A method and storage medium for target matching based on feature points and convolution optimization calculation
Through the method based on SURF feature point and convolution optimization calculation, the accuracy and efficiency of target recognition in the prior art are solved, and efficient target recognition without relying on big data models is achieved, so as to accurately identify the position, angle and scale of the target. .
Patent Information
- Application Number
- CN202210181610.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-25
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-02-25
AI Technical Summary
Existing deep learning-based object recognition technology relies on big data model training, collecting and labeling data is time-consuming and laborious, and it is impossible to accurately identify the real area, angle and scaling scale of the target.
The target matching method based on SURF feature points and convolution optimization calculation is adopted. The target features are automatically extracted by selecting the image area at one time, and the image pyramid is constructed. The feature points matching and optimization are carried out, and the attention mechanism is introduced. Perform convolutional calculations to obtain the position, angle and scale information of the target.
It realizes efficient target recognition without relying on big data model training, and can accurately identify the actual area, rotation angle and scaling scale of the target, with high robustness and accuracy.
Smart Images

Figure CN114549861B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method for target recognition, in particular to a method and a storage medium for target matching based on feature points and convolution optimization calculation. Background Art
[0002] Currently, the most eye-catching target recognition technology based on deep learning relies on big data model training. The quality of the model depends to a large extent on the training data set. The collection and labeling of the training data set is a time-consuming and laborious task. In addition, the general target recognition deep learning model can only locate the target in a regular rectangular frame, and cannot accurately locate the real area of the target. At the same time, it is also impossible to accurately identify the angle and scaling of the target. Summary of the invention
[0003] The present invention provides a method for target matching based on feature points and convolution optimization calculation, which aims to not rely on big data model training, and can automatically complete the extraction of selected target features and the construction of image pyramid by only selecting a target in a certain area of an image at one time, and can quickly complete the full-image target matching for the newly collected image, and output the position, scale, and angle information of the matching target. Based on the characteristics of SURF features that remain invariant to rotation, scale transformation, and brightness, and the introduction of the guiding convolution calculation of the attention mechanism, an efficient algorithm that can perform target matching well in multiple dimensions such as position, angle, and scale is proposed.
[0004] In order to achieve the above objectives, the present invention provides a method for target matching based on feature points and convolution optimization calculation, and the specific steps are as follows:
[0005] A method for target matching based on feature points and convolution optimization calculation, comprising the following steps:
[0006] (1) Select a template image Temp, which contains the target to be identified and has no cluttered background.
[0007] (2) Extract the SURF feature feature1 of the Temp image, and construct the Temp image pyramid based on downsampling, Gaussian smoothing, and rotation;
[0008] (3) Extract SURF feature feature2 from the matching image Src;
[0009] (4) Use FLANN matcher to match feature points of feature1 and feature2;
[0010] (5) If the number of matching feature points obtained in step (4) is greater than the set threshold, feature point calculation is performed, and a set of scale, angle, and coordinate information of the target matched by Src is calculated based on the matched feature points, and ROI extraction is performed on the matching target area according to the calculation results of the feature points;
[0011] (6) Using feature 1 extracted in step (2) as the attention mechanism, guide the image pyramid constructed in step (2) to perform convolution calculation on the ROI to obtain a new set of scale, angle, and coordinate information;
[0012] (7) Calibrate and check the two sets of scale, angle, and coordinate information obtained in step (5) and step (6) to obtain a final set of results;
[0013] (8) If the number of matching feature points obtained in step (4) is less than the set threshold, jump directly to step (6), use feature1 extracted in step (2) as the attention mechanism, guide the image pyramid constructed in step (2) to perform convolution calculation on Src, obtain a set of scale, angle, and coordinate information, and use this set of information as the final set of results.
[0014] Preferably, the method for calculating the feature points comprises the following steps:
[0015] (a) Based on the FLANN matcher, the matching feature points MatchPoint of feature1 and feature2 are obtained. The elements in the MatchPoint exist in pairs. The transformation matrix H1 is calculated based on the MatchPoint. The transformation matrix H1 is a 3-row × 3-column transformation matrix. The transformation matrix H1 is used to express the mapping relationship between the paired feature points in the MatchPoint. When calculating the transformation matrix H1, the RANSAC algorithm is used to remove some erroneous matching points.
