ORB feature matching tracking method based on FPGA

Abstract

The invention relates to the technical field of target tracking of computer vision, and provides an ORB feature matching tracking method based on an FPGA, and the method comprises the steps: detecting FAST feature points, carrying out FAST corner detection, response value calculation and maximum suppression on a 7 * 7 sliding pixel matrix, and outputting effective signals of the FAST feature points; feature point descriptors are generated, BRIEF descriptors are calculated on a 31 * 31 sliding pixel matrix, and combined output of the BRIEF descriptors and pixel coordinates is enabled through FAST feature point effective signals; and feature point matching and updating: storing feature point data of the target, the current state and the secondary state, traversing and matching the target feature points and the current state feature points, and updating the target feature point data. According to the target tracking method based on ORB feature point matching, the target tracking method based on ORB feature point matching is applied to an FPGA platform through software and hardware collaborative design, the resource limitation of an embedded platform and the requirement for processing mass image data in real time are considered, the target template is updated in real time, outdated features are removed, new features are introduced, the effectiveness of the template is ensured, and the reliability of matching tracking is improved.

Description

Technical Field

[0001] The present application belongs to the technical field of target tracking in computer vision, and specifically relates to an ORB feature matching and tracking method based on FPGA. Background Art

[0002] The continued support of national science and technology innovation policies has significantly promoted the development of computer vision, particularly in the field of target tracking. Target tracking technology is a hot topic in computer vision research, widely used in a variety of fields, including video surveillance, autonomous driving, robotic perception, drone aerial photography, and missile guidance, and possesses significant civilian and military applications. In the civilian sector, with the growing demand for smart cities and public safety, target tracking technology has been widely adopted, particularly in areas such as pedestrian tracking, autonomous vehicles, and intelligent video surveillance. These applications have improved public safety and traffic efficiency, promoted the development of smart city infrastructure, and enhanced the real-world value of target tracking technology. With technological advancements and policy guidance, target tracking technology is expected to play a greater role in more industries and scenarios.

[0003] Although target tracking algorithms on traditional computer platforms have made significant progress in processing speed, anti-interference capabilities, and accuracy, their application in practical environments such as embedded platforms remains limited. Embedded platforms often face resource constraints, so the design of target tracking algorithms must not only consider the computational complexity of the algorithm itself but also match the computing power, storage capacity, and power consumption constraints of the hardware platform. Furthermore, with the increasing popularity of high-definition video applications, the real-time processing of high-definition images and videos has become increasingly prominent, placing more stringent demands on the computing performance, data transmission bandwidth, and real-time processing capabilities of the processing hardware. Therefore, how to efficiently implement target tracking and image processing on resource-constrained embedded platforms has become a hot topic and a major challenge in current research.

[0004] FPGAs (Field Programmable Gate Arrays) are ideal platforms for image processing applications such as target tracking due to their high customizability, powerful parallel computing capabilities, and low power consumption. FPGAs enable hardware-level parallel processing, significantly improving data processing speed and allowing for optimization based on application requirements. In particular, FPGAs offer low latency and high-speed real-time processing for high-resolution image and video processing, meeting the demands of applications requiring high real-time performance and accuracy. The ORB (Oriented FAST and Rotated BRIEF) algorithm is an efficient feature point detection and description algorithm widely used in image matching and target tracking. Combining FAST feature point detection with BRIEF descriptors, ORB offers low computational complexity and high processing efficiency, making it suitable for resource-constrained environments. ORB's ability to stably extract and match feature points and its robustness under conditions such as occlusion and changing lighting conditions make it widely used in dynamic target tracking and image matching. Deploying the ORB algorithm on an FPGA platform leverages the FPGA's advantages in parallel processing and hardware acceleration. Hardware-level optimization of key steps in the ORB algorithm (such as feature point detection, descriptor generation, and matching) significantly improves processing speed, reduces latency, and achieves low-power operation. However, the ORB matching tracking algorithm has certain limitations. Its tracking module lacks a real-time update mechanism and is easily affected by noise and environmental changes, resulting in reduced tracking accuracy or mismatches. Summary of the Invention

