Synchronous localization and mapping method and device based on A-KAZE feature recognition and RatSLAM

Abstract

The invention relates to a synchronous localization and map construction method and device based on A-KAZE feature recognition and RatSLAM, and the method comprises the following steps: obtaining speedometer information and an RGB image of a current scene, and building a topological experience map by using an improved RatSLAM algorithm; the improved RatSLAM algorithm is a RatSLAM algorithm in which an A-KAZE feature recognition algorithm is introduced into a local view module, and in the improved RatSLAM algorithm, the local view module extracts feature information of a grey-scale map obtained by converting an RGB image through the A-KAZE feature recognition algorithm, and then the feature information is matched with a historical visual template; if the matching is successful, carrying out loopback detection to further correct the established topological experience map; and otherwise, storing the grey-scale map as a new visual template. Compared with the prior art, the method has the advantages that the important structure and texture information in the environment can be stably captured, the coverage rate and the accuracy of loopback detection are improved, and then the integrity and the accuracy of the map are enhanced.

Description

Technical Field

[0001] The invention belongs to the technical field of bionics and computer vision, and in particular relates to a synchronous positioning and map building method and device based on A-KAZE feature recognition and RatSLAM. Background Art

[0002] RatSLAM is a simultaneous localization and mapping (SLAM) technology based on bionic principles. It is inspired by the navigation mechanism of the hippocampus of rodents (such as mice). By analyzing sensor data, such as visual and motion information, it realizes the autonomous positioning and mapping of robots in unknown environments. One of the key features of RatSLAM is to represent the environment in an online incremental manner. RatSLAM builds a pose cell network that can be updated by three methods: Attractor Dynamics, Path Integration, and Local View Calibration, allowing the robot to dynamically update the map during movement and observation. However, the traditional RatSLAM has the following shortcomings in practical applications: First, it has poor adaptability to complex environments (such as lighting changes, occlusions or dynamic objects) when extracting image information. This is because there are errors in the robot's estimation of its own position and environment in complex environments. These errors accumulate over time to form cumulative errors; second, the computational complexity is high. When processing large-scale environments or high-resolution images, a large amount of computing resources are required for feature extraction and matching.

[0003] Chinese invention patent CN113223099B discloses a method for improving the environmental adaptability of RatSLAM based on a biological vision model, which converts images in RGB color space to HSI color space, converts them to RGB space after dual-channel processing through I channel and S channel to generate visual templates, uses SAD algorithm to match visual templates, and uses closed-loop detection to improve the topological empirical map established by RatSLAM, thereby improving the accuracy of the established topological map. However, the above method still has certain limitations. For example, under extreme lighting conditions, the processing of I channel cannot completely eliminate the impact of lighting changes on image matching accuracy; in scenes with complex textures or blurred edges, the processing of S channel has limited enhancement of edge characteristics; SAD algorithm is more sensitive to lighting changes, and when lighting conditions change greatly, mismatching or missed matching may occur. Therefore, it is necessary to redesign a synchronous positioning and map construction method based on RatSLAM to further improve the integrity and accuracy of the established topological map. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a synchronous positioning and map construction method and device based on A-KAZE feature recognition and RatSLAM, so as to further improve the integrity and accuracy of the established topological map.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] The present invention provides a synchronous positioning and map construction method based on A-KAZE feature recognition and RatSLAM, comprising the following steps:

[0007] Obtain odometer information and RGB images of the current scene, and use the improved RatSLAM algorithm to build a topological empirical map;

[0008] The improved RatSLAM algorithm is a RatSLAM algorithm that introduces the A-KAZE feature recognition algorithm in the local view module. In the improved RatSLAM algorithm, the local view module extracts feature information of the grayscale image obtained by converting the RGB image through the A-KAZE feature recognition algorithm, and then matches it with the historical visual template; if the match is successful, loop detection is performed to further correct the established topological experience map; otherwise, the grayscale image is stored as a new visual template;

[0009] The feature information includes feature points and their main directions. The specific process of extracting feature information through the A-KAZE feature recognition algorithm is as follows: through the Perona-Ma1ik diffusion equation, different conduction functions are used to perform nonlinear diffusion filtering on the grayscale image to obtain a series of images of different scales, and a nonlinear scale space is constructed. Each conduction function is used to control the diffusion rate and adjust the diffusion intensity according to the local gradient change of the image; local maximum points are found in the nonlinear scale space as feature points, and the main directions of each feature point are generated through vector superposition operations.