[0016] (b) Calculate the four mapping points of the four vertices of Temp in Src according to the transformation matrix H1. The four mapping points are connected to form a closed area. The content of the area is the matching target range of Temp in Src. Then, the matching feature points are screened according to the four mapping points to further eliminate the erroneous matching points outside the matching target range.
[0017] (c) sorting the further screened matching feature points in descending order of similarity, and selecting the top n pairs of matching feature points in the sorting results;
[0018] (d) Using the optimization algorithm to optimize the first n pairs of matching feature points, we can find 6 pairs of optimal feature points;
[0019] (e) After finding the best 6 pairs of matching feature points, recalculate a more accurate transformation matrix H2, and update the matching target range of Temp in Src based on the new transformation matrix H2;
[0020] (f) intercepting the optimized matching target range as the target ROI;
[0021] (g) Select 2 pairs of feature points from the 6 pairs of matching feature points and calculate the rotation angle and scaling scale of the matching target.
[0022] Preferably, during the optimization operation, taking the simulated annealing algorithm as an example, finding several pairs of optimal matching feature points from n pairs of matching feature points, the optimization steps are as follows:
[0023] (a) 7 pairs of matching feature points are ,in, Represents the first feature points, Represents the first feature points, and define the solution space as ,in , that is, 3 pairs of feature points;
[0024] (b) The objective function formula is as follows:
[0025]
[0026]
[0027]
[0028]
[0029]
[0030] in, Representing feature points and The Euclidean distance between
[0031] (c) The new solution is generated by randomly selecting two pairs from the remaining feature point pairs:
[0032] The two pairs in , the new solution is as follows:
[0033]
[0034] (d) The objective function difference is the difference between the solution before the transformation and the objective function after the transformation. The formula is as follows:
[0035]
[0036] (e) The acceptance criterion for the new solution follows the Metropolis acceptance criterion, which is based on the difference in the objective function and the probability
[0037] accept
[0038] As the new current solution
[0039] , the acceptance criterion formula is as follows:
[0040]
[0041] (f) There are two termination conditions for the optimization algorithm. Condition 1 is to find the optimal solution, and condition 2 is that the current solution satisfies the following conditions:
[0042] .
[0043] Preferably, the method of selecting 2 pairs of feature points from 6 pairs of matching feature points and calculating the rotation angle of the matching target is as follows:
[0044] (a) Assume that the two pairs of feature points are selected as ;
[0045] (b) By Calculate the angle angle1 between the line connecting the two points and the horizontal coordinate system, by Calculate the angle angle2 between the line connecting the two points and the horizontal coordinate system;
[0046] (c) The rotation angle of the matching target compared to the template image can be obtained through angle1 and angle2.
[0047] Preferably, the method of selecting 2 pairs of feature points from 6 pairs of matching feature points and calculating the scaling scale of the matching target is as follows:
[0048] (a) Assume that the two pairs of feature points are selected as ;
[0049] (b) Obtain the scaling scale of the matching target according to the following formula:
[0050]
[0051] In the formula, Representing feature points and The Euclidean distance between .
[0052] Preferably, the feature 1 extracted in step (2) is used as an attention mechanism to guide the image pyramid constructed in step (2) to perform convolution calculation, wherein the convolution kernel of the convolution calculation is the image pyramid constructed by Temp, and the convolution is directly performed on Src or ROI, and the convolution is accelerated by parallel operation, and the convolution result is the similarity between Temp and a partial area of Src or ROI;
[0053] A new set of scale, angle, and coordinate information is obtained through convolution calculation, and calibrated and checked with the first set of scale, angle, and coordinate information obtained based on SURF feature points.
[0054] Preferably, the method of the attention mechanism includes: the role of the attention mechanism is to guide the convolution calculation, with the SURF feature point of Temp as the main factor. During the convolution calculation, the convolution value within the 8*8 window of the feature point is weighted, thereby reducing the influence of factors such as noise and enhancing the robustness of the convolution result.
