A VSLAM backend optimization method based on maximum cross-correlation entropy of multi-convex combination

Abstract

The application belongs to the field of visual simultaneous localization and mapping (VSLAM), and particularly relates to a VSLAM back-end optimization method based on a maximum cross-correlation entropy of a plurality of convex combinations. The method comprises the following steps: sensor information reading: camera image information reading and preprocessing, so as to obtain pixel information of the image; front-end visual odometry: rough camera motion is estimated according to information of adjacent images, initial values are provided for the back-end, it is judged whether each frame is a key frame, and a preliminary local map is established; back-end optimization: camera poses measured by visual odometry at different times and information of loop detection are accepted, and optimization is performed on the camera poses and the information, so that a globally consistent trajectory and map are obtained; loop detection: it is detected and judged whether the robot has arrived at a previous position, if loop detection is detected, information is provided to the back-end for processing, and a global optimization thread is triggered; and mapping: a map corresponding to a task requirement is established according to an estimated trajectory.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of visual simultaneous localization and mapping (VSLAM), and particularly relates to a VSLAM backend optimization method based on a maximum cross-correlation entropy of a plurality of convex combinations. BACKGROUND

[0002] Thanks to the development of modern science and technology and artificial intelligence, robot technology has become a major research trend to replace traditional productivity. The visual simultaneous localization and mapping (VSLAM) algorithm in the background of artificial intelligence is a product that conforms to the trend of the times and the needs of the times. The VSLAM algorithm has been widely used in human life and has brought great convenience. A typical application is automatic driving, which plays an important role in daily driving, mine and forestry investigation, and car applications for delivering food and goods. The structure of the VSLAM algorithm is mainly divided into front-end and back-end, and the process is mainly divided into five steps: sensor information reading, front-end visual odometry, back-end nonlinear optimization, loop detection and mapping. The front-end is mainly used to collect data through sensors, and the back-end is responsible for further processing the sensor information obtained by the front-end to obtain accurate pose information and map. The back-end optimization module is of great significance to improve the accuracy of VSLAM positioning and mapping, and can be divided into filtering-based methods and nonlinear optimization-based methods. At present, the nonlinear optimization-based method is the mainstream method.

[0003] Most of the existing nonlinear back-end optimization methods assume Gaussian noise and select mean square error (MSE) as the cost function. In the optimization process, MSE regards all data samples as equally important and cannot distinguish between outliers and inliers. When outliers appear in the sample, the performance of the nonlinear back-end optimization algorithm based on MSE will decrease significantly. Using robust cost functions such as Huber, Cauchy and German-Mcclure (G-M), Welsch kernel is a common method to solve the above problems. Some researchers have proposed a general robust cost function that can obtain different robust kernel functions by setting different function shape control parameters, so that the above cost function is only a special case of adaptive cost function. Some researchers have extended the general cost function in an adaptive way to avoid manually selecting kernel functions for specific situations. However, these robust kernel function methods have certain limitations. When the sample close to the true value is non-Gaussian distribution, the performance of the method based on robust kernel function often decreases, which requires higher statistical measure to describe this distribution. In the past few decades, information theory learning (ITL) has attracted widespread attention and various ITL-based methods have been proposed. The maximum cross-correlation entropy (MCC) in information theory can obtain high-order information and has been successfully applied to data classification, image processing, adaptive filtering and other fields under non-Gaussian noise. However, a single MCC function cannot fully cover non-Gaussian noise, and the method of using multiple convex combination maximum cross-correlation entropy can solve this problem. In the field of SLAM, researchers have applied MCC to filter-based back-end optimization algorithms. Since filter-based back-end optimization algorithms are not suitable for large-scale scenarios, it is of great significance to apply information theory to the study of nonlinear back-end optimization. SUMMARY

[0004] The present application provides a back-end optimization method based on multiple convex combination maximum cross-correlation entropy to solve the problem of non-Gaussian noise.