[0010] Furthermore, the expression of the nonlinear diffusion filtering is as follows:

[0011]

[0012] Where L is the image brightness, is the image gradient, c(x, y, t) is the conduction function, (x, y) represents the pixel point on the image, and t represents the time.

[0013] Furthermore, nonlinear diffusion filtering is performed through three conduction functions, including a first conduction function g1, a second conduction function g2 and a third conduction function g3, and the expressions are as follows:

[0014]

[0015]

[0016]

[0017] in, Represents the image L obtained after Gaussian filtering σ The gradient of , σ is the scale parameter, which represents the standard deviation of the Gaussian kernel, and k is the contrast factor, which is used to control the diffusion level.

[0018] Furthermore, the expression of the time for Gaussian filtering of the image is as follows:

[0019]

[0020] Among them, σ i represents the scale parameter of the i-th image in the nonlinear scale space, t i represents the Gaussian filtering time of the ith image in the nonlinear scale space.

[0021] Furthermore, the scale parameter σ of the i-th image in the nonlinear scale space is i The calculation formula is as follows:

[0022] σ i (o, s) = σ02 o+s / S , o∈[0…O-1], s∈[0…S-1], i∈[0…N]

[0023] Wherein, o and s are the discrete indices of octave and sub-level respectively obtained by discretizing the nonlinear scale space, σ0 represents the initial reference value of the scale parameter, O and S are the number of octave and sub-level respectively, and N=0*S represents the total number of images in the nonlinear scale space.

[0024] Furthermore, the Hessian matrix is ​​used to find the local maximum point as the feature point. The expression of the Hessian matrix is ​​as follows:

[0025]

[0026] Among them, L Hessian is the Hessian matrix, σ is the scale parameter, L xx , L yy and L xy They represent the second-order derivatives of the image pixel grayscale L in the x direction, y direction, and x and y directions respectively.

[0027] Furthermore, the specific process of using the Hessian matrix to find local maximum points is as follows: compare the Hessian matrix eigenvalue of a pixel point with the Hessian matrix eigenvalues ​​of all its image domain and scale domain adjacent points. When the Hessian matrix eigenvalue of the pixel point is greater than that of all its adjacent points, the pixel point is a local maximum point.

[0028] Furthermore, the specific process of generating the main direction of the feature point is as follows:

[0029] The search circle is divided with the current feature point as the center and the integer multiple of its corresponding scale parameter as the radius;

[0030] With the current feature point as the center, a fan-shaped sliding window is set and a vector superposition operation is performed in the fan-shaped sliding window to obtain a vector. The fan-shaped sliding window traverses the entire search circle, and the angle of the longest vector obtained is the main direction of the feature point.

[0031] Furthermore, the vector superposition operation is: Gaussian weighting of the first-order differential values ​​of all feature points in the fan-shaped sliding window, including the first-order derivative in the y direction and the first-order derivative in the x direction.

[0032] The present invention also provides an electronic device, comprising a memory, a processor, and a program stored in the memory, wherein the processor implements the above method when executing the program.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. The present invention introduces the A-KAZE feature recognition algorithm in the local view module of the RatSLAM algorithm to extract the feature information of the grayscale image obtained by converting the RGB image, and then matches it with the historical visual template. If the match is successful, loop detection is performed to further correct the established topological empirical map, otherwise, the grayscale image is stored as a new visual template; the feature information includes feature points and their main directions. The specific process of extracting feature information by the A-KAZE feature recognition algorithm is as follows: the grayscale image is subjected to nonlinear diffusion filtering by using different conduction functions through the Perona-Malik diffusion equation to obtain a series of images of different scales, and a nonlinear scale space is constructed. Each conduction function can be used to control the diffusion rate and the local gradient of the image is adjusted according to the local gradient of the image. The diffusion intensity is adjusted by adjusting the degree of change, so as to achieve the effect of image smoothing and edge preservation, and the diffusion can be adapted to the local structure of the image; the local maximum points are found as feature points in the nonlinear scale space, and the main direction of each feature point is generated by vector superposition operation; compared with traditional feature extraction algorithms, A-KAZE has better robustness and stability, especially when processing images with large scale and perspective changes. The extracted feature points not only have good scale invariance and rotation invariance, but also can capture important structure and texture information in the environment. Introducing the feature information extracted by A-KAZE into the map construction process of RatSLAM can improve the coverage and accuracy of loop detection, thereby enhancing the integrity and accuracy of the map.

