A video image stitching method based on descriptor matching

Abstract

The present invention discloses a video image stitching method based on descriptor matching, which belongs to the technical field of video image stitching, and comprises the following steps: step 1, a drone carries two depth cameras to collect environmental information, and completes video image initialization; step 2, performs timestamp fault diagnosis on the two depth cameras, and completes timestamp alignment and correction of the two depth cameras; step 3, designs a stitching descriptor, and extracts the stitching descriptor from the primary video image collected by the primary depth camera; step 4, designs a hierarchical region scoring algorithm, matches the extracted stitching descriptor with the secondary video image collected by the secondary depth camera, calculates a similarity score, and then obtains an overlapping region; step 5, initializes a blank image, and completes video image stitching. The present invention can find accurate overlapping regions and realize video image stitching.

Description

Technical Field

[0001] The invention belongs to the technical field of video image splicing, and in particular relates to a video image splicing method based on descriptor matching. Background Art

[0002] In today's society, electronic technology and multimedia technology have made rapid progress and development. Video, as a way of information dissemination, plays an important role in people's production and life. However, the field of view of a single video acquisition device is relatively small, and the field of view captured is generally around 60 degrees. Therefore, using multiple single video acquisition devices to achieve image stitching can solve the defect of the small field of view of the above-mentioned single video acquisition device, so as to obtain a large-viewing angle image video.

[0003] Among the existing technologies for realizing video image stitching, there are several methods: a video image stitching method, system, electronic device and storage medium, a real-time stitching method for drone video images based on spherical transformation, and a study on key technologies for image stitching and target detection for railway freight monitoring videos. Although these methods can realize the video image stitching function, they have the following disadvantages: first, they do not consider the time consistency problem in the process of stitching multiple video images; second, the real-time performance is not strong during feature matching, resulting in mis-stitching and large stitching errors. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a video image stitching method based on descriptor matching, which uses a drone equipped with two depth cameras to collect environmental information, designs a stitching descriptor, uses the grayscale value and depth value of the pixel point to calculate the value in the stitching descriptor, and uses a graded area scoring method to match the extracted stitching descriptor with the video image collected by the secondary depth camera to calculate the similarity score. The present invention can find accurate overlapping areas and realize video image stitching.

[0005] The technical solution of the present invention is as follows:

[0006] A video image stitching method based on descriptor matching comprises the following steps:

[0007] Step 1: The drone is equipped with two depth cameras to collect environmental information and complete video image initialization;

[0008] Step 2: Perform timestamp fault diagnosis on the two depth cameras and complete timestamp alignment and correction of the two depth cameras;

[0009] Step 3: Design a splicing descriptor and extract the splicing descriptor from the first-level video image collected by the first-level depth camera;

[0010] Step 4: Design a hierarchical region scoring algorithm to match the extracted splicing descriptor with the secondary video image captured by the secondary depth camera, calculate the similarity score, and then obtain the overlapping area;

[0011] Step 5: Initialize a blank image and complete video image stitching.

[0012] Furthermore, in step 1, the two depth cameras carried by the drone are divided into a primary depth camera and a secondary depth camera, which respectively collect a primary video image with a size of 480 * 640 and secondary video images ,The positions of the primary depth camera and the secondary depth camera are fixed, and there is an overlapping area during video acquisition;

[0013] The video image initialization process is: define the depth value of each pixel in the first-level video image as After graying the video image, the gray value of each pixel in the first-level video image is ; Define the depth value of each pixel in the secondary video image containing the environment as After graying the video image, the gray value of each pixel in the secondary video image is .

[0014] Furthermore, the specific process of step 2 is as follows:

[0015] Step 2.1: Initialize the distance between the two depth cameras and signal transmission speed , calculate the delay time according to the following formula :

[0016] ;

[0017] Step 2.2: 3 seconds after the first-level depth camera and the second-level depth camera are turned on, they simultaneously send their respective timestamps to each other and receive the timestamp information sent by the other party. , calculate the alignment timestamp according to the following formula :

[0018] ;

[0019] Step 2.3, compare the alignment timestamp and the current timestamp. If they are consistent, it means that the timestamps of the two depth cameras are aligned consistently. If they are inconsistent, the current timestamp is corrected based on the alignment timestamp, that is, the value of the current timestamp is adjusted to the value of the alignment timestamp.

