A method for optimizing the pose graph of a monocular camera in a drone
By combining Gaussian difference and Gaussian blur techniques with the LM algorithm to optimize the pose map of a UAV monocular camera, the problems of inaccurate feature point extraction and discontinuous image processing in traditional methods are solved, achieving higher pose estimation accuracy and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HEFEI UNIV OF TECH
- Filing Date
- 2024-06-15
- Publication Date
- 2026-05-26
AI Technical Summary
Traditional UAV pose graph optimization methods suffer from inaccurate feature point extraction, weak contrast, and inability to handle discontinuous images, which affect the accuracy and stability of detection.
Gaussian difference point interest detection and Gaussian blur techniques are used to process image frame data. By constructing a Gaussian pyramid and downsampling, low-contrast and edge response feature points are removed. The pose parameters are optimized by combining the LM algorithm and outlier loss function to handle outliers, thereby optimizing discontinuous images.
It improves the robustness and accuracy of pose estimation, reduces computational load and data storage pressure, expands the range of pose graph optimization, and enhances the robustness and convergence speed of the system.
Smart Images
Figure CN118674637B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of UAV pose graph optimization technology, and in particular to a method for optimizing the pose graph of a UAV monocular camera. Background Technology
[0002] Over the past few years, the manufacturing cost of quadcopter drones has been decreasing while their performance has been improving, gradually achieving commercialization and personalization. However, traditional pose graph optimization methods on this platform have many problems. For example, previous feature extraction methods have issues such as feature points located at the edges of image frames and weak contrast, resulting in inaccurate feature information that affects the accuracy and stability of detection. Furthermore, pose graph optimization methods can only process a continuous segment of image frames, and cannot process images with discontinuous scales. Summary of the Invention
[0003] This invention discloses a method for optimizing the pose graph of a monocular camera in an unmanned aerial vehicle (UAV). The specific method is as follows:
[0004] During the flight of the drone, image frame data is continuously collected using a monocular camera;
[0005] Gaussian difference point interest detection is performed on continuously acquired image frame data;
[0006] Pose graph optimization is performed for discontinuous image frame data.
[0007] Furthermore, Gaussian difference point interest detection is performed on the continuously acquired image frame data. The specific method is as follows:
[0008] Blur processing, Gaussian pyramid construction, and downsampling are performed on different image frame data;
[0009] Candidate feature points are determined by subtracting Gaussian blurred images of adjacent scales;
[0010] Accurately determine the position of candidate feature points on the image frame;
[0011] Remove low-contrast candidate feature points and edge response feature points to obtain feature points.
[0012] Furthermore, blurring, Gaussian pyramid construction, and downsampling are performed on different image frame data. The specific methods are as follows:
[0013] Initialize the acquired image frame data;
[0014] Convolve the initialized image frame data to obtain a blurred image;
[0015] By changing the blur scale, several blurred images with the same resolution but different degrees of blur can be obtained;
[0016] Construct Gaussian pyramids using the same blurred images with the same resolution but different degrees of blur.
[0017] The last blurred image in the Gaussian pyramid is selected for downsampling, and the downsampled image data is used as the initial image for the next blurred image.
[0018] Furthermore, when convolving the initialized image frame data, the convolution kernel function is defined as:
[0019]
[0020] In the formula, σ is the scale, and x and y are the input images;
[0021] The formula for convolving the image frame data I(x,y) to obtain the blurred image L(x,y,σ) is as follows:
[0022] L(x,y,σ)=G(x,y,σ)*I(x,y)
[0023] In the formula, * denotes the convolution operation;
[0024] The specific methods for changing the fuzziness scale are as follows:
[0025] Set a scale factor k for each pyramid layer so that the Gaussian blur scale of each layer increases progressively, as shown in the following expression:
[0026] σ i =k i σ0
[0027] Where i represents the number of convolutions in the current layer;
[0028] The specific method for constructing a Gaussian pyramid is as follows:
[0029] Several identical blurred images with the same resolution but different degrees of blur are constructed into an image set;
[0030] Within the same image set, starting from the lowest-level image, the blur scale is gradually increased. Each increase in blur level yields a new image L(x,y,σ). i The expression is as follows:
[0031] L(x,y,σ i )=G(x,y,σ i )*I(x,y)
[0032] Within the m-th image set, starting from the initial image, the images are blurred multiple times to generate images L[m,n] of different scales, as expressed below:
[0033] L[m,n]=G(x,y,σ i)*L[n,n-1]
[0034] Where n represents the index of the current scale in the image set, and L[m,0] is the initial image of the m-th image set;
[0035] Downsampling, the specific method is as follows:
[0036] The last blurred image from the Gaussian pyramid is selected for downsampling and used as the initial image for the next image set, as shown in the following expression:
[0037] L[m+1,0]=Downsample(L[m,n max ])
[0038] Where, n max It is the maximum label in the current image set. When m = 0, L[0,0] = I(x,y) is the initial input image.