[0005] The present invention aims to provide an FPGA-based ORB feature matching tracking method that effectively migrates complex ORB feature-based target tracking algorithms to an FPGA platform. This method fully considers the resource limitations of embedded platforms and can meet the requirements of real-time processing of large amounts of image data. By introducing a dynamic update mechanism, the target feature template is updated in real time, removing outdated features and introducing new valid features, thereby ensuring the continued validity of the tracking template and significantly improving the accuracy and reliability of tracking matching.

[0006] The FPGA-based ORB feature matching and tracking method of the present invention includes video image format conversion, FAST feature point detection, feature point descriptor generation, feature point storage, target feature point update, coordinate mapping module and display control module, including the following steps:

[0007] Step 1: Video image format conversion: the color video format data input by the image acquisition module is converted into a grayscale image after three frames of buffering. The color video format data is used for the display control module, and the grayscale video image data is used for the subsequent ORB feature point matching and tracking algorithm.

[0008] Step 2, FAST (Features from Accelerated Segment Test) feature point detection, generate a 7*7 sliding pixel matrix; perform FAST corner point detection in the 7*7 sliding pixel matrix to determine whether the center pixel of the matrix is ​​a FAST corner point; calculate the FAST response value of the center pixel in the 7*7 sliding pixel matrix; in the 7*7 sliding pixel matrix, according to the FAST corner point enable signal generated by the FAST corner point detection submodule and the FAST response value generated by the FAST response value calculation submodule, perform maximum value suppression and output the final FAST feature point valid signal; this process adopts the FPGA pipeline processing mechanism. Whenever there is a pixel input, the FAST feature point detection module will update the 7*7 sliding pixel matrix in parallel at each stage, and generate the corresponding FAST corner point enable signal and FAST response value at the same time, and finally output the FAST feature point valid signal through maximum value suppression;

[0009] Step 3: Generate feature point descriptors and generate a 31*31 sliding pixel matrix. In the 31*31 sliding pixel matrix, select 256 pairs of pixels and compare their sizes to generate a 256-bit BRIEF (Binary Robust Independent Elementary Features) descriptor. Combine the 32-bit pixel coordinate positions generated by the row and field counters and the 256-bit BRIEF descriptor to generate a 288-bit feature point descriptor, and the feature point valid signal output by the FAST feature point detection module enables the output of the feature point descriptor. This process makes full use of the pipeline processing mechanism and parallelization characteristics of the FPGA. Whenever there is a pixel input, the feature point descriptor module will generate the corresponding feature point descriptor in real time and work in parallel with the FAST feature point detection module. The output of the descriptor is controlled by the feature point valid signal, and the result is finally written into the current feature point memory.

[0010] Step 4, including feature point storage and feature point matching and updating; the feature point storage includes three memories, namely the target feature point memory, the current feature point memory and the next feature point memory; the feature point data detected in the first frame will be stored in the current feature point memory, and then transferred to the target feature point memory as the target feature point data; the feature point data detected in the subsequent frames will first be stored in the current feature point memory, and then feature point matching will be performed with the target feature point memory, and the feature point data with successful matching will be stored in the next feature point memory; after the matching is completed, the data in the next feature point memory will be transferred to the target feature point memory to update the target feature point data;

[0011] Feature point matching update, including feature point matching degree traversal calculation and target feature point template matching update; by dynamically eliminating outdated feature points and introducing new data, the real-time validity of the feature point template is maintained, the matching accuracy is improved, and the tracking loss problem caused by excessive target changes is solved; the feature point matching degree traversal calculation is responsible for controlling the read and write operations of the target feature point memory, the current state feature point memory, and the next state feature point memory, and performing traversal calculation of the feature point matching degree; according to the relationship between the calculated matching degree and the preset matching degree threshold, the target feature point matching update part adopts different matching strategies and updates the target feature point memory; through the synergy of storage and matching update, the feature point template can be updated in real time, significantly improving matching accuracy and robustness;

