A traversable map construction method for legged robots

Abstract

The present invention discloses a method for constructing a traversable map for a legged robot, comprising: preprocessing the acquired raw point cloud information, inertial measurement information, longitude and latitude information, and camera image information; performing error state Kalman filtering on the preprocessed point cloud information and inertial measurement information to obtain the rough robot pose; then obtaining the precise robot pose through graph optimization; constructing and processing a 2.5D grid topographic map through an extended Kalman filter to obtain a smooth grid topographic map without holes; extracting the normal vector within each unit grid in the grid topographic map, calculating the slope, slope change rate and roughness, and traversability within each unit grid, and maintaining the traversable map for the legged robot in real time through an extended Kalman filter. The present invention implements adaptive calculation of traversability for different terrains and materials in the real world, making the resulting traversable map more credible.

Description

Technical Field

[0001] The present invention relates to the technical field of environment perception and map construction of legged robots, and in particular to a traversable map construction method for legged robots. Background Art

[0002] With the rapid development of intelligent mobile robot technology, traversability maps are becoming increasingly important in autonomous robot navigation. This is particularly true for legged robots, such as humanoid robots and quadrupeds, that operate in complex indoor and outdoor environments. Compared to wheeled robots and unmanned vehicles, legged robots face more complex and unpredictable terrain, and therefore rely heavily on real-time, accurate traversability information to ensure efficient and safe operation. Traversability maps for legged robots dynamically reflect the traversability of each detailed area in the environment, enabling robots to quickly identify obstacles and traversability costs, select optimal paths, avoid impassable areas, and reduce energy consumption and time costs. Therefore, developing a traversability map for legged robots that can be generated and updated in real time is of great practical significance.

[0003] Mainstream approaches for constructing traversability maps include extracting traversable areas from 3D maps and calculating grid traversability from 2.5D grid maps (mostly elevation maps). Each approach has its own advantages and disadvantages. 3D maps provide rich environmental details and are suitable for accurate traversability assessment in complex terrain. However, they require high computational resources, are complex to generate and maintain, and are inefficient in real-time applications. 2.5D grid maps are computationally efficient, easy to construct, and update, making them suitable for robots to perform real-time traversability assessments in dynamic environments. However, they only represent terrain elevation information and cannot distinguish other terrain characteristics, such as slope and ground material, in detail, resulting in low accuracy in the resulting traversability maps. To address these issues, research has focused on accurately and rapidly constructing traversability maps. Kang Ke et al. (CN118392156A) calculated the height, slope, curvature, and roughness of each grid in an elevation map to obtain an initial accessibility cost map. They then inflated the obstacles in the initial accessibility cost map based on the edges of the rectangle approximated by the robot to construct a final accessibility cost map. Li Chao et al. (CN118329008A) improved the accuracy of the map by integrating depth camera data into an existing elevation map. The depth camera continuously collected point cloud data of the terrain, which was pre-processed by a perception and computing unit and then registered with the previous frame's point cloud to obtain an accurate relative pose between frames. The elevation map was then updated based on the relative pose to obtain a local terrain map. During registration, the robot's positioning information was used to obtain the initial relative pose of the two registered point clouds. Li Zhijun et al. (CN116147642A) provided a method and system for constructing a reachability map for a quadruped robot that integrates terrain and force. By fusing visual and inertial information, they created a two-dimensional grid map containing height information and set a height threshold to determine terrain accessibility. Li Chuanpeng (CN117830991A) and others proposed a complex scene perception method and system for a quadruped robot based on multimodal fusion. The terrain ahead was reconstructed through a deep neural network, and super-pixel images were obtained by clustering RGB images. Then, a multi-layer perceptron network was used to process the traversable area in the image coordinate system, and then a cost map was generated.