[0055] Furthermore, it also includes multi-target matching, and the multi-target matching method includes:
[0056] (a) When the background of the matching target image is relatively simple, multiple target regions Objs can be extracted through simple Blob analysis, and then SURF features are extracted from Objs respectively based on parallel computing. The subsequent processing flow is the same as described above;
[0057] (b) When the background of the target image is relatively complex, the SURF feature of Temp is directly used as the attention mechanism to guide the template image pyramid to perform convolution calculation on Src, and a set of scale, angle, and coordinate information is obtained as the final result.
[0058] A device for a target matching method based on feature points and convolution optimization calculation, comprising: a memory for storing a computer program; a processor for executing the computer program to implement any of the above-mentioned target matching methods based on feature points and convolution optimization calculation.
[0059] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for target matching based on feature points and convolution optimization calculation described in any one of the above items is implemented.
[0060] It can be seen from the above description of the present invention that, compared with the prior art, the present invention has the following beneficial effects:
[0061] (1) The method of the present invention selects SURF feature points that remain invariant to rotation, scale transformation, and brightness to process the image, and iteratively optimizes the matching feature points, especially introducing the simulated annealing method to obtain the optimal feature point pair, thereby obtaining accurate scale, angle, and coordinate information. At the same time, the SURF feature points are used as an attention mechanism to guide the convolution calculation, and the extracted ROI area is secondary processed to obtain a second set of scale, angle, and coordinate information for calibration and verification, and finally the most accurate multi-dimensional information is obtained. The algorithm has extremely high robustness and accuracy.
[0062] (2) The method of the present invention does not rely on big data model training. It can automatically complete the extraction of target features and the construction of an image pyramid by only selecting a target in a certain area of an image at one time. It can quickly complete high-precision recognition of targets in the entire image for newly collected images. It can not only accurately locate the actual area where the target is located, but also effectively identify the zoom scale and rotation angle of the target. BRIEF DESCRIPTION OF THE DRAWINGS
[0063] Figure 1 It is a flow chart of a target matching method based on feature points and convolution optimization calculation;
[0064] Figure 2 Schematic diagram of SURF features extracted from a template image in an embodiment;
[0065] Figure 3 Schematic diagram of SURF features extracted from the image to be matched in the embodiment;
[0066] Figure 4 Schematic diagram of mathematical modeling of simulated annealing method in the embodiment;
[0067] Figure 5 Schematic diagram of the matching target area before and after updating in the embodiment;
[0068] Figure 6 Schematic diagram of calculating the rotation angle and scaling scale of the matching target in the embodiment;
[0069] Figure 7 is a schematic diagram of a multi-objective matching algorithm in an embodiment;
[0070] Figure 8 Schematic diagram of target matching results in the embodiment. DETAILED DESCRIPTION
[0071] The present invention will be further described with reference to the accompanying drawings.
[0072] like Figures 1 to 8 As shown in the figure, a target matching method based on feature points and convolution optimization calculation is shown in the flowchart. Figure 1 As shown, the specific steps include:
[0073] (1) Select a template image Temp. Temp is generally an image of a region that contains only the target to be identified and tries not to contain cluttered background. The target is correctly placed and not skewed. Figure 2 As shown in the target;
[0074] (2) Extract the SURF feature feature1 of the Temp image. The extracted feature effect is as follows: Figure 2 As shown, the image pyramid of Temp is constructed based on downsampling, Gaussian smoothing and rotation;
[0075] (3) Extract SURF feature feature2 from the matching image Src. The extracted feature effect is as follows Figure 3 As shown;
[0076] (4) Use FLANN matcher to match feature points of feature1 and feature2;
[0077] (5) If the number of matching feature points obtained in step (4) is greater than the set threshold of 10, feature point calculation is performed. Based on the matched feature points, a set of scale (Scale), angle (angle_final), and coordinate (target center point X, Y value) information of Src matching the target is calculated, and ROI extraction is performed on the matching target area according to the calculation results of the feature points;
[0078] The specific process of feature point calculation is as follows:
[0079] (a) Based on the FLANN matcher, the matching feature points MatchPoint of feature1 and feature2 are obtained. The elements in MatchPoint exist in pairs. According to MatchPoint, the transformation matrix H1 is calculated. The transformation matrix H1 is a 3-row × 3-column transformation matrix that expresses the mapping relationship between paired feature points in MatchPoint. In addition, when calculating the transformation matrix H1, the RANSAC algorithm is used to eliminate some erroneous matching points;
[0080] (b) Calculate the four mapping points of the four vertices of Temp in Src according to the transformation matrix H, where the coordinates of the four vertices of Temp are:
[0081] ,
[0082] is the width of the image Src to be matched, is the height of the image to be matched Src. These four mapping points are connected to form a closed area, and the content of this area is the matching target range of Temp in Src. The matching feature points are screened according to the four mapping points to further eliminate the wrong matching points outside the matching target range.