[0012] Step 5: Coordinate mapping: receiving the successfully matched feature point position data from the feature point matching module and mapping these position data to the horizontal and vertical directions respectively, thereby calculating the coordinate positions of the four corners of the tracking frame;

[0013] Step 6: Display control, which receives the color video format data input from the image acquisition module and the tracking frame coordinate information from the coordinate mapping module, draws the corresponding tracking frame on the color video and outputs the result.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] 1. Video image data uses three-frame buffering to improve data stability and solve the problem of screen tearing;

[0016] 2. The non-blocking algorithm based on stream processing realizes real-time processing and algorithm acceleration of large-scale video image data by performing parallel calculations while pixel data is input;

[0017] 3. The target feature point information can be updated in real time, that is, outdated feature points caused by noise are removed and new target feature points are added, thereby ensuring the real-time validity of the feature template and improving the matching reliability during the tracking process. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Flowchart of ORB feature point matching and tracking method

[0019] Figure 2 Detailed structural diagram of the ORB feature point matching and tracking method

[0020] Figure 3 Traverse the matching flowchart for feature points

[0021] Figure 4 Update the flow chart for the target feature points DETAILED DESCRIPTION

[0022] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings.

[0023] The ORB feature matching tracking method based on FPGA, the ORB feature point matching tracking method flow chart is as follows Figure 1 As shown, the detailed structure diagram of the ORB feature point matching and tracking method is as follows Figure 2 As shown, it includes the following steps:

[0024] Step 1: Convert the video format image. The color video format data input by the image acquisition module is divided into two channels after three frames of buffering. One channel is used for grayscale conversion for subsequent ORB feature point matching and tracking algorithm, and the other channel is used for superimposed display of tracking frames.

[0025] Original formula for RGB to YCbCr:

[0026] Y=0.299R+0.587G+0.114B

[0027] Cb=0.568(BY)+128=-0.172R-0.339G+0.511B+128

[0028] Cr=0.713(RY)+128=0.511R-0.428G+128

[0029] Since floating-point operations cannot be used directly in FPGAs and occupy a high number of resources, fixed-point operations are used instead of floating-point operations for optimization. The formula is multiplied by 256 and then shifted to the right by 8 bits. The formula is as follows:

[0030] Y=(77*R+150*G+29*B)>>8

[0031] Cb=(-43*R-85*G+128*B+32768)>>8

[0032] Cr=(128*R-107*G-21*B+32768)>>8

[0033] The above R, G, and B correspond to the red, green, and blue components of the three primary color model respectively, and Y, Cb, and Cr correspond to the brightness, blue chromaticity, and red chromaticity of the YCbCr color model respectively;

[0034] Step 2: FAST feature point detection generates a 7*7 sliding pixel matrix. Within the 7*7 sliding pixel matrix, the absolute value of the difference between the center pixel and the 16 surrounding pixels is calculated and compared with a preset threshold. When the absolute value of the difference between nine consecutive pixels is greater than the threshold, the center pixel is identified as a valid FAST corner point. By setting a reasonable threshold, noise points are effectively filtered out and corner points with significant features are detected.

[0035] The FAST response value is an important indicator for evaluating the quality of FAST corner points. In a 7*7 sliding pixel matrix, the absolute value of the difference between the central pixel and the 16 surrounding pixels is calculated, and the sum of the absolute values ​​of these differences is used as the FAST response value of the central pixel. This score reflects the significance and stability of the corner point and provides a basis for subsequent maximum value suppression processing.

