculated after screening feature point matching is compared. It can be seen that in most image pairs, the present invention is slightly better than the classical GMS method, and is more conducive to the calculation of the homography matrix after feature matching.

[0095] Some steps in the embodiments of the present invention can be implemented by software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk, etc.

[0096] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A feature point matching and screening method based on adaptive regional motion statistics, characterized by including the following steps: Step 1: Divide the image into several regions using the watershed algorithm based on the relationship of pixel gray values in the image where feature point matching is located; Step 2: For the regions obtained in Step 1, use the constraint conditions of the maximally stable extremal regions to screen out the maximally stable extremal regions with relatively small changes in regional area caused by changes in pixel gray thresholds; Step 3: Remove the overlapping parts in the regions of the maximally stable extremal regions and mark the region numbers; Step 4: Statistically analyze the number of feature point matches and corresponding relationships in each region; Step 5: Obtain the feature point matching mapping relationship Map for each region in the two images after Step 4 ab and Map ba , traverse the mapping Map ab , find the mapping in which the region in Image I a corresponds to the region in Image I b with the most matching feature points, and then verify whether this mapping relationship is also the mapping in Map ba in which the region in Image I b corresponds to the region in Image I a with the most matching feature points, and retain the feature point matching within the region; The said Step 5 utilizes a dual associative container, including: Map ab During traversal, if the key is α, the corresponding mapping of the value is Map ab [α], find the mapping with the largest value among them, denoted as Map ab [α][β], and at the same time calculate Map ab The sum of the values of [α], that is, the number of feature point matches in region α is denoted as sum α ; Verify Map ba Verify whether the mapping of this group [β][α] is in the Map ba is also the largest in [β]. If satisfied, judge the Map ab whether the value of [α][β] is greater than the preset threshold τ below: where average is the number of average feature point matches contained in the area in Figure I a ​ 2. The feature point matching and screening method for adaptive region motion statistics according to claim 1, wherein The specific content of the said Step 1 is: Step 11: Initialize the data structures, including: a region stack, a history stack, and a boundary heap. The region stack is a data structure for storing pixel information of image regions under different gray thresholds. The history stack is a data structure for recording the process of increasing the threshold in the region stack. The boundary heap is a data structure for storing the set of boundary pixels of the regions; A marker chunk is first placed in the region stack. When the marker chunk pops out, Step 1 ends, and the gray threshold corresponding to the marker chunk is set to 256; Step 12: Set the starting pixel as the current pixel. The gray value of the starting pixel is the current gray threshold, and mark it as visited; Step 13: Add an empty chunk to the region stack, and the value of the empty chunk is the current pixel gray value; Step 14: Sequentially access the four-neighborhood of the current pixel, mark the unvisited neighborhood pixels as visited. If the pixel gray value is not less than the current gray threshold, put it into the boundary heap; if the pixel gray value is lower than the current gray threshold, put the current pixel into the top of the region stack, set this neighborhood pixel as the current pixel, and return to Step 13; Step 15: Accumulate the current pixel into the chunk at the top of the region stack, pop out the boundary heap pixel. If the heap is empty, end; Step 16: If the popped boundary gray value is equal to the gray value of the current chunk at the top of the region stack, set the popped pixel as the current pixel and return to Step 14; if the popped boundary gray value is higher than the gray value of the current chunk at the top of the region stack, process all components in the region stack until the gray values of all chunks in the region stack are higher than the current popped boundary gray value; Step 17: If the popped boundary gray value is less than the gray value of the second chunk from the top of the region stack, record the top chunk of the stack, set the gray value of the top chunk of the stack as the gray value of this boundary, and set the boundary pixel as the current pixel and return to Step 14; otherwise, if the popped boundary gray value is not less than the gray value of the second chunk from the top of the region stack, add the top chunk of the region stack to the history stack, merge the top chunk of the region stack and the second chunk from the top of the region stack, and return to Step 16.

3. The feature point matching and screening method for adaptive regional motion statistics according to claim 1, wherein The specific content of the said Step 2 is: After Step 1 ends, the regions under each threshold are obtained. Use the following formula to determine whether the region is a maximally stable extremal region: where i is the grayscale value threshold, Q i is a certain region when the threshold is i, delta is the change of the grayscale threshold, and q(i) is the change rate of the region Q i when it is less than the set maximum change rate, then this region is considered the maximum stable extreme region.

4. The feature point matching and screening method for adaptive region motion statistics according to claim 3, wherein The change delta of the gray threshold is 2.

5. The feature point matching and screening method for adaptive regional motion statistics according to claim 1, wherein The specific content of the said Step 3 is: Based on Steps 1 and 2, it is known that the region at a certain threshold is gradually expanded from the regions with thresholds smaller than it within the region. The operation of retaining the largest region is as follows: Initialize the matrix of the image size, mark all the values in the matrix as -1, traverse the region set in reverse order. When the label starts from 0, determine whether the region is a smaller region by traversing whether the pixel coordinates in the region are already marked in the matrix. If the matrix coordinates are already marked, it means that there is a larger region than this region that has been marked, so remove this region and proceed to traverse the next region.

6. The feature point matching and screening method for adaptive regional motion statistics according to claim 1, wherein The specific steps of step 4 are as follows: Two images I where the feature points are located a , I b After the steps 1, 2, and 3, the image matrix M after region division has been obtained a , M b , the set of feature point matches to be screened is matches, where the i-th match matches i = {f ia , f ib}, f ia , f ib correspond to the coordinates of the corresponding feature points of the i-th feature point match in the image I a , I b respectively; Initialize the mapping relationship Map ab and Map ba , taking Map ab [α][β] represents the number of feature point matches in area β corresponding to area α with label I a in I with label α b in I with label β; Traverse the feature point matching set matches, and for f in the feature point matching set matches ia , f ib Check the corresponding labels in the image matrices M a , M b respectively. If the corresponding value is -1, discard this pair of matches; if the corresponding labels of M a , M b are α and β, then increment the values corresponding to Map ab [α][β] and Map ba [β][α] by 1. If there is no such corresponding relationship, new mappings need to be added to Map ab and Map ba with an initial value of 1. After traversal, the feature statistics corresponding to the region are obtained.

7. The feature point matching and screening method for adaptive region motion statistics according to claim 2, wherein The four-neighborhood of the current pixel is the right, bottom, left, and upper neighborhoods.

Images