An extraterrestrial object rover positioning method based on terrain profile matching

Abstract

The application discloses a method for locating an extraterrestrial rover based on terrain profile matching and belongs to the technical field of space robots. The application is composed of two threads of relative positioning by an odometer and absolute positioning by projection search. The relative positioning is always run, and the absolute positioning is run at a proper time. The relative positioning thread runs a visual odometer, receives an RGB image, and accumulates an initial value of the pose of the rover camera by estimating interframe motion. The purpose of the absolute positioning is to correct the accumulated error of the relative positioning in the process of a task or to find the global positioning when real-time positioning is lost. In the process of the rover running, whether there is a clear mountain profile in the field of view as a reference to ensure the positioning reliability is judged by an elevation map and the position of the rover. According to the camera internal parameter and the external parameter, a map point cloud is projected to each pose node to obtain a projection map. The projection map of each pose is matched with the image profile, the highest matching curve is selected, and the final global optimal pose is selected.

Description

Technical Field

[0001] This invention relates to a method for locating extraterrestrial rovers based on terrain contour matching, and more particularly to a method for locating rovers in situations where the mission area is uneven, lacks texture, and has limited sensor conditions, thus belonging to the field of space robotics technology. Background Technology

[0002] With the continuous development of deep space exploration technology, various wheeled rovers have been sent to the lunar and Martian surfaces to perform exploration, image acquisition, and soil sample collection tasks. Traditional rovers are controlled through a space-to-ground loop, where sensor information is transmitted back to the ground station for expert decision-making, and then control commands are transmitted back. Due to the time delay in space-to-ground transmission, this control method is cumbersome and time-consuming. Therefore, in recent years, autonomous rovers have become a research hotspot. Among them, accurate positioning is a prerequisite for path planning and obstacle avoidance and is one of the most important modules of the rovers. However, the positioning of lunar and Martian rovers faces the following challenges: (1) The uneven surface of the lunar and Martian surfaces leads to large errors in wheeled odometers, and the uneven terrain easily causes the rovers to bounce, resulting in inaccurate visual positioning; (2) The lack of lunar surface environmental features and the uniform texture lead to large errors in relative positioning methods; (3) Due to power and load limitations, the types and number of sensors equipped on rovers are limited, and most small rovers are only equipped with color cameras, which poses a challenge to the robustness of the positioning system. Under the above constraints, to ensure the accuracy and robustness of the positioning system, it is necessary to utilize prior map information for global positioning. In the initial stages of extraterrestrial missions, orbiting satellites are often used to create digital elevation maps of the mission area. These elevation maps, as prior map information, can provide a reference for the rover's global positioning.

[0003] Current research on autonomous localization technology for lunar rovers can be broadly categorized into two types: relative localization methods based on odometry and absolute localization methods based on pattern recognition. Visual odometry is a subproblem of SLAM (Simultaneous Localization and Mapping). It directly acquires 3D point information from the environment using stereo vision algorithms or depth sensors, estimates the inter-frame motion of the camera using the motion of matching 2D points in the image, and accumulates the motion of each frame to obtain the camera's pose relative to the initial moment. Some studies have proposed a visual localization scheme for rovers in lunar environments. This scheme uses visual and inertial odometry for relative localization, effectively addressing the problem of poor robustness in localization under low-texture conditions on the lunar surface. However, odometry is a relative localization method. In lunar exploration scenarios, rovers typically need to travel distances of several kilometers, making accumulated errors non-negligible. Global localization is then required to correct for these accumulated odometry errors.