[0004] However, the above methods still do not solve the problem of holes and burrs in the actual use of elevation maps, and the calculation methods of the traversability of different terrains are not differentiated, resulting in the low practicality of the traversability map finally constructed. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the present invention proposes a method for constructing a traversable map for a legged robot.

[0006] The technical solution of the present invention includes the following contents:

[0007] A method for constructing a traversable map for a legged robot comprises the following steps:

[0008] S1. Obtain the original point cloud information around the robot through the lidar, obtain the inertial measurement information of the robot body through the inertial measurement unit, obtain the longitude and latitude information of the robot body through the satellite signal receiver, and obtain the image information in front of the robot through the camera;

[0009] S2. Preprocessing the original point cloud information, inertial measurement information, latitude and longitude information, and camera image information obtained in step S1; wherein the preprocessing includes original point cloud information preprocessing, IMU pre-integration processing, UTM coordinate system projection of latitude and longitude information, and image semantic information extraction;

[0010] S3. Performing error state Kalman filtering on the point cloud information and inertial measurement information preprocessed in step S2 to obtain a rough pose of the robot; using the projected longitude and latitude information as a priori constraints and the rough pose of the robot as the factor to be optimized, obtaining the precise pose of the robot through a graph optimization method;

[0011] S4, constructing a 2.5D grid topographic map using the extended Kalman filter method for the point cloud information preprocessed in step S2 and the precise position and posture of the robot obtained in step S3, and processing the topographic map using a convolutional neural network to obtain a smooth grid topographic map without holes;

[0012] S5. For the grid topographic map obtained in step S4, extract the normal vector within each unit grid, and calculate the slope, slope change rate and roughness within each unit grid based on the normal vector and the precise position of the robot obtained in step S3; combine the image semantic information obtained in step S2 to calculate the passability within each unit grid, and maintain a leg-foot robot passable map in real time through extended Kalman filtering.

[0013] Furthermore, step S2 includes the following sub-steps:

[0014] S21. Preprocessing of the original point cloud information: performing outlier filtering and motion distortion correction on the original point cloud information, and removing dynamic point clouds to obtain a reliable point cloud;

[0015] S22, IMU pre-integration processing: Based on the IMU angular velocity and acceleration data, the relative motion between the two frames during measurement is calculated to obtain the robot posture change;

[0016] S23. Projection of longitude and latitude information into the UTM coordinate system: The satellite receiver interprets the message sent by the Global Navigation Satellite Positioning System (GNSS) as longitude and latitude information based on the WGS84 geocentric coordinate system, and then converts the longitude and latitude information into the UTM coordinate system, i.e., the XYZ three-axis coordinate information in the robot's local coordinate system.

[0017] S24. Image semantic information extraction: Extract the features of the original image information through a convolutional neural network, and use the fully connected layer to output the category and bounding box of the object to achieve semantic information extraction.

[0018] Furthermore, step S3 includes the following sub-steps:

[0019] S31, inferring the initial pose of the current frame based on the credible point cloud obtained in step S2 and the pose change between the two frames of the robot obtained by IMU pre-integration; transforming the initial pose to the coordinate system of the previous frame, and constructing an error model through the geometric registration relationship between the two frames of point clouds to obtain the optimal robot pose state, that is, the rough pose of the robot;

[0020] S32. Using a graph optimization method, with the rough posture of the robot as the vertex and the longitude and latitude information obtained in step S2 as the constraint edge, a graph model is constructed to solve the graph optimization problem and obtain the precise posture state of the robot.

[0021] Furthermore, step S4 includes the following sub-steps:

[0022] S41, based on the credible point cloud obtained in step S2 and the precise position state of the robot obtained in step S3, obtain a 2.5D grid topographic map by using an extended Kalman filter; wherein each grid contains a number of points, an average height value, and a variance estimate obtained by the Kalman filter;

[0023] S42. The raster topographic map obtained in step S41 is vectorized and input into a GAN network. Global features of the image are extracted by downsampling, and the spatial resolution of the image is gradually restored by upsampling. A vectorized, smoothed, and completed raster topographic map is output.