[0005] The present application adopts the following technical scheme: a back-end optimization method based on multiple convex combination maximum cross-correlation entropy, comprising:

[0006] S100: sensor information reading: camera image information reading and preprocessing to obtain pixel information of the image;

[0007] S200: front-end visual odometry: estimating the rough camera motion according to the information of adjacent images to provide initial values for the back-end, judging whether each frame is a key frame and establishing a preliminary local map;

[0008] S300: back-end optimization: accepting camera poses measured by visual odometry at different times and information of loop detection, and optimizing them to obtain a globally consistent trajectory and map;

[0009] S400: loop detection: detect whether the robot has reached the previous position, if the loop is detected, the information is provided to the backend for processing, triggering the global optimization thread;

[0010] S500: mapping: according to the estimated trajectory, a map corresponding to the task requirements is established.

[0011] In some embodiments, step S100 comprises:

[0012] S101: real-time acquisition of image information by camera sensor;

[0013] S102: filtering operation on image information;

[0014] S103: de-warping after filtering;

[0015] S104: after pre-processing of image information, it is passed to the visual odometry module.

[0016] In some embodiments, the estimation of the rough camera motion process in step S200 comprises:

[0017] S201: determine feature points and obtain feature point descriptors;

[0018] S202: inter-image feature matching;

[0019] S203: screening of false matches;

[0020] S204: after obtaining the matching feature points of adjacent frames, estimate the inter-frame camera motion;

[0021] S205: obtain the relative change between frames, and connect the inter-frame changes to obtain a preliminary trajectory map.

[0022] In some embodiments, step S200 comprises determining whether each frame is a key frame:

[0023] 1) more than a specified number of images since the last global repositioning;

[0024] 2) local map construction is in an idle state, or more than a specified number of images have been inserted since the last key frame;

[0025] 3) the current frame tracks less than a specified number of map cloud points;

[0026] 4) the current frame tracks less than a specified value of reference key frame K_ref cloud points.

[0027] In some embodiments, step S300 comprises:

[0028] S301: Construct a backend optimization cost function based on MCMCC for camera pose and landmarks.

[0029]

[0030] In the formula, It consists of a correlation entropy function with two different kernel parameters. It is a multiconvex combination coefficient. ; and It is a kernel parameter; e ij =z ij - h (Ti, Pi) is called the reprojection error, e ij It is the first i The first frame j The reprojection error corresponding to each landmark point It is the first j Road sign In the i The observations of a frame are usually represented by pixel coordinates; P j It is the first j The three-dimensional coordinates of each landmark in the world coordinate system ; It is the transformation matrix, which is the first... i The camera pose corresponding to the frame;

[0031] S302: Solve using graph optimization methods, with the goal of finding the optimal camera pose T. i and map point coordinates P j , so that the objective function Minimum.

[0032] In some embodiments, step S400 includes:

[0033] S401: Perform closed-loop detection simultaneously with S200 to determine whether the robot has previously passed through the same place;

[0034] S402: If the same location is detected, a global optimization process is triggered to optimize the pose of all keyframes and the coordinates of landmarks.

[0035] In some embodiments, step S402 includes:

[0036] In the global map, a cost function based on the maximum cross-correlation entropy is constructed using the poses of all cameras and the 3D coordinates of landmarks in the loop closure.

[0037]

[0038] The goal is to find the optimal camera pose T. i and map point coordinates Pj minimizing an objective function .

[0039] In some embodiments, step S500 comprises: according to the camera pose at each time and the information of the target in the space, establishing a corresponding map according to requirements.

[0040] Compared with the prior art, the present application has the following beneficial effects:

[0041] 1. The present application adopts a multi-convex combination maximum cross-correlation entropy (MCMCC) to construct a VSLAM backend optimization function, and iteratively solves the optimization problem through a Levenberg-Marquardt method, so that the accuracy of the backend optimization is improved in a non-Gaussian noise scene, and the robustness is stronger.

[0042] 2. The present application is compared with other kernel function-based backend optimization algorithms, analyzed, and improved algorithms show better performance in indoor and outdoor scenes.