[0039] Furthermore, candidate feature points are determined using the following method:
[0040] Subtracting Gaussian blurred images of different scales is calculated using the following formula:
[0041] DL[m,n] = L[m,n+1] - L[m,n]
[0042] Traverse the points and select a 3×3×3 range within the current scale DL[m,n], the previous scale DL[m,n-1], and the next scale DL[m,n+1] of the pixel for detection. Determine whether it is the maximum gray value of this region. If it is, add it as a local maximum to the candidate feature point sequence.
[0043] Furthermore, the precise location of candidate feature points on the image frame is determined using the following method:
[0044] A second-order fit is performed on the difference Gaussian function using Taylor expansion:
[0045]
[0046] Where x is the displacement vector of the keypoint, and DL is the difference Gaussian value at the keypoint location. and These are the gradient and Hessian matrix of the deep learning algorithm, respectively.
[0047] By solving the above equations, the precise locations of the key points can be obtained:
[0048]
[0049] in, Used to indicate the location of candidate feature points.
[0050] Furthermore, low-contrast candidate feature points and edge response feature points are removed, specifically as follows:
[0051] An evaluation function is constructed using the two eigenvalues of the Hessian matrix. Let the two eigenvalues of the Hessian matrix be λ1 and λ2, and the ratio of the eigenvalues be:
[0052]
[0053] like
[0054] r <Threshold1
[0055] Threshold1 is the first set threshold.
[0056] Then mark the candidate feature point as an edge response point and remove it from the candidate feature point list.
[0057] The formula for calculating the contrast value of feature points is as follows:
[0058]
[0059] like
[0060]
[0061] Threshold2 is the second set threshold.
[0062] If a candidate point is deemed unstable, it will be removed from the candidate point list.
[0063] Furthermore, the specific method for pose graph optimization for discontinuous image frame data is as follows:
[0064] Calibrate discontinuous image frame data;
[0065] For continuous image frame data, a first residual function is constructed; for discontinuous image frame data, a second residual function is constructed; the first residual function r1(θ) i ,θ j ) and the second residual function r2(θ) i ,θ j The expression is as follows:
[0066]
[0067]
[0068] Where T(θ) i ) and T(θ j) represent the relative motion transformation matrices of sensor frames i and j, respectively; t2v(T) is the form of the corresponding translation and rotation vectors. This represents the expected value of the relative attitude of adjacent nodes;
[0069] The LM algorithm is used to adjust the attitude parameters of the UAV to minimize the overall first residual function and the second residual function.
[0070] Furthermore, the LM algorithm is used to adjust the UAV attitude parameters. The specific method is as follows:
[0071] Construct the Hessian approximation matrix H and the gradient function g, as follows:
[0072] H = J T J+λI
[0073] Where J is the Jacobian matrix of the residual function with respect to the attitude parameters, and I is the identity matrix;
[0074] The gradient function g is expressed as follows:
[0075] g = J T r
[0076] Where r is the first residual function and the second residual function;
[0077] The Hessian matrix can be approximated using the Jacobian matrix, and then a system of linear equations can be solved to obtain the increments needed to update the attitude parameters, as shown in the following expression:
[0078] (J T J+λI)δθ=-J T r
[0079] The attitude parameters are updated by solving for an incremental δθ as follows:
[0080] θ←θ+δθ
[0081] The updated step size can be obtained by adding the calculated increment to the current attitude parameters. The error observation function is expressed as:
[0082]
[0083] The effectiveness of the update can be determined by comparing the error values before and after the update.
[0084] Furthermore, if the error value decreases after the update, decrease the LM parameter λ and use the Newton-Gauss method to quickly approach the optimal solution; if the error value increases, manually increase the LM parameter λ to make the algorithm use gradient descent more.