[0020] Furthermore, the specific process of step 3 is as follows:

[0021] Step 3.1: Extract the pixel images of the 639th and 640th columns in the primary video image to form a splicing descriptor. ;

[0022] Step 3.2: Splice the descriptor Divided into 60 intervals, each interval is used as a splicing descriptor Description unit, that is, each description unit is a sub-image of size 8*2;

[0023] Step 3.3, calculate the grayscale mean and depth mean of each interval sub-image pixel; The calculation formula for the grayscale mean and depth mean of the sub-image pixels in the interval is as follows, where: :

[0024] ;

[0025] ;

[0026] in, For the The grayscale mean of the pixels in the sub-image of the interval; For the The mean depth of the sub-image pixels in the interval; For the Interval subgraph The gray value of each pixel; For the Interval subgraph The gray value of each pixel;

[0027] Step 3.4: Concatenate descriptors There are 60 intervals, and 60 intervals correspond to 60 description unit values, that is, The interval corresponds to Descriptor values; by and Compute splice descriptors No. Descriptor value , the specific calculation formula is shown as follows:

[0028] .

[0029] Furthermore, the specific process of step 4 is as follows:

[0030] Step 4.1: The splicing descriptor extracted from the primary video image needs to be matched with the secondary video image captured by the secondary depth camera. The order is to move from the leftmost side of the secondary video image and calculate the similarity score; initialize the similarity threshold , if the difference between the previous similarity score and the next similarity score When it is greater than the similarity threshold, the primary search is completed;

[0031] Step 4.2, perform a secondary search; assume that the first column of the previous similarity score is , is an odd number, then and The columns form a new subgraph, namely Subgraph, match the concatenated descriptor and the new subgraph to calculate the similarity score and get the The similarity scores of the columns are compared The similarity scores of the columns and The size of the similarity score of the column, if the The similarity score of the column is greater than The similarity score of the columns is is the right boundary of the real overlapping area, otherwise Listed as the right boundary of the true overlap area;

[0032] Step 4.3: After steps 4.1 and 4.2, the left boundary of the accurate real overlapping area is obtained. and right border ; The specific expression is as follows:

[0033] ;

[0034] ;

[0035] in, The first video image List; is the first column in the secondary video image; The first video image List; The second video image List.

[0036] Furthermore, the specific process of step 4.2 is as follows:

[0037] Step 4.2.1. Extraction The matching unit of the sub-graph; The subgraph is divided into 60 intervals, each interval as The matching unit of the subgraph calculates Interval The grayscale mean of the sub-image pixels and depth mean , the specific calculation formula is as follows:

[0038] ;

[0039] ;

[0040] in, For the Interval Sub-image The gray value of each pixel; For the Interval Sub-image The depth value of each pixel;

[0041] Each The subgraph has 60 intervals, and the 60 intervals correspond to 60 matching unit values, that is, The interval corresponds to Matching unit values; by and calculate The sub-graph Matching cell values , the specific calculation formula is shown as follows:

[0042] ;

[0043] Step 4.2.2: Calculate the splicing descriptor Describes the average value of the cell and The average of the subgraph matching cell values , the calculation formula is as follows:

[0044] ;

[0045] ;

[0046] Step 4.2.3: Calculate the splicing descriptor Splicing unit and Similarity scores between subgraph matching units , the calculation formula is as follows:

[0047] .

[0048] Furthermore, the specific process of step 4.2 is as follows: Initialize the size of the blank image to ; Select all the first-level video images and the second-level video images The pixel points of 640 columns are stitched together to complete the image stitching task.

[0049] The beneficial technical effects brought about by the present invention are as follows.

[0050] 1. When realizing video image stitching, the present invention fixes the positions of the two depth cameras, aligns and corrects the timestamps, and avoids stitching errors caused by timestamp errors during the video image stitching process. Compared with the method of directly performing video image stitching, the scheme of the present invention adds a timestamp fault diagnosis mechanism, has stronger compatibility, gets rid of the dependence of the depth camera on the central controller, and at the same time improves the accuracy of video image stitching and reduces errors.

[0051] 2. The present invention designs a splicing descriptor, and uses the grayscale value and depth value of the pixel point to calculate the value in the splicing descriptor. Compared with the method of matching using other features such as SIFT, the scheme of the present invention has a small number of features and high calculation efficiency, which can further improve the real-time performance of video image splicing and improve the utilization rate of computing resources, thereby further reducing hardware costs and having high economic efficiency.