[0024] Furthermore, step S5 includes the following sub-steps:

[0025] S51, performing plane fitting on a plurality of point clouds in each unit grid according to the grid topographic map obtained in step S4, and taking the normal vector of the fitted plane as the normal vector of the unit grid;

[0026] S52, calculating the slope and roughness of each grid based on the normal vector of each grid extracted in step S51 and the normal vectors in its neighboring grids;

[0027] S53 , weightedly calculating the passability of each cell according to the parameters of each cell grid and its surrounding grids obtained in step S52 .

[0028] Furthermore, in step S53, the weight parameters in the weighted calculation of the passability are determined by the image semantic information obtained in step S2, and the weights of the parameters corresponding to different semantic areas are different.

[0029] In summary, the beneficial effects of the present invention are as follows: for robots performing tasks in complex indoor and outdoor environments, a method for generating and updating a 2.5D legged robot traversable map in real time by utilizing multi-sensor information is invented. On the basis of retaining information related to traversability, the present invention chooses to construct a 2.5D grid map with a smaller amount of information. Compared with a three-dimensional traversable map, it is lighter, consumes less computing and storage costs, and can be deployed on a robot for real-time operation. The present invention innovatively uses an adversarial neural network to solve the problem of holes and burrs that appear in elevation maps in practical applications, and provides more accurate and realistic information about the robot's surroundings for subsequent calculation of traversability. When calculating traversability, the present invention innovatively proposes to determine the weights of each information parameter in the traversability calculation formula in different semantic areas based on semantic information, thereby realizing adaptive calculation of traversability for different terrains and materials in the real world, making the obtained traversable map more credible. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a framework diagram of a method for constructing a traversable map for a legged robot according to the present invention;

[0031] Figure 2 A diagram summarizing the coordinate relationships of an embodiment of the present invention;

[0032] Figure 3 This is a flow chart of the algorithm of the adversarial neural network used to complete the elevation map in the present invention. DETAILED DESCRIPTION

[0033] The present invention will be described in detail below with reference to specific embodiments. The embodiments described below are illustrative and intended only to explain the present invention, and are not to be construed as limiting the present invention. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the scope of the present invention. Such modifications and improvements are all within the scope of protection of the present invention.

[0034] Robots that perform tasks in complex indoor and outdoor environments rely on real-time, accurate traversability information to ensure their efficiency and safety. This paper addresses the issues of holes and glitches in elevation maps and the inability to distinguish between traversability calculation methods for different terrains during actual use. A traversability map construction method for legged robots is proposed. Figure 1 This is the framework diagram of the method. The present invention mainly includes five steps: acquiring multi-sensor information, preprocessing the original information, calculating the robot's position, constructing a grid terrain map, and calculating the passability. The following will be described in detail:

[0035] S1. Obtain multi-sensor information. Obtain the original point cloud information around the robot through the laser radar, obtain the inertial measurement information of the robot body through the inertial measurement unit, obtain the longitude and latitude information of the robot body through the satellite signal receiver, and obtain the image information in front of the robot through the camera. In this embodiment, Figure 2 As shown in Figure 1, the relationship between multiple coordinate systems is described, including the robot body coordinate system B, the IMU (Inertial Measurement Unit) center coordinate system I, the lidar coordinate system L, the camera coordinate system C, the satellite positioning receiver coordinate system G, the UTM (Universal Transverse Mercator) grid world coordinate system W, and the robot odometry coordinate system O. The transformation matrix between each coordinate system is given. For example, the transformation matrix from the UTM world coordinate system to the robot odometry coordinate system O is Generally, the inertial measurement unit (IMU) is installed at the center of mass of the robot, that is, the robot body coordinate system coincides with the origin of the IMU center coordinate system I, and the directions of all axes are the same.