[0085] The specific method is expressed as follows:
[0086] λ←λ / v
[0087] λ←λ×ν
[0088] Where v>1;
[0089] If the error converges to the set range after iteration, or if the set maximum number of iterations is reached, the iteration stops, and the current set of pose parameters θ is used. i Output as the optimal solution.
[0090] Due to the adoption of the above solutions, the present invention has the following beneficial effects:
[0091] 1. The process of Gaussian difference point interest detection on continuously acquired image frame data suppresses noise in the image frame, effectively smooths the noise in the image, and the processed pose map is smoother with smaller gradient changes, thus improving the robustness of pose estimation.
[0092] 2. Gaussian blur can help extract more stable and salient features, and can extract pose features more accurately, thereby improving the accuracy of pose estimation.
[0093] 3. Gaussian blur reduces the resolution of image frames while preserving features, thereby reducing the number of iterations and computational load in the optimization process, and also reducing the overall data storage pressure.
[0094] 4. The pose graph optimization process expands the scope of pose graph optimization, and can also perform optimization processing on discontinuous images.
[0095] 5. The LM algorithm is used to solve nonlinear problems in optimization processing, and the LM parameters significantly improve the convergence speed;
[0096] 6. Finally, it is not sensitive to the requirements of the initial parameters and converges effectively within a wide range of initial values, thus improving the robustness of the system algorithm.
[0097] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0098] The accompanying drawings of this invention are described below.
[0099] Figure 1 A schematic diagram of the process for constructing an image pyramid.
[0100] Figure 2 This is a schematic diagram of the loss function determination process.
[0101] Figure 3 This is a schematic diagram of the process of the present invention. Detailed Implementation
[0102] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0103] A method for optimizing the pose graph of a UAV monocular camera, such as Figure 3 As shown, the specific steps are as follows:
[0104] S1. During the flight of the UAV, image frame data is continuously collected through a monocular camera.
[0105] S2. Perform Gaussian difference point interest detection on the continuously acquired image frame data.
[0106] The specific processing method for step S2 is as follows:
[0107] S21. Initialization: Set the input image as I(x,y); set the initial scale σ0 and scale factor k. σ0 is between 1.0 and 2.0, and 1.6 is usually chosen.
[0108] The Gaussian scale space kernel function used is defined as follows:
[0109]
[0110] S22. Convolving the input image yields a blurred image. Convolving the image I(x,y) yields the blurred image L(x,y,σ).
[0111] L(x,y,σ)=G(x,y,σ)*I(x,y)
[0112] * indicates a convolution operation.
[0113] Compared to the original image, this image retains the same resolution but undergoes a high-frequency filter to reduce the impact of high-frequency components.
[0114] S23, By continuously changing σ i The value of continuously increases the degree of ambiguity.
[0115] Set a scale factor k for each pyramid level, which is usually taken as... This causes the Gaussian blur scale to gradually increase in each layer:
[0116] σ i =k i σ0
[0117] Where i represents the number of convolutions in the current layer.
[0118] By repeatedly performing the above steps, you can obtain several images with the same resolution but different degrees of blur, thus generating a sequence of Gaussian blurred images.
[0119] S24. Construct each layer of the Gaussian pyramid, such as Figure 1 As shown. The images resulting from convolution are compiled into an image set. Images within the same image set have the same spatial resolution but different blur scales. Within this set, starting from the original image, the blur scale is progressively increased. Each increase in blur level yields a new image L(x,y,σ). i ):
[0120] L(x,y,σ i )=G(x,y,σ i )*I(x,y)
[0121] Within the m-th image set, starting from the initial image, the images are blurred multiple times to generate images L[m,n] of different scales:
[0122] L[m,n]=G(x,y,σ i )*L[m,n-1]
[0123] Where n represents the label of the current scale in the image set, and L[m,0] is the initial image of the m-th image set.
[0124] S25. Downsampling: For each image set, the last blurred image is selected for downsampling and used as the initial image for the next image set. This reduces the image's resolution while preserving its features; a reduction of 1 / 4 or 1 / 2 of the resolution can be chosen.
[0125] The initial image for the next image set:
[0126] L[m+1,0]=Downsample(L[m,n max ])
[0127] Where, n max It is the highest label in the current image set.
[0128] Where, when m=0, L[0,0]=I(x,y) is the initial input image.