[0052] 3. The present invention adopts a graded area scoring method to match the extracted stitching descriptors and the video images captured by the secondary depth camera to calculate the similarity score. Compared with other methods, the calculation accuracy of the scheme of the present invention is higher, and the scoring error caused by only one scoring can be avoided, thereby finding the accurate overlapping area and realizing the video image stitching function. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The flowchart of the video image stitching method based on descriptor matching of the present invention. DETAILED DESCRIPTION

[0054] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0055] like Figure 1 As shown, a video image stitching method based on descriptor matching mainly includes the following steps:

[0056] Step 1: The drone is equipped with two depth cameras to collect environmental information and complete video image initialization; the specific process is as follows:

[0057] The two depth cameras carried by the drone are divided into a primary depth camera and a secondary depth camera. The two depth cameras collect environmental information and collect primary video images with a size of 480 * 640. and secondary video images , used for subsequent video image stitching; the positions of the first-level depth camera and the second-level depth camera are fixed, and there is an overlapping area during video acquisition, but due to the influence of other factors such as mechanical looseness, the area of ​​the overlapping area is not constant, so it is necessary to further determine the overlapping area;

[0058] Complete the video image initialization and define the depth value of each pixel in the first-level video image containing the environment as After graying the video image, the gray value of each pixel in the first-level video image is Similarly, the depth value of the environment contained in each pixel in the secondary video image is defined as After graying the video image, the gray value of each pixel in the secondary video image is ;

[0059] Step 2: Perform timestamp fault diagnosis on the two depth cameras and complete timestamp alignment and correction of the two depth cameras;

[0060] The timestamps of the first-level depth camera and the second-level depth camera are compared and controlled according to the central controller before use. In order to eliminate the dependence of the depth camera on the central controller and ensure that when a time error occurs, the video image stitching task can be completed by relying on the consistency of the timestamps between the two depth cameras, the timestamp fault diagnosis method is adopted to complete the alignment and correction of the timestamps of the two depth cameras through the timestamp fault diagnosis;

[0061] The specific process of using timestamp fault diagnosis to complete the timestamp alignment and correction of two depth cameras is as follows:

[0062] Step 2.1: Initialize the distance between the two depth cameras and signal transmission speed , calculate the delay time according to the following formula :

[0063] (1);

[0064] Step 2.2: 3 seconds after the first-level depth camera and the second-level depth camera are turned on, they simultaneously send their respective timestamps to each other and receive the timestamp information sent by the other party. , calculate the alignment timestamp according to the following formula :

[0065] (2);

[0066] Step 2.3, compare the alignment timestamp and the current timestamp. If they are consistent, it means that the timestamps of the two depth cameras are aligned consistently. If they are inconsistent, the current timestamp is corrected based on the alignment timestamp, that is, the value of the current timestamp is adjusted to the value of the alignment timestamp.

[0067] Step 3: Design a splicing descriptor and extract the splicing descriptor from the first-level video image collected by the first-level depth camera;

[0068] The key to video image stitching is to identify the overlapping area. For traditional methods, SIFT (Scale Invariant Feature Transform) feature points are often extracted for image matching. However, there are many SIFT feature points, and it is difficult to ensure real-time performance during calculation. For video image stitching, it is easy to cause inaccurate stitching. Therefore, the present invention designs a stitching descriptor to identify the overlapping area.

[0069] During the video image stitching process, the width of the overlapping area is 480 pixels; the overlapping area includes the left border and the right border. For the overlapping area, the left border and the secondary video image Align the first column in the video image to the right margin and level The last column in is aligned;

[0070] Design a splicing descriptor and extract it from the first-level video image Extract the splicing descriptor from the repository. The specific steps are as follows:

[0071] Step 3.1: Extract the pixel images of the 639th and 640th columns in the primary video image to form a splicing descriptor. ;

[0072] Step 3.2: Splice the descriptor Divided into 60 intervals, each interval is used as a splicing descriptor Description unit, that is, each description unit is a sub-image of size 8*2;

[0073] Step 3.3, calculate the grayscale mean and depth mean of each interval sub-image pixel; The calculation formula for the grayscale mean and depth mean of the sub-image pixels in the interval is as follows, where: :

[0074] (3);

[0075] (4);

[0076] in, For the The grayscale mean of the pixels in the sub-image of the interval; For the The mean depth of the sub-image pixels in the interval; For the Interval subgraph The gray value of each pixel; For the Interval subgraph The gray value of each pixel;

[0077] Step 3.4: Concatenate descriptors There are 60 intervals, and 60 intervals correspond to 60 description unit values, that is, The interval corresponds to Descriptor values; by and Compute splice descriptors No. Descriptor value , the specific calculation formula is shown as follows:

[0078] (5).