[0036] To avoid overcrowding of corner points in an image, which could affect subsequent processing efficiency, the FAST Maximum Suppression submodule suppresses feature points within a 7×7 sliding window. This submodule combines the feature point enable signal output by the FAST Corner Detection submodule with the FAST score output by the FAST Response Value Calculation submodule. Through maximum suppression, it removes redundant feature points, ensuring a more balanced distribution of feature points, reducing computational effort, and improving matching robustness and accuracy. Through this mechanism, the FAST Feature Point Detection module not only achieves efficient local feature point detection but also provides reliable support for subsequent feature matching and descriptor generation, enhancing overall image processing performance.

[0037] Step 3: Generate feature point descriptors. Generate a 31*31 sliding pixel matrix. This module preselects 256 pairs of pixel coordinates in the window matrix. Whenever a new pixel flows into the sliding pixel matrix, it extracts the corresponding 256 pairs of pixel values ​​and compares their grayscale values. If the grayscale value of the former is greater than the latter, it is recorded as 1, otherwise it is recorded as 0. Finally, a 256-bit binary vector is generated as the BRIEF descriptor of the center pixel of the sliding pixel matrix.

[0038] To accurately identify the pixel's coordinate position in the image, the module dynamically calculates the horizontal and vertical coordinates of the current pixel using row and column counters. When a new pixel flows into the sliding pixel matrix, the horizontal and vertical coordinates need to be increased by 15 to ensure that the coordinate values ​​describe the BRIEF descriptor of the center pixel of the sliding pixel matrix, considering that the pixel needs to pass through 15 rows of pixels plus 15 pixels to reach the center of the sliding pixel matrix.

[0039] Finally, the 256-bit BRIEF descriptor generated by the 31*31 sliding pixel matrix is ​​combined with the coordinate position information provided by the row and column counters to form a complete feature point descriptor. Whether the feature point descriptor is output is controlled by the valid enable signal provided by the FAST feature point module to ensure that descriptors are generated and output only for valid feature points.

[0040] Step 4, feature point matching and updating, including feature point storage and feature point matching and updating; feature point storage includes three memories, namely, target feature point memory, current feature point memory and next feature point memory; the feature point data detected in the first frame will be stored in the current feature point memory and then transferred to the target feature point memory as target feature point data; the feature point data detected in subsequent frames will first be stored in the current feature point memory and then feature point matched with the target feature points in the target feature point memory, and the feature point data of the successfully matched feature points will be stored in the next feature point memory; after the matching is completed, the data in the next feature point memory will be transferred to the target feature point memory to update the target feature point data;

[0041] When calculating the matching degree of feature points through traversal, the feature point data in the target feature point memory and the feature point data in the current feature point memory are traversed and the Hamming distance is calculated. The matching degree between the two feature points with lower Hamming distance is higher. The two feature points with the highest matching degree in the current feature memory are selected as candidate feature points, and the feature points in the target feature point memory corresponding to them are called source feature points. Feature point matching and target feature point update operations are performed based on the matching degree of the two candidate feature points and the preset high matching degree threshold and low matching degree threshold.

[0042] During the feature point data update process, if the matching degree of both candidate feature points is lower than the low matching degree threshold, the matching is rejected; if the matching degree of only one candidate feature point is higher than the high matching degree threshold, only this feature point is matched and its data is written to the next-state feature point memory; if the matching degree of both candidate feature points is lower than the high matching degree threshold and higher than the low matching degree threshold, the feature point data in the current target feature point memory is read and written to the next-state feature point memory; if the matching degree of both candidate feature points is higher than the high matching degree threshold and the difference is large, the forward matching phase is entered;

[0043] In the forward matching stage, the two candidate feature points are traversed with the feature point data in the target feature point memory to calculate the matching degree, and the two feature points with the highest matching degree are selected; if any of the two feature points with the highest matching degree is the source feature point, the candidate feature point is matched and written to the secondary feature point memory; otherwise, the source feature point data is written to the matching feature point memory; after traversing the target feature point memory, the data in the secondary feature point memory is transferred to the target feature point memory to complete the target feature point update operation;

[0044] Step 5: Coordinate mapping: Receive the position data of successfully matched feature points from the feature point matching module and map these position data to the horizontal and vertical directions respectively. The module calculates the coordinate positions of the four corners of the tracking frame by screening the maximum and minimum horizontal and vertical coordinate positions of the matched feature points and resets this module after each frame of video transmission cycle ends.