[0129] S26. Generate a Gaussian difference image and detect candidate feature points.
[0130] After completing the set number of downsampling iterations, the m-th image set is selected, and a Gaussian difference image is generated by subtracting Gaussian blurred images of adjacent scales. Subtracting Gaussian blurred images of different scales eliminates low-frequency components in the image; the remaining frequency components are correlated with image edges, and the calculation formula is as follows:
[0131] DL[m,n] = L[m,n+1] - L[m,n]
[0132] We calculate the DL blurred image, and then we iterate through the points in it. For each pixel, we select a 3×3×3 range in the current scale DL[m,n], the previous scale DL[m,n-1], and the next scale DL[m,n+1] for detection. We determine whether it is the maximum gray value in this region. If it is, we add it as a local maximum to the candidate feature point sequence.
[0133] S27. Locate the position of candidate feature points.
[0134] Among the candidate feature points, their positions on the image frame are further precisely determined. A Taylor expansion is used to perform a quadratic fitting of the difference Gaussian function:
[0135]
[0136] Where x is the displacement vector of the feature point, and DL is the difference Gaussian value at the feature point location. and These are the gradient and Hessian matrix of the deep learning algorithm, respectively.
[0137] By solving the above equations, the precise location of the feature points can be obtained:
[0138]
[0139] Used to indicate the estimated location.
[0140] S28. Remove low-contrast feature points and edge response feature points.
[0141] At the edges of an image, firstly, due to the incompleteness of their neighborhood, it is impossible to determine whether local extrema exist; secondly, they are easily affected by noise and interference, introducing uncertainty and thus affecting the accuracy and stability of detection; moreover, edge points usually do not have true feature information, but are caused by the edges of certain objects or structures in the image. Therefore, candidate feature points of edge responses are generally not considered in image processing.
[0142] Since the image gradient changes drastically at the edges, an evaluation function can be constructed using the two eigenvalues of the Hessian matrix to make a judgment.
[0143] Let the two eigenvalues of the Hessian matrix be λ1 and λ2 respectively, and the eigenvalue ratio be:
[0144]
[0145] Set a threshold Threshold1, usually small, to distinguish edge response points and interest points. If r < Threshold1, then mark this candidate feature point as an edge response point and remove it from the list of candidate feature points.
[0146] Feature points with low contrast may represent noise or unreliable regions, and their features are often not sufficient for subsequent image processing tasks, so feature points with small contrast values are removed.
[0147] The calculation formula for the contrast value of a feature point is as follows:
[0148]
[0149] Set a threshold Threshold2, usually small. If this candidate point is considered unstable, then remove it from the list of candidate points.
[0150] The finally remaining list of candidate feature points is confirmed as feature points, and the subsequent optimization algorithm is configured with their orientations.
[0151] S3. For discontinuous image frame data, perform pose graph optimization.
[0152] In step S3, motion estimation is performed through feature matching based on the feature points provided above. By continuously comparing the positions of the same feature point in different image frames, a series of pose parameters are obtained to estimate the position of the UAV body.
[0153] First, construct an observation function to visualize the error degree between the map constructed by each parameter adopted and the actual map.
[0154] Formally, this function is defined as follows:
[0155]
[0156] θ1,...,θ N : Represent pose parameters;
[0157] represents the relative pose residual;
[0158] represents the optimization for the parameters θ1,...,θ N to make the subsequent objective function reach the minimum value;
[0159] This represents the summation of all relative pose measurements, where M is the number of relative pose measurements;
[0160] It is the corresponding information matrix, which is usually composed of the covariance matrix of sensor measurements;
[0161] ρ(·) is a robust cost function used to handle observation residuals, thereby reducing the impact of outliers and improving the stability and robustness of the optimization. It is used to mitigate the influence of outliers.
[0162] The Huber loss function can be chosen as the loss function, and the judgment process is as follows: Figure 2 As shown:
[0163]
[0164] r: Relative attitude residual;
[0165] δ: Hyperparameter, obtained through experience or experimentation.
[0166] The Huber loss function is more robust to handling outliers. It behaves like a mean squared error near zero and like an absolute error far from zero. This property allows the Huber loss function to maintain a good balance in handling outliers without excessively affecting the optimization process.
[0167] This includes a relative attitude residual function, which represents the difference between the estimated and actual values, and optimizes the overall attitude estimation by minimizing the sum of squares of all attitude residuals.