[0043] 3. The parameters suggested by the present application can be used as a reference for indoor and outdoor experimental parameters. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 A VSLAM technology roadmap is shown;

[0045] Figure 2 A research motivation is shown;

[0046] Figure 3 A re-projection error map is shown;

[0047] Figure 4 KITTI_09 sequence images are shown;

[0048] Figure 5 Image pixel gray scale distribution is shown;

[0049] Figure 6 EuRoC_MH04 trajectory error map based on MCMCC is shown;

[0050] Figure 7 EuRoC_MH04 trajectory error map based on Huber is shown;

[0051] Figure 8 EuRoC_MH04 trajectory error map based on GM is shown;

[0052] Figure 9 EuRoC_MH04 trajectory error map based on Cauchy is shown;

[0053] Figure 10EuRoC_MH04 sequence: comparison of the error of the improved algorithm with other kernel-based algorithms: line chart

[0054] Figure 11 EuRoC_MH04 sequence: comparison of the error of the improved algorithm with other kernel-based algorithms: column chart

[0055] Figure 12 EuRoC_MH04 sequence: comparison of the error of the improved algorithm with other kernel-based algorithms: box chart

[0056] Figure 13 KITTI_10 sequence: comparison of the error of the improved algorithm with other kernel-based algorithms: line chart

[0057] Figure 14 KITTI_10 sequence: comparison of the error of the improved algorithm with other kernel-based algorithms: column chart

[0058] Figure 15 KITTI_10 sequence: comparison of the error of the improved algorithm with other kernel-based algorithms: box chart

[0059] Figure 16 EuRoC parameter setting: influence of local parameters influence of global parameters : 0.7 / 1.4, = 0.7;

[0060] Figure 17 EuRoC parameter setting: influence of local parameters influence of global parameters : 0.5 / 1.0, = 0.7;

[0061] Figure 18 EuRoC parameter setting: influence of multi-convex combination coefficients influence of local parameters : 0.5 / 1.0, influence of global parameters : 0.7 / 1.4

[0062] Figure 19 KITTI parameter setting: influence of local parameters influence of global parameters : 1.0 / 2.0, = 0.7;

[0063] Figure 20 KITTI parameter setting: influence of local parameters influence of global parameters : 0.5 / 1.0, = 0.7;

[0064] Figure 21 KITTI parameter setting: the impact of multi-convex combination coefficient local parameters: 0.5 / 1.0, global parameters: 1.0 / 2.0 DETAILED DESCRIPTION

[0065] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0066] A VSLAM backend optimization method based on multi-convex combination maximum cross-correlation entropy, comprising the following steps.

[0067] S100: sensor information reading: camera image information reading and preprocessing, pixel information of the image can be obtained, and filtering (median filtering, Gaussian filtering, etc.) operation and distortion removal operation are performed on the image;

[0068] S101: real-time acquisition of image information (pixel information of the image) by a camera sensor;

[0069] S102: after obtaining the image data, filtering operation is performed on the image, which is generally median filtering or Gaussian filtering. The filtering function is mainly to remove noise in the image and reduce interference;

[0070] S103: after filtering, distortion removal processing is performed. The camera lens is a convex lens. Due to the refraction, loss, etc. of the lens, the straight line in the real environment becomes a curve in the image. Distortion mainly includes pillow distortion and barrel distortion. In order to solve this problem, distortion removal processing is performed on the image;

[0071] S104: after the camera image is preprocessed, it is transmitted to the visual odometry module;

[0072] S200: front-end visual odometry: according to the information of adjacent images, rough camera motion (transformation matrix T) and road marker three-dimensional coordinates P are estimated, and initial values (including camera pose T and road marker three-dimensional coordinates P) are provided for the back-end. Determine whether each frame is a key frame and establish a preliminary local map.

[0073] How to determine the key frame:

[0074] (1) The distance from the last global repositioning needs to exceed the specified number of images (such as 20 frames). ​​

[0075] (2) The local map construction is in idle state, or has more than a specified number (e.g. 20) of images since the last keyframe insertion.

[0076] (3) The current frame tracks less than a specified number (e.g. 50) of map cloud points.

[0077] (4) The current frame tracks less than a specified value (e.g. 90%) of the reference keyframe K_ref cloud points.