[0083] (c) sorting the further screened matching feature points in descending order of similarity, and selecting the top n pairs of matching feature points in the sorting results;
[0084] (d) Use the optimization algorithm to optimize the first n pairs of selected matching feature points and find 6 pairs of optimal feature points.
[0085] (e) After finding the best 6 pairs of matching feature points, a more accurate transformation matrix H2 is recalculated, and the matching target range of Temp in Src is updated based on the new transformation matrix H2. The schematic diagram before and after the update is shown in the figure. Figure 5 As shown, the dotted line framed area is the matching target range before optimization, and the solid line framed area is the matching target range after optimization;
[0086] (f) intercepting the optimized matching target range as the target ROI;
[0087] (g) Select 2 pairs of feature points from the 6 pairs of matching feature points and calculate the rotation angle and scaling of the matching target, such as Figure 6 As shown;
[0088] The optimization algorithm selects simulated annealing method. Taking the simulated annealing algorithm as an example, we take the search for 3 pairs of optimal matching feature points from 7 pairs of matching feature points as an example. The distribution of points is as follows: Figure 4 As shown in the figure, the mathematical model established for this optimization problem is specifically described, and the process is as follows:
[0089] (a) 7 pairs of matching feature points are ,in, Represents the first feature points, Represents the first feature points, and define the solution space as ,in , that is, 3 pairs of feature points.
[0090] (b) The objective function formula is as follows:
[0091]
[0092]
[0093]
[0094]
[0095]
[0096] in, Representing feature points and The Euclidean distance between .
[0097] (c) The new solution is generated by randomly selecting two pairs of feature point pairs from the remaining pairs to replace The two pairs in , the new solution is as follows:
[0098]
[0099] (d) The objective function difference is the difference between the solution before the transformation and the objective function after the transformation. The formula is as follows:
[0100]
[0101] (e) The acceptance criterion for the new solution follows the Metropolis acceptance criterion, which is based on the difference and probability of the objective function. accept As the new current solution , the acceptance criterion formula is as follows:
[0102]
[0103] (f) There are two termination conditions for the optimization algorithm. Condition 1 is to find the optimal solution, and condition 2 is that the current solution satisfies the following conditions:
[0104]
[0105] Figure 4 The points on the two pictures represent the extraction of a feature point. S1 to S7 in scr are the seven feature points extracted from the image to be matched, and T1 to T7 in temp are the seven feature points on the template image. The feature points are paired, for example, T1 and S1, these two points are corresponding, and S2 and T2 are corresponding.
[0106] The specific process of selecting 2 pairs of feature points from 6 pairs of matching feature points and calculating the rotation angle of the matching target is as follows:
[0107] (a) Assume that the two pairs of feature points are selected as ;
[0108] (b) by Calculate the angle angle1=-30° between the line connecting the two points and the horizontal coordinate system. Calculate the angle angle2=50° between the line connecting the two points and the horizontal coordinate system;
[0109] (c) By using angle1 and angle2, we can get the rotation angle of the matching target compared to the template image: angle_final=80°. Figure 6 shown.
[0110] The specific process of selecting 2 pairs of feature points from 6 pairs of matching feature points and calculating the matching target scaling is as follows:
[0111] (a) Assume that the two pairs of feature points are selected as ,;
[0112] (b) The scaling scale of the matching target is obtained according to the following formula, where: Representing feature points and The Euclidean distance between .