[0045] Step 6: Display control is used to draw a tracking frame on the color video stream to display the tracking position output by the coordinate mapping module. This module calculates the coordinates of the input pixel in real time through the row and field counters. When the pixel is located on the tracking frame, its output value is replaced with 24'hff0000 to achieve a highlighted display effect.

[0046] In the above step 4, in order to improve the accuracy of feature point matching and solve the problem of tracking loss caused by large changes in the tracking target, an improved feature point matching and updating mechanism is proposed; through the feature point matching and updating process, outdated feature point data is dynamically eliminated and new feature point data is introduced, thereby maintaining the real-time validity of the feature point template and improving the reliability of matching during the tracking process; its feature point traversal matching flow chart is as follows Figure 3 As shown, the target feature point update flow chart is as follows Figure 4 As shown; including the following steps:

[0047] Step 1: When the pixel input position exceeds the search area of ​​the target image, the feature point storage module notifies the feature point matching module through a handshake signal that the detection is complete and starts the feature point matching degree traversal calculation; read the i-th feature point P from the target feature point memory i , and take out the feature point data one by one from the current feature point memory and send it to the matching degree calculation module, and compare it with the current target feature point P i Calculate the matching degree and select the first two feature points with the highest matching degree as candidate feature points, which are recorded as P1 and P2 respectively. The corresponding matching degrees are recorded as ξ1 and ξ2 respectively. After the traversal is completed, reset the matching degree calculation module and wait for the feature point update to be completed.

[0048] Step 2: Preset a high matching threshold Tr1 and a low matching threshold Tr2; and adopt the following update strategy based on the results between the matching degree of the candidate feature point and the high matching threshold and the low matching threshold:

[0049] If ξ1 <T r2 ,ξ2 <T r2 , then refuse to match the feature point and delete the current feature point data;

[0050] If ξ1>T r1 ,ξ2 <T r2, then P1 is selected as the matching feature point, its feature point data is written into the secondary feature point memory, and the coordinate position is output to the coordinate mapping module;

[0051] If T r2 <ξ1 <T r1 , T r2 <ξ2 <T r1 , then retain P i The feature point data is written into the secondary feature point memory, but the coordinate position is not output to the coordinate mapping module;

[0052] If ξ1>T r1 ,ξ2>T r1 , and ξ2 / ξ1 <T r1 / 2, then select P1 as the matching feature point, write its feature point data into the secondary feature point memory, and output the coordinate position to the coordinate mapping module;

[0053] If ξ1>T r1 ,ξ2>T r1 , and ξ2 / ξ1>T r1 / 2, then execute step 3 for further judgment;

[0054] Step 3: Forward search: P1 and P2 are traversed and matched with the feature points in the target feature point memory one by one, and the top two feature points P with the highest matching degree are selected. 11 、P 12 and P 21 、P 22 As their candidate feature points, according to whether their candidate feature points contain P i , adopt the following update strategy:

[0055] If P 11 , P 12 Contains P i , and P 21 , P 22 Does not contain P i , then P1 is selected as the matching feature point, and its data is written into the secondary feature point memory, and the coordinate position is output to the coordinate mapping module;

[0056] If P 21 、P 22 Contains P i , and P 11 、P 12 Does not contain P i , then P2 is selected as the matching feature point, and its data is written into the secondary feature point memory, and the coordinate position is output to the coordinate mapping module;

[0057] If P 11 , P12 , P 21 , P 22 Does not contain P i , then retain P i As matching feature points, the data is written into the matching feature point memory, and the coordinate position is output to the coordinate mapping module;

[0058] If P 11 、P 12 and P 21 、P 22 P i , then P1 and P2 are selected as matching feature points at the same time, and their data are written into the matching feature point memory, and the coordinate positions are output to the coordinate mapping module;

[0059] After the feature point traversal and update are completed, the feature point data in the matching feature point memory is transferred to the target feature point memory to complete the update of the target feature point;

[0060] Step 4: After the target feature point is updated, determine whether the target feature point memory has been traversed. If not, return to the target feature point traversal state and repeat steps 2 and 3. If the traversal is completed, enter the traversal end state and complete the target feature point template update.