[0078] S201: Determine feature points by FAST algorithm, find special points (e.g. corner points, edge points, etc.) in the image, and obtain feature point descriptors;

[0079] S202: Perform feature matching between images;

[0080] S203: Screen for false matches. Calculate the Hamming distance of the feature point descriptors, remove the feature points with large differences, and retain the accurate matching pairs;

[0081] S204: After obtaining the matching feature points of adjacent frames, estimate the camera motion between frames. Calculate the rotation vector R and displacement vector t between the two frames of images;

[0082] S205: Obtain the relative changes between frames. Connecting the frame changes in sequence is a preliminary trajectory map;

[0083] S300: Back-end (non-linear) optimization: accept camera poses measured by visual odometry at different times, as well as loop detection information, and optimize them to obtain a globally consistent trajectory and map (optimization is divided into local optimization and global optimization);

[0084] S301: Construct a back-end optimization cost function based on MCMCC for camera poses and road points

[0085]

[0086] wherein, consists of two different kernel parameter correlation entropy functions, is a multi-convex combination coefficient, ; and is a kernel parameter.e ij =z ij - h (Ti, Pi) is called the re-projection error, because it indicates the error between the camera observation value and the two-dimensional coordinate in the i-th frame obtained by projecting the three-dimensional map point using the transformation matrix . The re-projection error is calculated as follows: i wherein, yes In the i The observations of a frame are usually represented by pixel coordinates. Pj is the three-dimensional coordinate of the landmark point in the world coordinate system. . It is the transformation matrix, which is the first... i The camera pose corresponding to the frame. The reprojection process is as follows: Figure 3 As shown in the figure, For the observed values, Reproject the map points to the corresponding points.

[0087] S302: Solve using graph optimization methods, also known as Bundle Adjustment; (during the process, a sliding window method may also be used to control the optimization data).

[0088] The goal is to find the optimal camera pose. and map point coordinates , so that the objective function Minimum.

[0089] During the optimization of the cost function formula, the Levenberg-Marquardt algorithm is used for iterative updates. and Since the special Euclidean group SE(3) formed by the transformation matrices cannot be directly differentiated, while its corresponding Lie algebra can be directly differentiated, for The corresponding Lie algebra is derived through iterative updates. The corresponding Lie algebra is , .

[0090] The Levenberg-Marquardt algorithm incorporates constraint terms into the objective function and uses Lagrange multipliers to construct the Lagrange function, expressed as follows:

[0091]

[0092] In the formula x The variables to be optimized are represented by Lie algebras, specifically the camera pose and map points. Let the radius be the confidence interval. f(x) Let the reprojection error function be... , yes f(x) about x The derivative of . D is the coefficient matrix. Definition for x The increment, Includes camera pose increment and map point increment The goal is to find the increment. Minimize the function.

[0093] Then, we can get

[0094]

[0095] Take the derivative of the above equation with respect to x and set it to zero, we get

[0096]

[0097] Simplify to , Simplify . To simplify the calculation, according to the literature, set D = I . Then, we can get

[0098]

[0099] Define , , we can get

[0100]

[0101] Where the Jacobian matrix J(x) is composed of the derivative of each error term e ij with respect to the optimization variables, denoted as

[0102]

[0103] The formula is a nonlinear equation system with respect to the variables , which can be solved by an iterative method.

[0104] S400: Loop detection: Detect whether the robot has passed the previous location. If loop detection is detected, provide information to the backend for processing, triggering the global optimization thread.

[0105] How to judge loop detection: By comparing the similarity between the current frame and the past key frame, if the similarity exceeds a certain threshold, it can be considered as loop detection.

[0106] S401: Perform loop detection while S200 is being performed, to determine whether the robot has previously passed the same place.

[0107] S402: If it is detected that the same place is passed through, a global optimization process is triggered to optimize the poses of all key frames and the coordinates of landmark points. In the global map, all camera poses and landmark point three-dimensional coordinates in the loop are used to construct a cost function based on maximum mutual correlation entropy (a fixed number of key frames can also be selected for optimization according to requirements to reduce the amount of calculation),

[0108]

[0109] The solution target is to find the optimal camera pose T i and map point coordinates P j , so that the objective function is minimized. The solving process is similar to S302 and will not be repeated here.

[0110] S500: Map building: according to the camera pose at each time and the information of the target in space, a corresponding map is established according to requirements.