[0168] A novel residual function for discontinuous images is constructed for both continuous and discontinuous images. During motion estimation, feature points between adjacent image frames are detected to be consistent and continuous. The discontinuous parts are specially calibrated, and when optimizing the relevant parts, this novel residual function can be used in the solver for calculation.
[0169] For continuously acquired image frames, the following residual function was constructed:
[0170]
[0171] For cases where image acquisition is discontinuous, the following residual function is constructed:
[0172]
[0173] in
[0174] T(θ i ) and T(θ j): These represent the relative motion transformation matrices for sensor frames i and j, respectively;
[0175] t2v(T): Converts the corresponding vectors into translation and rotation vectors.
[0176] The expected value of the relative attitude of adjacent nodes;
[0177] The LM algorithm is used iteratively to adjust the attitude parameters in order to minimize the error function.
[0178] First, select a set of initial attitude parameters θ. i And set an initial LM parameter λ. A maximum number of iterations can be manually set to prevent the algorithm from looping indefinitely when there is no optimal solution.
[0179] Then, a Hessian matrix approximation matrix H and a gradient function are constructed, which enables the algorithm to smoothly transition between gradient descent and the Gauss-Newton method.
[0180] The approximate matrix H is expressed as follows:
[0181] H = J T J+λI
[0182] Where J is the Jacobian matrix of the residual function with respect to the attitude parameters, and I is the identity matrix.
[0183] The gradient function g is expressed as follows:
[0184] g = J T r
[0185] The Hessian matrix can be approximated using the Jacobian matrix, and then a system of linear equations can be solved to obtain the increments we need to update the state.
[0186] (J T J+λI)δθ=-J T r
[0187] We obtain an increment δθ, and then we update the attitude parameters.
[0188] θ←θ+δθ
[0189] The current attitude parameters plus the calculated increment can be used to obtain the updated step size. This step size guides the algorithm in which direction to move in order to minimize the observation function.
[0190] The error observation function is expressed as:
[0191]
[0192] The effectiveness of the update can be determined by comparing the error values before and after the update.
[0193] If the updated error value decreases, decrease the LM parameter λ to allow the algorithm to use the Newton-Gaussian method more frequently to quickly approach the optimal solution; if the error value increases, it indicates that the optimization effect is slow, and the LM parameter λ needs to be manually increased to allow the algorithm to use the gradient descent method more frequently, expand the search range in the search space, and find a better solution.
[0194] Specific adjustment strategies could be:
[0195] λ←λ / ν
[0196] λ←λ×ν
[0197] Where ν>1 is a predefined parameter.
[0198] After each iteration, we need to make a judgment. If the obtained error converges to the range we set, or if the maximum number of iterations is reached, we stop iterating and use the current set of pose parameters θ. i Output as the optimal solution.
[0199] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for optimizing the pose graph of a monocular camera on a UAV, characterized in that, The specific method is as follows: During the flight of the drone, image frame data is continuously collected using a monocular camera; Gaussian difference point interest detection is performed on continuously acquired image frame data; Pose graph optimization is performed on discontinuous image frame data; Gaussian difference point interest detection is performed on continuously acquired image frame data. The specific method is as follows: Blur processing, Gaussian pyramid construction, and downsampling are performed on different image frame data; Candidate feature points are determined by subtracting Gaussian blurred images of adjacent scales; Accurately determine the position of candidate feature points on the image frame; Remove low-contrast candidate feature points and edge response feature points to obtain feature points; The specific method for pose graph optimization for discontinuous image frame data is as follows: Calibrate discontinuous image frame data; For continuous image frame data, a first residual function is constructed; for discontinuous image frame data, a second residual function is constructed; the first residual function... Second residual function The expression is as follows: in, and Representing sensor frames respectively and The relative motion transformation matrix; To convert the corresponding vectors into translation and rotation vectors; This represents the expected value of the relative attitude of adjacent nodes; The LM algorithm is used to adjust the UAV attitude parameters to minimize the overall first and second residual functions; The LM algorithm is used to adjust the attitude parameters of the UAV. The specific method is as follows: Constructing the Hessian approximation matrix and gradient function The expression is as follows: in, Let Jacobian matrix be the residual function with respect to attitude parameters. It is the identity matrix; gradient function The expression is as follows: in, It is the residual function; The Hessian matrix can be approximated using the Jacobian matrix, and then a system of linear equations can be solved to obtain the increments needed to update the attitude parameters, as shown in the following expression: Solve for an increment The attitude parameters are updated as follows: The updated step size can be obtained by adding the calculated increment to the current attitude parameters. The error observation function is expressed as: : Indicates attitude parameters; : Represents relative attitude residual; : indicates for parameters Optimize to minimize the objective function. : indicates summing over all relative pose measurements, where It is the number of relative pose measurements; : is the corresponding information matrix; It is a robust cost function; The effectiveness of the update is judged by comparing the error values before and after the update. If the error value decreases after the update, reduce the LM parameter. The Newton-Gaussian method can be used to quickly approximate the optimal solution; if the error value increases, the LM parameter can be manually increased. This encourages the algorithm to use gradient descent more often. The specific method is expressed as follows: in, ; If the error obtained after iteration converges to the set range, or the set maximum number of iterations is reached, the iteration stops, and the current set of pose parameters is used. Output as the optimal solution.