[0079] Step 4: Design a hierarchical region scoring algorithm to extract the spliced ​​descriptors and the secondary video images captured by the secondary depth camera. Match and calculate the similarity score to get the overlapping area. The specific process is as follows:

[0080] Step 4.1: The spliced ​​descriptor extracted from the primary video image needs to be matched with the secondary video image captured by the secondary depth camera. The order is to move from the leftmost side of the secondary video image and compare them in turn to calculate the similarity score. For example, the spliced ​​descriptor and the (1,2) sub-image composed of the first and second columns of the secondary video image are matched to calculate the similarity score. The obtained similarity score is the similarity score of the second column. Therefore, only the even columns have similarity scores. Initialize the similarity threshold. , if the difference between the previous similarity score and the next similarity score When it is greater than the similarity threshold, the primary search is completed;

[0081] Step 4.2: In order to further obtain the precise boundary of the overlapping area, a secondary search is required. Assume that the first column of the previous similarity score is , is an odd number, then and The columns form a new subgraph, namely Subgraph, match the concatenated descriptor and the new subgraph to calculate the similarity score and get the The similarity scores of the columns are compared The similarity scores of the columns and The size of the similarity score of the column, if the The similarity score of the column is greater than The similarity score of the column is is the right boundary of the real overlapping area, otherwise Listed as the right edge of the true overlap area.

[0082] The specific process of calculating the similarity score is:

[0083] Step 4.2.1. Extraction The matching unit of the sub-graph; The subgraph is divided into 60 intervals, each interval as The matching unit of the subgraph calculates Interval The grayscale mean of the sub-image pixels and depth mean , the specific calculation formula is as follows:

[0084] (6);

[0085] (7);

[0086] in, For the Interval Sub-image The gray value of each pixel; For the Interval Sub-image The depth value of each pixel;

[0087] Each The subgraph has 60 intervals, and the 60 intervals correspond to 60 matching unit values, that is, The interval corresponds to Matching unit values; by and calculate The sub-graph Matching unit values , the specific calculation formula is shown as follows:

[0088] (8);

[0089] Step 4.2.2: Calculate the splicing descriptor Describes the average value of the cell and The average of the subgraph matching cell values , the calculation formula is as follows:

[0090] (9);

[0091] (10);

[0092] Step 4.2.3: Calculate the splicing descriptor Splicing unit and Similarity scores between subgraph matching units , the calculation formula is as follows:

[0093] (11).

[0094] The present invention specifically takes the (1,2) subgraph as an example, and the specific calculation steps for calculating the similarity score are as follows:

[0095] First, extract the matching unit of the (1,2) sub-image. Divide the (1,2) sub-image into 60 intervals, and use each interval as the matching unit of the (1,2) sub-image. Calculate the grayscale mean of the pixel points of the (1,2) sub-image according to formula (6) and formula (7): and depth mean , when calculating, Substitute (1,2);

[0096] Then, according to formula (8), and As the matching cell value of the (1,2) subgraph ; Each (1,2) subgraph has 60 descriptor values.

[0097] Then, the splicing descriptor is calculated according to formula (9) and formula (10): The average value of the splicing unit and the average of the (1,2) subgraph matching units ;

[0098] Finally, the splicing descriptor is calculated according to formula (11): Similarity score between the concatenated unit and the (1,2) subgraph matching unit ;

[0099] Step 4.3: After the above steps, the left boundary of the real overlapping area can be accurately obtained. and right border , the specific expression is as follows:

[0100] (12);

[0101] (13);

[0102] in, The first video image List; is the first column in the secondary video image; The first video image List; The second video image List.

[0103] Step 5: Initialize a blank image and complete video image stitching. The specific process is:

[0104] Initialize a blank image. The size of the blank image is ;

[0105] Select all primary video images and secondary video images The pixel points of 640 columns are stitched together to complete the image stitching task.