[0036] S2. Preprocessing the original information. Preprocess the original point cloud information, inertial measurement information, latitude and longitude information, and camera image information obtained in step S1. The preprocessing includes preprocessing the original point cloud information, IMU pre-integration processing, UTM coordinate system projection of the latitude and longitude information, and extraction of image semantic information.

[0037] S21. Preprocessing of the raw point cloud: First, perform outlier filtering. This involves setting a fixed search radius and checking the number of neighboring points within that radius for each point in the point cloud. If the number of neighboring points falls below a preset threshold, the point is identified as an outlier and filtered out. Motion distortion correction is then performed. This involves calculating the change in the LiDAR's position during each scan and converting the point cloud data to the same coordinate system (the robot's odometry coordinate system, O) to correct for motion distortion. Dynamic point clouds are then removed to obtain a reliable point cloud.

[0038] The specific formula for motion distortion correction is:

[0039]

[0040] Among them, t i , t j , t k Represent the time corresponding to the three frames respectively, and Represents the robot's posture changes from time k to time i and from time j to time i.

[0041] S22. Preprocessing method for inertial unit information: Calculate the relative position change of the robot between two adjacent frames according to the IMU pre-integration formula:

[0042]

[0043] Among them, i, j, k represent three frames respectively, a k and ω k is the acceleration and angular velocity of the kth frame in the I coordinate system, and is the bias of the accelerometer and gyroscope, n a and n g is the noise of the accelerometer and gyroscope. ij is the position change of the robot from frame i to frame j, Δv ij is the speed change of the robot from frame i to frame j, Δq ij is the change in the robot's posture from the i-th frame to the j-th frame.

[0044] S23. Preprocessing method for satellite positioning information: The satellite receiver interprets the message sent by the Global Navigation Satellite System (GNSS) as longitude and latitude information based on the WGS84 geocentric coordinate system. This method converts the longitude and latitude information into the UTM coordinate system, that is, the XYZ three-axis coordinate information in the robot's local coordinate system.

[0045] Specifically, the satellite positioning information gives the longitude and latitude and altitude h, are converted to Cartesian coordinates (X GNSS ,Y GNSS ,Z GNSS ):

[0046]

[0047] The positive direction of the X-axis is due east, the positive direction of the Y-axis is due north, and the positive direction of the Z-axis is vertically upward.

[0048] S24. Image information preprocessing method: Image semantic information extraction method: Extract the features of the original image information through a convolutional neural network, and use the fully connected layer to output the category and bounding box of the object to achieve semantic information extraction.

[0049] Specifically, the acquired image information is segmented and detected using the YOLOv5 visual 2D target detection algorithm to identify semantic information such as cement roads, grass, and steps, and its detection frame is projected into the point cloud coordinate system L according to the pyramid rule and BEV (Bird's-Eye-View) rule.

[0050] S3. Calculate the robot's pose. Perform an error-state Kalman filter on the point cloud information and inertial measurement information preprocessed in step S2 to obtain the robot's rough pose. Using the projected longitude and latitude information as prior constraints and the robot's rough pose as the factor to be optimized, the robot's precise pose is obtained through graph optimization. This is described in detail below:

[0051] S31. Method for obtaining the rough pose of the robot: First, the nominal true value of the current frame is inferred based on the trusted point cloud in step S2 and the pose change between two frames of the robot obtained by IMU pre-integration. That is, use the pose x of the previous frame k Add the relative posture change δx of the robot between two adjacent frames calculated in step S22.