[0035] 2. The present invention constructs three conduction functions for nonlinear diffusion filtering, including the first conduction function g1, the second conduction function g2 and the third conduction function g3, wherein g1 increases with The modulus increases and decreases rapidly, and the high-contrast edges can be retained preferentially; g2 also decreases with the increase of the gradient modulus, but the rate of decrease is slower than the exponential form. This form can still retain a certain diffusion effect under larger gradient values, and can retain relatively subtle edges or weak edge information; g3 is a threshold function. When the gradient modulus is equal to 0, the function value is 1, which means that when there is no local change in the image, the conduction function is the largest, allowing a larger diffusion to smooth the image. Once the gradient is greater than 0, the conduction function decreases rapidly, thereby suppressing diffusion. This design can effectively smooth the interior of the region while retaining boundary information. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the process to improve the RatSLAM algorithm;

[0037] Figure 2 The effect of different contrast factors k is shown (from left to right, the value of the contrast factor k changes from small to large);

[0038] Figure 3Generate a schematic diagram for the main directions of feature points;

[0039] Figure 4 This is the result of A-KAZE feature recognition;

[0040] Figure 5 Compare the results for empirical mapping,

[0041] Among them, (5a) is the actual route, (5b) is the empirical map drawn by the original RatSLAM, and (5c) is the empirical map drawn by the improved RatSLAM with the introduction of the A-KAZE algorithm;

[0042] Figure 6 It is the algorithm loop performance data;

[0043] Figure 7 is the map trajectory accuracy evaluation result,

[0044] Among them, the title APE wrt translation part (m) (with SE (3) Umeyama alignment) means the error calculated by “absolute trajectory error (APE) relative to the translation part, and by SE (3) Umeyama alignment”, (7a) is the result of automatic alignment of the estimated value obtained by the method of the present invention with the true data value using the evo tool, the horizontal and vertical axes represent the distance in the real environment, the unit is meter, reference represents the reference object, that is, the dotted line is the true value of the route, the colored line is the estimated value, the redder the color, the larger the error here, and the bluer the color, the smaller the error, (7b) is the absolute trajectory error calculation result, max represents the maximum error value (Maximum Error), mean represents the average value of all error values ​​(Mean Error), median represents the median of all error values ​​(Median Error), min represents the minimum error value (Minimum Error), rmse represents the root mean square error (Root M ean S q uare E rr o r) ， sse represents the sum of squared errors ( S um o f S q uared E rr o rs) ， std stands for standard deviation. The fluctuating gray solid line represents the size of the error value at each moment, and the other straight lines are the values ​​of the corresponding indicators. DETAILED DESCRIPTION

[0045] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0046] Example:

[0047] This embodiment provides a method for synchronous positioning and map construction based on A-KAZE feature recognition and RatSLAM, comprising the following steps:

[0048] Obtain odometer information and RGB images of the current scene (from the robot camera or dataset), and use the improved RatSLAM algorithm to draw a topological empirical map.

[0049] The improved RatSLAM algorithm is a RatSLAM algorithm that introduces the A-KAZE feature recognition algorithm in the local view module. Figure 1 As shown in the figure, the local view module extracts the feature information of the grayscale image obtained by converting the RGB image through the A-KAZE feature recognition algorithm, and then matches it with the historical visual template; if the match is successful, loop detection is performed to further correct the drawn topological experience map; otherwise, the grayscale image is stored as a new visual template.

[0050] The feature information includes feature points and their main directions. The specific process of extracting feature information through the A-KAZE feature recognition algorithm is as follows:

[0051] First, the Perona-Malik diffusion equation (Equation (1)) is used to perform nonlinear diffusion filtering on the grayscale image using different conduction functions to obtain a series of images of different scales and construct a nonlinear scale space:

[0052]

[0053]

[0054] Where L is the image brightness, is the image gradient, c(x, y, t) is the conduction function, which is used to control the diffusion rate and adjust the diffusion intensity according to the local gradient change of the image, so as to achieve the effect of image smoothing and edge preservation, and make the diffusion adapt to the local structure of the image. The conduction function can be a scalar or a tensor, (x, y) represents the pixel point on the image, and time t is used as the scale parameter. The larger its value, the simpler the image representation. Represents the image L obtained after Gaussian filtering σ The gradient of , σ is the scale parameter, which represents the standard deviation of the Gaussian kernel.