[0113]
[0114] (6) Using feature 1 extracted in step (2) as the attention mechanism, guide the image pyramid constructed in step (2) to perform convolution calculation on the ROI to obtain a new set of scale, angle, and coordinate information;
[0115] The specific process of convolution calculation is as follows:
[0116] The convolution kernel of the convolution calculation is the image pyramid constructed by Temp. It directly performs convolution on Src or ROI, and accelerates the convolution through parallel computing. The convolution result is the similarity between Temp and Src or part of the ROI. A new set of scale, angle, and coordinate information is obtained through convolution calculation, and calibrated and checked with the first set of scale, angle, and coordinate information obtained based on SURF feature points to ensure the accuracy of the matching results, thereby improving the robustness of the algorithm. It can also obtain a set of scale, angle, and coordinate information when target matching based on SURF feature points fails, thereby reducing the missed detection rate. In addition, the introduction of convolution, supplemented by the attention mechanism, can achieve multi-target matching. The schematic diagram of the multi-target matching algorithm is shown below. Figure 7 shown.
[0117] The specific content of the attention mechanism is:
[0118] The role of the attention mechanism is to guide the convolution calculation, with Temp's SURF feature points as the main factor. During the convolution calculation, the convolution value within the 8*8 window of the feature point is weighted, thereby reducing the influence of factors such as noise and enhancing the robustness of the convolution result.
[0119] The introduction of convolution, supplemented by the attention mechanism, can achieve multi-target matching. The specific method is:
[0120] (a) When the background of the matching target image is relatively simple, multiple target regions Objs can be extracted through simple Blob analysis, and then SURF features are extracted from Objs respectively based on parallel computing. The subsequent processing flow is the same as described above.
[0121] (b) When the background of the matching target image is relatively complex, the SURF feature of Temp is directly used as the attention mechanism to guide the template image pyramid to perform convolution calculation on Src, and a set of scale, angle, and coordinate information is obtained as the final result.
[0122] (7) Calibrate and check the two sets of scale, angle, and coordinate information obtained in step (5) and step (6) to obtain a final set of results.
[0123] The calibration check is done by:
[0124] (a) Calculate the error between two sets of scale, angle, and coordinate values. When the error range exceeds the set threshold, it means that there is a big problem with the calculation result, so the result is discarded. The scale threshold is 0.1, the angle threshold is 2°, and the coordinate threshold is 10 pixels.
[0125] (b) When the errors of the two sets of scales, angles, and coordinates are all less than the set threshold, it means that the calculation results are basically correct, and the average value of the two sets of multi-dimensional information is taken as the final set of results.
[0126] (8) If the number of matching feature points obtained in step (4) is less than the set threshold, then jump directly to step (6), and use feature1 extracted in step (2) as the attention mechanism to guide the image pyramid constructed in step (2) to perform convolution calculation on Src, and obtain a set of scale, angle, and coordinate information, and use this set of information as the final set of results. The target matching result is as follows: Figure 8 shown.
[0127] A device for a target matching method based on feature points and convolution optimization calculation, comprising: a memory for storing a computer program; a processor for executing the computer program to implement any of the above-mentioned target matching methods based on feature points and convolution optimization calculation.
[0128] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for target matching based on feature points and convolution optimization calculation described in any one of the above items is implemented. The computer-readable storage medium mentioned here includes random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the technical field.
[0129] Professionals may further realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention, and the description of the above embodiments is only used to help understand the technical solution and its core idea of the present invention. It should be pointed out that for ordinary technicians in the technical field, without departing from the principle of the present invention, the present invention can also be improved and modified in a number of ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.