2. The method for optimizing the pose graph of a UAV monocular camera as described in claim 1, characterized in that, Blurring, Gaussian pyramid construction, and downsampling are performed on different image frame data. The specific methods are as follows: Initialize the acquired image frame data; Convolve the initialized image frame data to obtain a blurred image; By changing the blur scale, several blurred images with the same resolution but different degrees of blur can be obtained; Construct Gaussian pyramids using the same blurred images with the same resolution but different degrees of blur. The last blurred image in the Gaussian pyramid is selected for downsampling, and the downsampled image data is used as the initial image for the next blurred image.
3. The method for optimizing the pose graph of a UAV monocular camera as described in claim 2, characterized in that, When convolving initialized image frame data, the convolution kernel function is defined as: In the formula, As a scale, Input image; For image frame data Perform convolution to obtain a blurred image. The formula is as follows: In the formula, Indicates the convolution operation; The specific methods for changing the fuzziness scale are as follows: Set the scale factor for each pyramid level This causes the Gaussian blur scale of each layer to gradually increase, as shown in the following expression: in Indicates the number of convolutions in the current layer; The specific method for constructing a Gaussian pyramid is as follows: Several identical blurred images with the same resolution but different degrees of blur are constructed into an image set; Within the same image set, starting from the lowest-level image, the blur scale is gradually increased. Each increase in blur level results in a new image. The expression is as follows: In the Within a set of images, starting from the initial image, multiple blurring processes are applied to generate images at different scales. The expression is as follows: in, This indicates the label of the current scale within the image set. It is the first The initial image of the image set; Downsampling, the specific method is as follows: The last blurred image from the Gaussian pyramid is selected for downsampling and used as the initial image for the next image set, as shown in the following expression: in, It is the largest label in the current image set, when hour, This is the initial input image.
4. The method for optimizing the pose graph of a UAV monocular camera as described in claim 1, characterized in that, The specific method for determining candidate feature points is as follows: Subtracting Gaussian blurred images of different scales is calculated using the following formula: Iterate through the points, at the current scale of each point. Previous scale and the next scale In the process, a 3×3×3 area is selected for detection to determine whether it is the maximum gray value in this area; if so, it is added as a local maximum to the candidate feature point sequence.
5. The method for optimizing the pose graph of a UAV monocular camera as described in claim 1, characterized in that, The specific method for accurately determining the position of candidate feature points on an image frame is as follows: A second-order fit is performed on the difference Gaussian function using Taylor expansion: in, It is the displacement vector of the key point. It is the difference Gaussian value at the key point location. and They are The gradient and Hessian matrix; By solving the above equations, the precise locations of the key points can be obtained: in, Used to indicate the location of candidate feature points.
6. The method for optimizing the pose graph of a UAV monocular camera as described in claim 5, characterized in that, The method for removing low-contrast candidate feature points and edge response feature points is as follows: Construct an evaluation function using the two eigenvalues of the Hessian matrix. Let the two eigenvalues of the Hessian matrix be... and Construct evaluation parameters for: like in, Then mark the candidate feature point as an edge response point and remove it from the candidate feature point list. The formula for calculating the contrast value of feature points is as follows: like in, Set a threshold for the second one; If a candidate point is deemed unstable, it will be removed from the candidate point list.