Claims

1. The ORB feature matching tracking method based on FPGA includes the following steps: Step 1: Video image format conversion: the color video format data input by the image acquisition module is converted into a grayscale image after three-frame buffering. The color video format data is used for the display control module, and the grayscale video image data is used for the subsequent ORB feature point matching and tracking algorithm. Step 2, FAST feature point detection, generate a 7*7 sliding pixel matrix; perform FAST corner point detection in the 7*7 sliding pixel matrix to determine whether the center pixel of the matrix is ​​a FAST corner point; calculate the FAST response value of the center pixel in the 7*7 sliding pixel matrix; in the 7*7 sliding pixel matrix, according to the FAST corner point enable signal generated by the FAST corner point detection submodule and the FAST response value generated by the FAST response value calculation submodule, perform maximum value suppression, and output the final FAST feature point valid signal; this process adopts the FPGA pipeline processing mechanism. Whenever there is a pixel input, the FAST feature point detection module will update the 7*7 sliding pixel matrix in parallel at each stage, and generate the corresponding FAST corner point enable signal and FAST response value at the same time, and finally output the FAST feature point valid signal through maximum value suppression; Step 3: Generate feature point descriptors to generate a 31*31 sliding pixel matrix; select 256 pairs of pixels in the 31*31 sliding pixel matrix and compare the sizes to generate a 256-bit BRIEF descriptor; combine the 32-bit pixel coordinate position generated by the row and field counters and the 256-bit BRIEF descriptor to generate a 288-bit feature point descriptor, and enable the output of the feature point descriptor by the feature point valid signal output by the FAST feature point detection module; this process makes full use of the pipeline processing mechanism and parallelization characteristics of the FPGA. Whenever there is a pixel input, the feature point descriptor module will generate the corresponding feature point descriptor in real time, and work in parallel with the FAST feature point detection module, and control the output of the descriptor through the feature point valid signal, and finally write the result into the current state feature point memory; Step 4, including two parts: feature point storage and feature point matching update; The feature point storage includes three memories, namely, the target feature point memory, the current feature point memory and the next feature point memory; the feature point data detected in the first frame will be stored in the current feature point memory, and then transferred to the target feature point memory as the target feature point data; the feature point data detected in the subsequent frames will first be stored in the current feature point memory, and then feature point matching will be performed with the target feature point memory, and the feature point data with successful matching will be stored in the next feature point memory; After the matching is completed, the data in the secondary feature point memory will be transferred to the target feature point memory, and the target feature point data will be updated; Feature point matching update, including feature point matching degree traversal calculation and target feature point template matching update; By dynamically removing outdated feature points and introducing new data, the real-time validity of the feature point template is maintained, the matching accuracy is improved, and the tracking loss problem caused by large changes in the target is solved; The feature point matching degree traversal calculation is responsible for controlling the read and write operations of the target feature point memory, the current state feature point memory and the next state feature point memory, and performing the traversal calculation of the feature point matching degree; according to the relationship between the calculated matching degree and the preset matching degree threshold, the target feature point matching update part adopts different matching strategies and updates the target feature point memory; Through the synergy of storage and matching update, the feature point template can be updated in real time, significantly improving the matching accuracy and robustness; Step 5: coordinate mapping, receiving the successfully matched feature point position data from the feature point matching module, and mapping these position data to the horizontal direction and the vertical direction respectively, so as to calculate the coordinate positions of the four corners of the tracking frame; Step 6: Display control, receiving the color video format data input from the image acquisition module and the tracking frame coordinate information from the coordinate mapping module, drawing the corresponding tracking frame on the color video and outputting the result.