Claims
1. A method for target matching based on feature points and convolution optimization calculation, characterized in that: The following steps are involved: (1) Select a template image Temp, which contains the target to be identified and has no cluttered background. (2) Extract the SURF feature feature1 of the Temp image, and construct the Temp image pyramid based on downsampling, Gaussian smoothing, and rotation; (3) Extract SURF feature feature2 from the matching image Src; (4) Use FLANN matcher to match feature points of feature1 and feature2; (5) If the number of matching feature points obtained in step (4) is greater than the set threshold, feature point calculation is performed, and a set of scale, angle, and coordinate information of the target matched by Src is calculated based on the matched feature points, and ROI extraction is performed on the matching target area according to the calculation results of the feature points; (6) Using feature 1 extracted in step (2) as the attention mechanism, guide the image pyramid constructed in step (2) to perform convolution calculation on the ROI to obtain a new set of scale, angle, and coordinate information; (7) Calibrate and check the two sets of scale, angle, and coordinate information obtained in step (5) and step (6) to obtain a final set of results; (8) If the number of matching feature points obtained in step (4) is less than the set threshold, jump directly to step (6), use feature1 extracted in step (2) as the attention mechanism, guide the image pyramid constructed in step (2) to perform convolution calculation on Src, obtain a set of scale, angle, and coordinate information, and use this set of information as the final set of results; The method for calculating the characteristic points comprises the following steps: (a) Based on the FLANN matcher, the matching feature points MatchPoint of feature1 and feature2 are obtained. The elements in the MatchPoint exist in pairs. The transformation matrix H1 is calculated based on the MatchPoint. The transformation matrix H1 is a 3-row × 3-column transformation matrix. The transformation matrix H1 is used to express the mapping relationship between the paired feature points in the MatchPoint. When calculating the transformation matrix H1, the RANSAC algorithm is used to remove some erroneous matching points. (b) Calculate the four mapping points of the four vertices of Temp in Src according to the transformation matrix H1. The four mapping points are connected to form a closed area. The content of the area is the matching target range of Temp in Src. Then, the matching feature points are screened according to the four mapping points to further eliminate the erroneous matching points outside the matching target range. (c) sorting the further screened matching feature points in descending order of similarity, and selecting the top n pairs of matching feature points in the sorting results; (d) Use the simulated annealing algorithm to optimize the first n pairs of selected matching feature points and find 6 pairs of optimal feature points; (e) After finding the best 6 pairs of matching feature points, recalculate a more accurate transformation matrix H2, and update the matching target range of Temp in Src based on the new transformation matrix H2; (f) intercepting the optimized matching target range as the target ROI; (g) Select 2 pairs of feature points from the 6 pairs of matching feature points and calculate the rotation angle and scaling scale of the matching target; The method of the attention mechanism includes: the function of the attention mechanism is to guide the convolution calculation, taking the SURF feature point of Temp as the leading factor, and during the convolution calculation, the convolution value within the 8*8 window of the feature point is weighted; The feature 1 extracted in step (2) is used as an attention mechanism to guide the image pyramid constructed in step (2) to perform convolution calculation, wherein the convolution kernel of the convolution calculation is the image pyramid constructed by Temp, and the convolution is directly performed on Src or ROI. The convolution is accelerated by parallel operation, and the convolution result is the similarity between Temp and a partial area of Src or ROI.
2. The target matching method based on feature points and convolution optimization calculation according to claim 1, characterized in that: The method of selecting 2 pairs of feature points from 6 pairs of matching feature points and calculating the rotation angle of the matching target is as follows: (a) The two pairs of feature points selected are ; (b) By Calculate the angle angle1 between the line connecting the two points and the horizontal coordinate system, by Calculate the angle angle2 between the line connecting the two points and the horizontal coordinate system; (c) The rotation angle of the matching target compared to the template image can be obtained through angle1 and angle2.
3. The target matching method based on feature points and convolution optimization calculation according to claim 1, characterized in that: The method of selecting 2 pairs of feature points from 6 pairs of matching feature points and calculating the scaling scale of the matching target is as follows: (a) The two pairs of feature points selected are ; (b) The scaling scale of the matching target is obtained according to the following formula: ; In the formula, Representing feature points and The Euclidean distance between .
4. A device for a target matching method based on feature points and convolution optimization calculation, characterized in that: include: Memory for storing computer programs; A processor, configured to execute the computer program to implement a method for target matching based on feature points and convolution optimization calculation as described in any one of claims 1 to 3.
5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for target matching based on feature points and convolution optimization calculation as described in any one of claims 1 to 3 is implemented.
Citation Information
Patent Citations
Virtual wall construction method and device based on machine vision, map construction method and movable electronic device
CN109313822A
Ultra-short-term photovoltaic power prediction method
CN112507793A
Dressing safety detection method and device, storage medium and computer equipment
CN112651371A
Robot for bone joint replacement surgery
CN114027985A
Template matching algorithm based on edge and gradient feature
WO2021017361A1