[0055] In this embodiment, nonlinear diffusion filtering is performed by three conduction functions, including a first conduction function g1, a second conduction function g2, and a third conduction function g3, and the expressions are as follows:

[0056]

[0057]

[0058]

[0059] Among them, k is the contrast factor, which is used to control the diffusion level and can determine how much edge information is retained. The larger its value, the less edge information is retained, such as Figure 2 As shown in FIG. 1 , from left to right, the image processing results of the parameter k changing from small to large are shown. The gradient value corresponding to 70% of the gradient histogram after smoothing the original image is taken as the empirical value of k.

[0060] Different g functions can be used for different application scenarios. g1 increases with VL σ The modulus increases and decreases rapidly. The gradient in the edge area is large, which makes g1 very small, and can suppress diffusion and retain the edge. Therefore, g1 can give priority to retaining high-contrast edges. g2 also decreases with the increase of the gradient modulus, but the speed of decrease is slower than the exponential form. Compared with the exponential function, this form can still retain a certain diffusion effect under larger gradient values, which is suitable for retaining more subtle edges or weak edge information. Therefore, g2 can give priority to retaining areas with larger widths. ge is a threshold function. When the gradient modulus is equal to 0, the function value is 1, which means that when there is no local change in the image (the gradient is 0, that is, the flat area), the conduction function is the largest, allowing a larger diffusion to smooth the image. Once the gradient is greater than 0, the conduction function decreases rapidly, thereby suppressing diffusion. This design can retain the edge information in the image without smoothing them out. Therefore, g3 can effectively smooth the interior of the area and retain the boundary information.

[0061] In order to construct a nonlinear scale space, a set of evolution times needs to be defined:

[0062]

[0063]

[0064] Among them, σ i represents the scale parameter of the i-th image in the nonlinear scale space, t irepresents the Gaussian filtering time of the i-th image in the nonlinear scale space. o and s are the discrete indices of octave and sub-level obtained by discretization of the nonlinear scale space, σ0 represents the initial reference value of the scale parameter, O and S are the number of octave and sub-level, respectively, and N = O*S represents the total number of images in the nonlinear scale space. In the Gaussian scale space, convolving the image with a Gaussian kernel with a standard deviation of σ is equivalent to convolving the image with a duration of After multiple iterations, the image is updated and each layer of the scale space is generated to achieve edge preservation and smoothing at multiple scales.

[0065] Then, local maximum points are found as feature points in the nonlinear scale space, and the main directions of each feature point are generated by vector superposition operation. A-KAZE feature point detection is similar to SIFT, which is achieved by finding local maximum points of Hessian after normalization at different scales. The Hessian matrix is ​​used to find local maximum points as feature points. Hessian The expression is as follows:

[0066]

[0067] Among them, L xx , L yy and L xy They represent the second-order derivatives of the image pixel grayscale L in the x direction, y direction, and x and y directions respectively.

[0068] The specific process of using the Hessian matrix to find local maximum points is as follows: compare the Hessian matrix eigenvalue of a pixel point with the Hessian matrix eigenvalues ​​of all its image domain and scale domain adjacent points. When the Hessian matrix eigenvalue of the pixel point is greater than that of all its adjacent points, the pixel point is a local maximum point.

[0069] In order to achieve image rotation invariance, it is necessary to determine the main direction of the feature point based on the local image structure. The specific process of generating the main direction of the feature point is as follows:

[0070] The search circle is divided with the current feature point as the center and the integer multiples of its corresponding scale parameter as the radius. The radius size affects the size of the image area considered in the detection and description process of the feature point. A smaller radius means that the feature point only focuses on local details and is more sensitive to local changes in the image, so more but smaller features may be detected; a larger radius means that the feature point considers a wider image area and can better capture the overall structure and texture information of the image, so larger but more stable features may be detected. Therefore, the radius size of the feature point is actually the result of a trade-off between feature point detection and description in the local scale space. A smaller radius can improve the sensitivity of the feature point and the ability to capture local details, but may cause instability of the feature point; a larger radius can enhance the stability of the feature point, but may lose some small features. In this embodiment, 6σ is taken. i is the radius.

[0071] With the current feature point as the center, set a fan-shaped sliding window and calculate the first-order differential values ​​of all feature points in the area enclosed by the fan-shaped sliding window and the search circle, including the first-order derivative L in the x direction. x and the first-order derivative L in the y direction y Gaussian weighting makes the response contribution close to the feature point large, while the response contribution far from the feature point is small. In this embodiment, the angle of the fan-shaped sliding window is 60°. Figure 3 As shown, the fan-shaped sliding window traverses the entire search circle, and the angle of the longest vector obtained is the main direction of the feature point. Through the above operation, the feature information of each frame image can be obtained, such as Figure 4 , the solid circle is the feature point of the current frame image, and the arrow is the feature direction of the feature point.