2. The FPGA-based ORB feature matching tracking method according to claim 1 is characterized in that: The feature point matching update includes two parts: feature point matching degree traversal calculation and target feature point template update; target feature point update is added in the feature point matching process, outdated feature point data is dynamically eliminated, and new feature point data is introduced, so as to maintain the real-time validity of the feature point template, improve the feature point matching accuracy and solve the tracking loss problem caused by excessive changes in the tracking target.

3. The feature point matching degree traversal calculation and target feature point template update according to claim 2, characterized in that: The following steps are included: Step 1: When the pixel input position exceeds the search area of ​​the target image, the feature point storage module notifies the feature point matching module of the completion of the detection through a handshake signal, and starts the traversal calculation of the feature point matching degree; Read the i-th feature point Pi from the target feature point memory, and take out the feature point data one by one from the current feature point memory and send them to the matching degree calculation module, calculate the matching degree with the current target feature point Pi, and select the first two feature points with the highest matching degree as candidate feature points, which are recorded as P1 and P2 respectively, and the corresponding matching degrees are recorded as ξ1 and ξ2 respectively; after the traversal is completed, reset the matching degree calculation module and wait for the feature point update to be completed; Step 2: Preset high matching threshold T r1 and low matching threshold T r2 ; According to the results between the matching degree of the candidate feature points and the high matching degree threshold and the low matching degree threshold, the following update strategy is adopted: If ξ1 <T r2 ,ξ2 <T r2 , then refuse to match the feature point and delete the current feature point data; If ξ1>T r1 ,ξ2 <T r2 , then P1 is selected as the matching feature point, its feature point data is written into the secondary feature point memory, and the coordinate position is output to the coordinate mapping module; If T r2 <ξ1 <T r1 , T r2 <ξ2 <T r1 , then keep P i The feature point data is written into the secondary feature point memory, but the coordinate position is not output to the coordinate mapping module; If ξ1>T r1 ,ξ2>T r1 , and ξ2-ξ1 <T r1 / 2, then select P1 as the matching feature point, write its feature point data into the secondary feature point memory, and output the coordinate position to the coordinate mapping module; If ξ1>T r1 ,ξ2>T r1 , and ξ2-ξ1>T r1 / 2, then execute step 3 for further judgment; Step 3: forward search: P1 and P2 are traversed and matched with the feature points in the target feature point memory one by one, and the first two feature points P with the highest matching degree are selected. 11 , P 12 and P 21 , P 22 As their candidate feature points, according to whether their candidate feature points contain P i , adopt the following update strategy: If P 11 , P 12 Contains P i , and P 21 , P 22 Does not contain P i , then P1 is selected as the matching feature point, and its data is written into the secondary feature point memory, and the coordinate position is output to the coordinate mapping module; If P 21 , P 22 Contains P i , and P 11 , P 12 Does not contain P i , then P2 is selected as the matching feature point, and its data is written into the secondary feature point memory, and the coordinate position is output to the coordinate mapping module; If P 11 , P 12 , P 21 , P 22 None of them contain P i , then keep P i As matching feature points, the data is written into the matching feature point memory, and the coordinate position is output to the coordinate mapping module; If P 11 , P 12 and P 21 , P 22 P i , then P1 and P2 are selected as matching feature points at the same time, and their data are written into the matching feature point memory, and the coordinate positions are output to the coordinate mapping module; After the feature point traversal update is completed, the feature point data in the matching feature point memory is transferred to the target feature point memory to complete the update of the target feature point; Step 4: After the target feature point is updated, determine whether the target feature point memory has been traversed; if not, return to the target feature point traversal state and repeat steps 2 and 3; if the traversal is completed, enter the traversal end state and complete the target feature point template update.

Images