[0106] In order to demonstrate the feasibility of the present invention, the following examples are given.

[0107] According to the above steps 1 to 3, environmental information is collected, video images are initialized, and timestamps of the two depth cameras are aligned and corrected. Then, the similarity score is calculated according to step 4 to obtain the overlapping area; in step 4.1, the similarity score is calculated by moving from the leftmost side of the secondary video image to the left side. For example, the similarity score is calculated by matching the (1,2) sub-image composed of the first and second columns of the spliced ​​descriptor and the secondary video image. The similarity score is the similarity score of the second column. Therefore, only the even columns have similarity scores. The subsequent sub-images of the secondary video image are (3,4), (5,6),..., (639,640). In this embodiment, the sub-image that meets the conditions when the primary search is completed is (5,6). When performing the secondary search, it is assumed that , then the 6th and 7th columns are combined into a new sub-image, namely, the (6,7) sub-image, and the concatenated descriptor and the new sub-image are matched to calculate the similarity score, and the similarity score of the 7 columns is obtained. The similarity score of the 6th column is compared with the similarity score of the 7th column. If the similarity score of the 7th column is greater than the similarity score of the 6th column, the 7th column is the right boundary of the true overlapping area, otherwise the 6th column is the right boundary of the true overlapping area. In this embodiment, the specific process of calculating the similarity score is as follows: first, extract the matching unit of the (6,7) sub-image; divide the (6,7) sub-image into 60 intervals, each interval is used as the matching unit of the (6,7) sub-image, and calculate the grayscale mean and depth mean of the pixel points of each interval (6,7) sub-image according to formula (6) and formula (7). Then, according to formula (8), the grayscale mean and depth mean of the pixel points of each interval (6,7) sub-image are used as each matching unit value of the (6,7) sub-image, and each (6,7) sub-image has 60 matching unit values. Then, the splicing descriptor is calculated according to formula (9) and formula (10): The average value of the splicing unit and the average of the (6,7) subgraph matching units ; Finally, the splicing descriptor is calculated according to formula (11) Similarity score between the concatenated unit and the (6,7) subgraph matching unit After the above steps, according to formula (12) and formula (13), the left boundary of the overlapping area can be accurately obtained. and right border ;in, It is the 633rd column in the first-level video image; is the first column in the secondary video image; It is the 640th column in the first-level video image; is the 7th column in the secondary video image. At this time, the size of the initialized blank image is ; Select the pixel points from the 8th to the 640th columns in all the first-level video images and the second-level video images for stitching to complete the image stitching task.

[0108] It should be noted that in the solution of the present invention, due to the existence of edge distortion, there are at least two columns in the non-overlapping area.

[0109] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A video image stitching method based on descriptor matching, characterized in that: The steps include: Step 1: The drone is equipped with two depth cameras to collect environmental information and complete video image initialization; Step 2: Perform timestamp fault diagnosis on the two depth cameras and complete timestamp alignment and correction of the two depth cameras; Step 3: Design a splicing descriptor and extract the splicing descriptor from the first-level video image collected by the first-level depth camera; the specific process is as follows: Step 3.1, extracting the pixel images of the 639th column and the 640th column in the primary video image to form a splicing descriptor O; Step 3.2, divide the splicing descriptor O into 60 intervals, each interval is used as a description unit of the splicing descriptor O, that is, each description unit is a sub-image of size 8*2; Step 3.3, calculate the grayscale mean and depth mean of each interval sub-image pixel; the calculation formula for the grayscale mean and depth mean of the j-th interval sub-image pixel is as follows, where j = 1, 2, ..., 60: in, is the grayscale mean of the pixels in the jth interval sub-image; is the depth mean of the pixels in the j-th interval sub-image; is the gray value of the i-th pixel in the j-th interval sub-image; is the gray value of the i-th pixel in the j-th interval sub-image; Step 3.4, the splicing descriptor O has 60 intervals, and the 60 intervals correspond to 60 description unit values, that is, the jth interval corresponds to the jth description unit value; through and Calculate the jth description unit value P of the spliced ​​descriptor O j , the specific calculation formula is shown as follows: Step 4: Design a hierarchical region scoring algorithm to match the extracted splicing descriptor with the secondary video image captured by the secondary depth camera, calculate the similarity score, and then obtain the overlapping area; Step 5: Initialize a blank image and complete video image stitching.