[0072] The extracted feature information and the ID of the current frame image are saved as a visual template. A visual template will be obtained for each input frame image. When a new visual template is obtained, it will be matched with all previous visual templates. If the match is successful, the loop detection is completed; if not, it is stored as a new visual template.

[0073] The Kitti series of datasets was developed by a team from the Karlsruhe Institute of Technology in the United States. The datasets are the most commonly used public datasets for evaluating performance in visual SLAM. To verify the effectiveness of the above method, this example uses the Kitti00 dataset for testing. The empirical map drawing results are shown in the figure below. Figure 5As shown in the figure, (5a) is the actual route, (5b) is the empirical map drawn by the original RatSLAM, and (5c) is the empirical map drawn by the improved RatSLAM with the A-KAZE algorithm. It can be seen that the empirical map drawn by the improved RatSLAM with the A-KAZE algorithm is closer to the actual route.

[0074] RatSLAM algorithms usually use template matching graphs to evaluate visual processing performance, such as Figure 6 . The horizontal axis in the figure represents the number of frames captured by the camera, the vertical axis represents the ID of the local visual scene list of RatSLAM, the blue line is the true value data, and the red line is the experimental test result. Each ID represents a specific scene. When a new scene appears, a new ID is generated; when returning to the previous scene, the ID will be restored to the previous value. The higher the overlap between the experimental curve and the true curve, the better the visual processing performance.

[0075] The Kitti00 dataset has a total of 5 loop segments. The RatSLAM optimization method based on A-KAZE can detect continuous and accurate loop segments. Based on the above data, the accuracy and recall commonly used in loop detection are used to quantitatively analyze the visual processing performance of this method, as shown in equations (9) and (10) respectively.

[0076]

[0077]

[0078] Among them, TP represents the number of closed loops correctly identified by the algorithm, FP represents the number of non-closed loops incorrectly identified as closed loops by the algorithm, and FN represents the number of closed loops incorrectly identified as non-closed loops by the algorithm. After calculation, the recall rate based on A-KAZE and RatSLAM is 91.83% and the accuracy rate is 89.82%.

[0079] To further compare the trajectory generated by the RatSLAM system with the known real trajectory and evaluate its positioning and map construction accuracy, the absolute trajectory error is calculated using the evo tool. The results are as follows: Figure 7 As shown. Among them, (7a) is the result of automatic alignment of the estimated value obtained by the method of the present invention with the true value of the data (reference value) using the evo tool, and (7b) is the result of absolute trajectory error calculation, where the root mean square error rmse (Root Mean Square Error) is 17.097995 meters, the total length of the data set is 3724 meters, and the normalized rmse is about 0.459%. It can be seen that the map constructed by the method of the present invention has a high accuracy.

[0080] In summary, the present invention combines the RatSLAM method with the A-KAZE feature extraction to optimize the performance of RatSLAM in map construction. First, a scale space is constructed on the basis of the image obtained by RatSLAM, and a nonlinear diffusion filter is applied instead of an ordinary Gaussian filter to obtain a multi-scale image and extract feature points and feature directions. These feature points have good scale invariance and rotation invariance, and can stably capture important structures and texture information in the environment at the same time. This multi-scale feature extraction method is similar to the human visual system, which performs focus adjustment on objects of different distances and sizes to fully perceive and understand the environment. At the same time, the features extracted by A-KAZE are usually descriptions of local textures or structures, which are similar to the importance of local features that humans pay attention to when identifying objects, such as corners, edges, etc. Then the extracted feature information is combined with the local view module of RatSLAM to complete loop detection, thereby updating the odometer accumulated error. By conducting comparative experiments with traditional methods, this combination can improve the coverage and accuracy of loop detection, thereby enhancing the integrity and accuracy of the map, and providing more reliable support for the autonomous navigation of the robot in an unknown environment.