[0111] To verify the performance of the present application, the present application is applied to the currently better performing SLAM framework ORB-SLAM3, and experiments are performed on the indoor dataset EuRoC and the outdoor dataset KITTI. The EuRoC dataset includes indoor factory, room with calibration version, and ordinary room scene. The KITTI dataset includes outdoor rural, urban and other different scenes. The effectiveness of the proposed algorithm is verified through experiments in different scenes.

[0112] The obtained trajectory is evaluated by using the evaluation tool EVO, and the evaluation index is the absolute pose error (APE), including the root mean square error rmse, the standard deviation std, and the variance sse. The trajectory accuracy of the proposed algorithm is compared with that of the original method of ORB-SLAM3 (backend optimization algorithm based on Huber) and the backend optimization algorithm based on GM and Cauchy.

[0113] The specific experimental process includes the following steps:

[0114] (1) The backend optimization function in ORB-SLAM3 is changed to the backend optimization function constructed based on the multi-convex combination maximum mutual correlation entropy (MCMCC), and the kernel parameters are set.

[0115] (2) The optimization variables are designed as vertices, and the constraint conditions (cost functions) are set as edges

[0116] (3) The VSLAM framework based on MCMCC is tested on the indoor dataset EuRoC and the outdoor dataset KITTI, and the trajectories obtained by the backend optimization algorithm based on other kernel functions and the proposed method are analyzed with the real trajectory by using EVO.

[0117] (4) Experimental results: The performance on two datasets is shown in the following table:

[0118] Table 1 Experimental results on indoor dataset EuRoC

[0119]

[0120] Table 2 Experimental results on outdoor dataset KITTI

[0121]

[0122] From the above results, it can be seen that whether it is an outdoor scene or an indoor scene, the method of the application has improved the trajectory accuracy. From Table 1, it can be seen that on the indoor dataset EuRoC, the improved algorithm based on MCMCC is superior to the algorithms based on the other three kernel functions in most cases. However, in the V101 sequence and the V102 sequence, the backend optimization algorithm based on MCMCC is slightly inferior to the backend optimization algorithm based on Huber. In the MH01 sequence, the backend optimization algorithm based on MCMCC is slightly inferior to the backend optimization algorithm based on GM. Upon closer observation, the MH01, V101, and V102 sequence data acquisition methods and scenes are relatively simple, and the noise tends to be closer to a Gaussian distribution. We can infer that if the re-projection error distribution between the tracking thread and the key frame in ORB-SLAM3 is close to a Gaussian distribution, or the data acquisition method is more stable, the performance of the backend optimization algorithm based on Huber or GM is better than the proposed algorithm, otherwise the performance of the proposed algorithm is better.

[0123] From Table 2, it can be seen that in the outdoor dataset KITTI, the algorithm based on MCMCC has improved performance in most sequences, and is the optimal algorithm. In the KITTI_01, KITTI_02, KITTI_07, and KITTI_09 sequences, the algorithm based on MCMCC does not achieve the optimal value. The analysis is that in sequences 01, 02, 07, and 09, there are many roads and trees, and the scene feature points are not clear. One of the characteristics of MCMCC is that it can suppress outliers, and in this type of scene, there are many groups of mis-matched feature points, and they are easily confused with correct matching results, so the trajectory accuracy decreases.

[0124] The experimental test on the public dataset tests the performance of the proposed method in VSLAM. Next, the focus will be on , , The influence on the performance of the proposed algorithm. The convergence speed is discussed by studying the change of the cost function value in the iteration process. In ORB-SLAM3, the optimization is mainly triggered in local BA and global BA, and the selection of local BA and global BA should be different. In addition, because the reprojection error distribution of indoor scenes and outdoor scenes is different, the setting of parameters will also be different. For this reason, the influence of different parameters on local BA and global BA is analyzed on the EuRoC dataset and the KITTI dataset respectively. On the EuRoC dataset, , and the influence on local BA and global BA is shown in Figure 16 , Figure 17 , Figure 18 and the influence on KITTI is shown in Figure 19 , Figure 20 , Figure 21 . The final recommended parameters are shown in Table 4.