[0052]

[0053] The nominal truth value is completely obtained by the IMU, and drift may occur when the robot moves a longer distance. The next step is to multiply the nominal truth value by the transformation matrix to transform it to the coordinate system of the previous frame, and construct an error model through the geometric registration relationship between the two frames of point clouds to obtain the optimal robot posture state. This step is also called error state Kalman filtering, which can be expressed as considering the nominal truth value obtained from the IMU state propagation model. The error state δx to be corrected and the error model of point cloud registration. The optimization problem can be expressed as follows:

[0054]

[0055] Among them, ||·|| represents the Mahalanobis distance norm, J k is the Jacobian matrix of the lidar measurement noise, which is generally a unit matrix, M k is the measurement noise of the lidar, which is set by the manual parameter of the lidar. The specific form of f(·) is as follows:

[0056]

[0057] By solving the optimization problem, we can get the optimal error state δx, and further get the real position x of the robot in the current frame. k+1 .

[0058] S32. About the method of obtaining the precise posture of the robot: The above process uses the optimization method of tightly coupling LiDAR and IMU to obtain the rough posture of the robot, and then integrates the satellite positioning information through the graph optimization method. Specifically, a graph model is constructed, whose vertices are the rough postures of the robot at different times, and the edges are the constraints composed of the UTM coordinates obtained in step S2. The goal of graph optimization is to optimize the structure of the entire graph by minimizing the errors between these vertices, and then obtain the most accurate posture of the robot. The specific formula can be expressed as:

[0059]

[0060] Among them, X i and X j is the robot pose, the initial value is the rough pose obtained in the previous step, Z ij is the relative observation value between the two frames obtained from the UTM coordinate system, f(X i ,X j ) is the predicted relative pose calculated based on the pose vertices, Ω i,j is the covariance matrix, which is used to express the confidence of the sensor measurement and is usually a symmetric positive definite matrix.

[0061] S4. Construct a grid topographic map. Using the preprocessed point cloud information obtained in step S2 and the precise robot pose obtained in step S3, a 2.5D grid topographic map is constructed using an extended Kalman filter. This map is then processed using a convolutional neural network to produce a smooth, hole-free grid topographic map. This is described in detail below:

[0062] S41. Construct a 2.5D grid terrain map by using the extended Kalman filter method: Based on the coordinate system conversion relationship described in step S1 and the precise posture of the robot obtained in step S3, the height value of the surrounding terrain centered on the robot body can be obtained from the trusted point cloud information in step S2. After the trusted point cloud is rasterized, the height of each unit grid is initialized with the average value of the point cloud height belonging to the grid, and is updated after receiving new point cloud information. During the update process, the height data in the newly received trusted point cloud is fused with the existing elevation estimate through the extended Kalman filter. The introduction of this filtering method can improve measurement accuracy while processing measurement noise. After each measurement update, each grid point of the terrain contains an estimated value and a covariance matrix to represent its uncertainty. When the robot moves, the height estimates of all grid points will be updated according to the changes in the robot state to ensure the consistency of the elevation map with the actual terrain. The update formula is:

[0063]

[0064] Where h is the height of the grid cell, is the variance of the height of this grid cell, p is the average height of this grid cell obtained from the new credible point cloud, is the variance of the height of the grid cell obtained from the new credible point cloud. Generally, the height values ​​obtained from the credible point cloud are considered to follow a Gaussian distribution, that is:

[0065]

[0066] The present invention uses a model that is proportional to the square of the measured distance to represent the variance of the height value. Among them, α is a custom scale parameter, and d is the distance from the grid midpoint to the lidar.

[0067] S42. A method for obtaining a smooth and hole-free raster topographic map: the raster topographic map obtained in step S41 is vectorized and input into a GAN network. The global features of the raster map are extracted by downsampling, and the spatial resolution of the raster map is gradually restored by upsampling. A vectorized, smoothed and completed raster topographic map is output. Figure 3 As shown, the following is a detailed description:

[0068] First, the input grid passes through a convolutional layer with 64 filters of size 5x5 and a stride of 1. The convolution operation slides the filters across the grid, extracting low-level features. A batch normalization layer then normalizes the convolution output to ensure numerical stability during training. Next, a ReLU activation function sets all negative values ​​to zero, introducing nonlinearity.