2. The video image stitching method based on descriptor matching according to claim 1, characterized in that: In step 1, the two depth cameras carried by the drone are divided into a primary depth camera and a secondary depth camera, which respectively collect a primary video image Q with a size of 480*640 1 and secondary video image Q 2 ,The positions of the primary depth camera and the secondary depth camera are fixed, and there is an overlapping area during video acquisition; The video image initialization process is as follows: define the depth value of each pixel in the primary video image containing the environment as H 1 After graying the video image, the gray value of each pixel in the first-level video image is S 1 ; Define the depth value of each pixel in the secondary video image containing the environment as H 2 After graying the video image, the gray value of each pixel in the secondary video image is S 2 .

3. The video image stitching method based on descriptor matching according to claim 1, characterized in that: The specific process of step 2 is: Step 2.1, initialize the distance I and signal transmission speed V between the two depth cameras, and calculate the delay time t according to the following formula: t = I / V; Step 2.2: 3 seconds after the first-level depth camera and the second-level depth camera are turned on, they simultaneously transmit their respective timestamps to each other and receive the timestamp information T sent by the other party. info , calculate the alignment timestamp T according to the following formula align : T align =F info +t; Step 2.3, compare the alignment timestamp with the current timestamp. If they are consistent, it means that the timestamps of the two depth cameras are aligned consistently. If they are inconsistent, the current timestamp is corrected based on the alignment timestamp, that is, the value of the current timestamp is adjusted to the value of the alignment timestamp.

4. The video image stitching method based on descriptor matching according to claim 1, characterized in that: The specific process of step 4 is as follows: Step 4.1, the splicing descriptor extracted from the primary video image needs to be matched with the secondary video image captured by the secondary depth camera, in the order of moving from the leftmost side of the secondary video image to calculate the similarity score; Initialize the similarity threshold α. If the difference ΔP between the previous similarity score and the next similarity score is greater than the similarity threshold, the primary search is completed. Step 4.2, perform a secondary search; assuming that the first column of the previous similarity score is m, and m is an odd number, then the m+1th and m+2th columns form a new subgraph, that is, the (m+1, m+2) subgraph, match the spliced ​​descriptor and the new subgraph to calculate the similarity score, and obtain the similarity score of the m+2th column, and compare the similarity score of the m+1th column with the similarity score of the m+2th column. If the similarity score of the m+2th column is greater than the similarity score of the m+1th column, the m+2th column is the right boundary of the true overlapping area, otherwise the m+1th column is the right boundary of the true overlapping area; Step 4.3: After steps 4.1 and 4.2, the left boundary of the accurate real overlapping area is obtained. and right border The specific expression is as follows: in, The 640-(m+2)th column in the first-level video image; is the first column in the secondary video image; It is the 640th column in the first-level video image; It is the m+2th column in the secondary video image.

5. The video image stitching method based on descriptor matching according to claim 4 is characterized in that: The specific process of step 4.2 is as follows: Step 4.2.1, extract the matching unit of the (m+1, m+2) sub-image; divide the (m+1, m+2) sub-image into 60 intervals, each interval as the matching unit of the (m+1, m+2) sub-image, and calculate the grayscale mean of the pixel points of the (m+1, m+2) sub-image in the jth interval and depth mean The specific calculation formula is as follows: in, is the gray value of the i-th pixel in the j-th interval (m+1, m+2); is the depth value of the i-th pixel of the j-th interval (m+1, m+2); Each (m+1, m+2) subgraph has 60 intervals, and the 60 intervals correspond to 60 matching unit values, that is, the jth interval corresponds to the jth matching unit value; and Calculate the j-th matching unit value of the (m+1, m+2) subgraph The specific calculation formula is shown as follows: Step 4.2.2: Calculate the average value of the descriptor unit value of the splicing descriptor O and the average of the (m+1, m+2) subgraph matching unit values The calculation formula is as follows: Step 4.2.3: Calculate the similarity score D between the splicing unit of the splicing descriptor O and the (m+1, m+2) sub-image matching unit (m +1，m+2) , the calculation formula is as follows:

6. The video image stitching method based on descriptor matching according to claim 5, characterized in that: The specific process of step 4.2 is: initialize the size of the blank image to 480*(1280-(m+2)); select the pixel points from the m+3th column to the 640th column in all the primary video images and the secondary video images for splicing to complete the image splicing task.

Images