[0125] Table 4 Algorithm parameter settings

[0126]

[0127] The structures, proportions, sizes, etc. shown in the drawings accompanying the present specification are merely used to cooperate with the content disclosed in the present specification, to be understood and read by those skilled in the art, and do not have technical significance, and any modification of the structure, change of the proportional relationship, or adjustment of the size, without affecting the effects and purposes that can be achieved by the present application, should still fall within the scope of the technology disclosed by the present application. At the same time, the terms such as "upper", "lower", "left", "right", "middle" and "one" used in the present specification are merely for the convenience of clear description, and are not used to limit the scope of the present application, and the change or adjustment of the relative relationship, without substantially changing the technical content, is also considered as the scope of the present application. Finally, it should be pointed out that: the above examples are only used to illustrate the technical solutions of the present application, and are not limited thereto; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A back-end optimization method based on the maximum cross-correlation entropy of multi-convex combinations, characterized in that, include: S100: Sensor Information Reading: Reading and preprocessing camera image information to obtain pixel information of the image; S200: Front-end visual odometry: Estimates rough camera motion based on information from adjacent images, provides initial values ​​to the back-end, determines whether each frame is a keyframe, and builds a preliminary local map; S300: Backend optimization: It accepts camera pose measurements from visual odometry at different times, as well as loop closure detection information, and optimizes them to obtain a globally consistent trajectory and map; Step S300 includes: S301: Construct a backend optimization cost function based on MCMCC for camera pose and landmarks. In the formula, It consists of a correlation entropy function with two different kernel parameters. These are multiconvex combination coefficients. ; and It is a kernel parameter; e ij =z ij - h (Ti, Pi) is called the reprojection error, e ij It is the first i The first frame j The reprojection error corresponding to each landmark point It is the first j Road signs In the i The observations of a frame are usually represented by pixel coordinates; P j It is the first j The three-dimensional coordinates of each landmark in the world coordinate system ; It is the transformation matrix, which is the first... i The camera pose corresponding to the frame; S302: Solve using graph optimization methods, with the goal of finding the optimal camera pose T. i and map point coordinates P j , so that the objective function Minimum; S400: Loop Detection: Detects whether the robot has reached a previous position. If a loop is detected, the information is provided to the backend for processing, triggering the global optimization thread. Step S400 includes: S401: Perform closed-loop detection simultaneously with S200 to determine whether the robot has previously passed through the same place; S402: If the same place is detected, a global optimization process is triggered to optimize the pose of all keyframes and the coordinates of landmarks. Step S402 includes: In the global map, a cost function based on the maximum cross-correlation entropy is constructed using the poses of all cameras and the 3D coordinates of landmarks in the loop closure. The goal is to find the optimal camera pose T. i and map point coordinates P j , so that the objective function Minimum; S500: Map Building: Based on the estimated trajectory, build a map that corresponds to the task requirements.

2. The back-end optimization method based on the maximum cross-correlation entropy of multi-convex combinations according to claim 1, characterized in that, Step S100 includes: S101: Acquires image information in real time via camera sensor; S102: Perform filtering operation on image information; S103: Distortion removal processing is performed after filtering; S104: After preprocessing, the image information is transmitted to the visual odometry module.

3. The back-end optimization method based on the maximum cross-correlation entropy of multi-convex combinations according to claim 1, characterized in that, The estimation of the approximate camera motion process in step S200 includes: S201: Determine feature points and obtain feature point descriptors; S202: Perform feature matching between images; S203: Filtering out false matches; S204: After obtaining the feature points matched between adjacent frames, estimate the camera motion between frames; S205: Obtain the relative changes between frames, and string together the changes between frames to obtain a preliminary trajectory map.

4. The back-end optimization method based on the maximum cross-correlation entropy of multi-convex combinations according to claim 1, characterized in that, The step S200 of determining whether each frame is a keyframe includes: 1) The number of frames since the last global relocation must exceed a specified number of frames; 2) The local map construction is in an idle state, or there have been more than a specified number of images since the last keyframe was inserted; 3) The current frame tracks fewer than the specified number of map cloud points; 4) The current frame tracks fewer cloud points than the specified value of the reference keyframe K_ref.

5. The back-end optimization method based on the maximum cross-correlation entropy of multi-convex combinations according to claim 1, characterized in that, Step S500 includes: building a corresponding map according to requirements based on the camera pose and target information in space at each moment.

Images