[0069] The next two convolutional layers gradually increase the depth of the network and the complexity of feature extraction. Convolutional layer 2 downsamples the spatial dimensions using 128 3x3 filters and convolution operations with a stride of 2, halving the size of the feature map. Convolutional layer 3 further extracts and processes high-level features of the grid using 128 3x3 filters and convolutions with a stride of 1. Each layer uses batch normalization and ReLU activation functions for normalization and nonlinear mapping.

[0070] After multiple layers of convolution, the network enters the dilated convolution layer. The purpose of dilated convolution is to expand the receptive field of the convolution kernel without increasing computational complexity, thereby capturing a wider range of contextual information. This is crucial for the global structure of the raster image. Dilated convolution is also equipped with batch normalization and ReLU activation.

[0071] After the dilated convolution, the network recovers the details and texture information of the grid through convolution layer 4. The convolution operation of this layer continues to extract features at a higher level, and the number of filters is kept at 256 to capture enough details.

[0072] To restore the feature map to the same spatial resolution as the input map, the network uses deconvolution operations. Deconvolution layers 1 and 2 use 128 and 64 filters respectively, and upsample the feature map by deconvolution operations with a stride of 2, doubling its size.

[0073] The core task of this network is to extract and generate a high-resolution, reliable raster map from the low-level features of the input raster image. Through multiple layers of convolution, dilated convolution, and deconvolution operations, the network is able to capture global context while preserving the details of the original terrain map, ultimately generating a smoothed and completed raster terrain map similar to the input raster.

[0074] S5. Calculate traversability. Extract the normal vectors within each unit grid of the grid topographic map obtained in step S4. Based on the normal vectors and the precise robot pose obtained in step S3, calculate the slope, slope change rate, and roughness within each unit grid. Combined with the image semantic information obtained in step S2, calculate the traversability within each unit grid, and maintain a real-time traversable map for the legged robot using an extended Kalman filter. This is described in detail below:

[0075] S51. Extract the normal vector of the unit grid. In the constructed grid map, each grid contains several point clouds. A plane is fitted from these point clouds. Specifically, the present invention uses the least squares method to fit a plane. The normal vector of the fitted plane is the normal vector of the grid. Finally, the normal vector is normalized to the unit normal vector n to facilitate subsequent analysis and processing.

[0076] S52: Calculate the slope and roughness of each grid based on the normal vector obtained in step S51. The specific calculation formula is as follows:

[0077] s=arccos(n,n z )

[0078]

[0079] Among them, s is the slope, r is the roughness, n is the grid normal vector, and n z is the ground normal vector, n i is the neighborhood grid normal vector, and m is the custom neighborhood size.

[0080] S53: Based on the normal vector slope and roughness of each unit grid and its surrounding grids obtained in step S52, weighted calculation of each unit cell's passability is performed. The specific calculation formula is as follows:

[0081]

[0082] Where trav is the traversability, with a value range of [0, 1]. a and b are custom scaling parameters whose values ​​are determined by the semantic information obtained in step S24. Different semantic regions correspond to different parameters. This design enables the robot to autonomously distinguish between different scenarios, such as stairs, grass, and concrete, when calculating traversability, making it more valuable in practical applications.