[0081] If the above method is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0082] The above description of the embodiments is to facilitate the understanding and use of the invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A synchronous positioning and map construction method based on A-KAZE feature recognition and RatSLAM, characterized in that: The following steps are involved: Obtain odometer information and RGB images of the current scene, and use the improved RatSLAM algorithm to build a topological empirical map; The improved RatSLAM algorithm is a RatSLAM algorithm that introduces the A-KAZE feature recognition algorithm in the local view module. In the improved RatSLAM algorithm, the local view module extracts feature information of the grayscale image obtained by converting the RGB image through the A-KAZE feature recognition algorithm, and then matches it with the historical visual template; if the match is successful, loop detection is performed to further correct the established topological experience map; otherwise, the grayscale image is stored as a new visual template; The feature information includes feature points and their main directions. The specific process of extracting feature information through the A-KAZE feature recognition algorithm is as follows: through the Perona-Malik diffusion equation, different conduction functions are used to perform nonlinear diffusion filtering on the grayscale image to obtain a series of images of different scales, and a nonlinear scale space is constructed. Each conduction function is used to control the diffusion rate and adjust the diffusion intensity according to the local gradient change of the image; local maximum points are found in the nonlinear scale space as feature points, and the main directions of each feature point are generated through vector superposition operations.

2. A method for synchronous positioning and mapping based on A-KAZE feature recognition and RatSLAM according to claim 1, characterized in that: The expression of the nonlinear diffusion filter is as follows: Where L is the image brightness, is the image gradient, c(x, y, t) is the conduction function, (x, y) represents the pixel point on the image, and t represents the time.

3. The method for synchronous positioning and mapping based on A-KAZE feature recognition and RatSLAM according to claim 1, characterized in that: Nonlinear diffusion filtering is performed through three conduction functions, including the first conduction function g1, the second conduction function g2 and the third conduction function g3, and the expressions are as follows: in, Represents the image L obtained after Gaussian filtering σ The gradient of , σ is the scale parameter, which represents the standard deviation of the Gaussian kernel, and k is the contrast factor, which is used to control the diffusion level.

4. The method for synchronous positioning and mapping based on A-KAZE feature recognition and RatSLAM according to claim 3, characterized in that: The expression for the time of Gaussian filtering of an image is as follows: Among them, σ i represents the scale parameter of the i-th image in the nonlinear scale space, t i represents the Gaussian filtering time of the ith image in the nonlinear scale space.

5. A method for synchronous positioning and mapping based on A-KAZE feature recognition and RatSLAM according to claim 4, characterized in that: The scale parameter σ of the i-th image in the nonlinear scale space is i The calculation formula is as follows: s i (o, s)=σ02 o+s / S ,o∈[0…O-1],s∈[0…S-1],i∈[0…N] Wherein, o and s are the discrete indices of octave and sub-level respectively obtained by discretizing the nonlinear scale space, σ0 represents the initial reference value of the scale parameter, O and S are the number of octave and sub-level respectively, and N=O*S represents the total number of images in the nonlinear scale space.

6. The method for synchronous positioning and mapping based on A-KAZE feature recognition and RatSLAM according to claim 1, characterized in that: The Hessian matrix is ​​used to find the local maximum point as the feature point. The expression of the Hessian matrix is ​​as follows: Among them, L Hessian is the Hessian matrix, σ is the scale parameter, L xx , L yy and L xy They represent the second-order derivatives of the image pixel grayscale L in the x direction, y direction, and x and y directions respectively.

7. The method for synchronous positioning and mapping based on A-KAZE feature recognition and RatSLAM according to claim 6, characterized in that: The specific process of using the Hessian matrix to find local maximum points is as follows: compare the Hessian matrix eigenvalue of a pixel point with the Hessian matrix eigenvalues ​​of all its image domain and scale domain adjacent points. When the Hessian matrix eigenvalue of the pixel point is greater than that of all its adjacent points, the pixel point is a local maximum point.

8. The method for synchronous positioning and mapping based on A-KAZE feature recognition and RatSLAM according to claim 1, characterized in that: The specific process of generating the main direction of feature points is as follows: The search circle is divided with the current feature point as the center and the integer multiple of its corresponding scale parameter as the radius; With the current feature point as the center, a fan-shaped sliding window is set and a vector superposition operation is performed within the fan-shaped sliding window to obtain a vector. The fan-shaped sliding window is traversed throughout the search circle, and the angle of the longest vector obtained is the main direction of the feature point.

9. The method for synchronous positioning and mapping based on A-KAZE feature recognition and RatSLAM according to claim 8, characterized in that: The vector superposition operation is: Gaussian weighting of the first-order differential values ​​of all feature points in the fan-shaped sliding window, including the first-order derivative in the y direction and the first-order derivative in the x direction.

10. An electronic device comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 9 is implemented.

Images