[0083] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for constructing a traversable map for a legged robot, characterized in that: The following steps are involved: S1. Obtain the original point cloud information around the robot through the lidar, obtain the inertial measurement information of the robot body through the inertial measurement unit, obtain the longitude and latitude information of the robot body through the satellite signal receiver, and obtain the image information in front of the robot through the camera; S2. Preprocessing the original point cloud information, inertial measurement information, latitude and longitude information, and camera image information obtained in step S1; wherein the preprocessing includes original point cloud information preprocessing, IMU pre-integration processing, UTM coordinate system projection of latitude and longitude information, and image semantic information extraction; S3. Performing error state Kalman filtering on the point cloud information and inertial measurement information preprocessed in step S2 to obtain a rough pose of the robot; using the projected longitude and latitude information as a priori constraints and the rough pose of the robot as the factor to be optimized, obtaining the precise pose of the robot through a graph optimization method; S4, constructing a 2.5D grid topographic map using the extended Kalman filter method for the point cloud information preprocessed in step S2 and the precise position and posture of the robot obtained in step S3, and processing the topographic map using a convolutional neural network to obtain a smooth grid topographic map without holes; S5. For the grid topographic map obtained in step S4, extract the normal vector within each unit grid, and calculate the slope, slope change rate and roughness within each unit grid based on the normal vector and the precise position of the robot obtained in step S3; combine the image semantic information obtained in step S2 to calculate the passability within each unit grid, and maintain a leg-foot robot passable map in real time through extended Kalman filtering.

2. The method for constructing a traversable map for a legged robot according to claim 1, wherein: The step S2 includes the following sub-steps: S21. Preprocessing of the original point cloud information: performing outlier filtering and motion distortion correction on the original point cloud information, and removing dynamic point clouds to obtain a reliable point cloud; S22, IMU pre-integration processing: Based on the IMU angular velocity and acceleration data, the relative motion between the two frames during measurement is calculated to obtain the robot posture change; S23. Projection of longitude and latitude information into the UTM coordinate system: The satellite receiver interprets the message sent by the Global Navigation Satellite Positioning System (GNSS) as longitude and latitude information based on the WGS84 geocentric coordinate system, and then converts the longitude and latitude information into the UTM coordinate system, i.e., the XYZ three-axis coordinate information in the robot's local coordinate system. S24. Image semantic information extraction: Extract the features of the original image information through a convolutional neural network, and use the fully connected layer to output the category and bounding box of the object to achieve semantic information extraction.

3. The method for constructing a traversable map for a legged robot according to claim 2, wherein: The step S3 includes the following sub-steps: S31, inferring the initial pose of the current frame based on the credible point cloud obtained in step S2 and the pose change between the two frames of the robot obtained by IMU pre-integration; transforming the initial pose to the coordinate system of the previous frame, and constructing an error model through the geometric registration relationship between the two frames of point clouds to obtain the optimal robot pose state, that is, the rough pose of the robot; S32. Using a graph optimization method, with the rough posture of the robot as the vertex and the longitude and latitude information obtained in step S2 as the constraint edge, a graph model is constructed to solve the graph optimization problem and obtain the precise posture state of the robot.

4. The method for constructing a traversable map for a legged robot according to claim 1, wherein: The step S4 includes the following sub-steps: S41, based on the credible point cloud obtained in step S2 and the precise position state of the robot obtained in step S3, obtain a 2.5D grid topographic map by using an extended Kalman filter; wherein each grid contains a number of points, an average height value, and a variance estimate obtained by the Kalman filter; S42. The raster topographic map obtained in step S41 is vectorized and input into a GAN network. Global features of the image are extracted by downsampling, and the spatial resolution of the image is gradually restored by upsampling. A vectorized, smoothed, and completed raster topographic map is output.

5. The method for constructing a traversable map for a legged robot according to claim 1, wherein: The step S5 includes the following sub-steps: S51, performing plane fitting on a plurality of point clouds in each unit grid according to the grid topographic map obtained in step S4, and taking the normal vector of the fitted plane as the normal vector of the unit grid; S52, calculating the slope and roughness of each grid based on the normal vector of each grid extracted in step S51 and the normal vectors in its neighboring grids; S53 , weightedly calculating the passability of each cell according to the parameters of each cell grid and its surrounding grids obtained in step S52 .

6. The method for constructing a traversable map for a legged robot according to claim 5, wherein: In step S53 , the weight parameters in the weighted calculation of the passability are determined by the image semantic information obtained in step S2 , and the weights of the parameters corresponding to different semantic regions are different.

Images