[0004] Global positioning refers to determining the rover's pose relative to a global coordinate system. Since the rover can only perform relative positioning locally, additional methods are needed to provide global positioning information. Global positioning methods fall into two categories. One is through satellite radio communication, where a lunar orbiter communicates with the rover, achieving meter-level positioning accuracy. However, this method relies on stable communication between the satellite and the rover, imposing strict constraints on the rover's trajectory and preventing real-time, on-the-go global positioning during missions. The other method uses known global features from a map as prior knowledge, matching the rover's local observations with these features. Some studies have used the rover's onboard star sensor to observe the sun's azimuth, obtaining accurate azimuth angle estimations, but the position accuracy is only at the kilometer level, failing to meet the rover's navigation requirements. Furthermore, other studies have used meteorite craters in satellite maps as landmarks, matching the reconstructed crater shapes from sensors such as vision and lidar with the craters on the map to obtain the rover's global positioning, with azimuth errors meeting navigation and obstacle avoidance standards. However, this approach requires a high variety of sensors on the rover, failing to meet the rover's lightweight design requirements. Summary of the Invention

[0005] The main objective of this invention is to provide a method for locating extraterrestrial celestial bodies based on terrain contour matching. This method uses satellite digital elevation maps and the rover's visual odometry for positioning. It generates a pose search network near the initial positioning value, projects the elevation point cloud onto each pose node, and performs feature matching between the projected contour and the camera observation contour to search for the optimal pose. This method has the advantages of strong robustness and high positioning accuracy.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] This invention discloses a method for locating an extraterrestrial rover based on terrain contour matching, consisting of two threads: odometry relative positioning and projection search absolute positioning. Relative positioning runs continuously, while absolute positioning runs selectively. The relative positioning thread runs visual odometry, receives RGB images, and accumulates the initial pose of the rover's camera by estimating inter-frame motion, providing real-time, continuous relative positioning. The purpose of absolute positioning is to correct accumulated errors in relative positioning during the mission, or to recover global positioning when real-time positioning is lost due to rover turbulence, rollover, etc. During rover operation, the presence of obvious mountain contours within the field of view is determined using elevation maps and the rover's orientation to ensure positioning reliability. When craters are present within the rover's field of view, the absolute positioning module is activated. The absolute positioning module first reconstructs a point cloud map from the digital elevation map through interpolation and scaling. Using the pose provided by relative positioning as the search center, a three-dimensional search network (planar coordinates x, y and orientation angle γ) is constructed. Based on camera intrinsic and extrinsic parameters, the map point cloud is projected onto each pose node to obtain a projected image. Contour extraction is performed between the projected image and the camera image, and feature matching is performed between the projected image and the image contours for each pose. The curve with the highest matching degree is selected, and the pose corresponding to its node is the final global optimal pose.

[0008] The present invention discloses a method for locating extraterrestrial celestial bodies based on terrain contour matching, comprising the following steps:

[0009] Step 1: By running the optical flow front-end, extract optical flow feature points, solve the inter-frame motion of the camera, and accumulate the inter-frame motion to give the rovers pose relative to the starting point, expressed in Cartesian coordinates and Euler angles as follows:

[0010] pose={x,y,z,γ,β,α} (1)

[0011] Where x, y, z are rectangular coordinates, γ, β, α are Euler angles, and the corresponding transformation order is zyx, with the directions following the right-hand rule.

[0012] Step 2: Process the digital elevation map, obtaining a 2D raster map through slicing. This raster map reflects the distribution of craters in the mission area. Based on the camera's field of view and the rover's pose obtained in Step 1, determine whether there are craters within the camera's field of view as a reference. In the planar map, the rover's six-DOF pose degenerates into a three-DOF pose.

[0013] pose'={x,y,γ} (2)

[0014] A ray on the plane is determined based on pose', representing the position of the rover and the direction of its camera field of view; the camera's line of sight is parallel to the rover's axis. Considering the camera's field of view angle fov, the camera's field of view is represented as several rays covering the field of view range. The expression for the i-th ray is:

[0015]

[0016] Where n is the total number of rays, and ray() is the mapping from orientation to ray; perform collision detection between the n rays and the black grid representing the crater in the grid map, calculate the ray length, i.e. the distance between the crater obstacle and the rovers; when a crater appears in the camera's field of view and the distance is less than the set threshold, proceed to step three and start the absolute positioning thread;

[0017] Step 3: Using the initial position given by relative positioning as the center, generate a 3D pose search network, represented as...

[0018]

[0019] Where, n i ,n j ,n k Let x be the search radius for each dimension of the network. i ,y j ,γ k ) represents the pose node;

[0020] First, the azimuth angle is sampled near its initial value.

[0021] γ k =γ0+δ γ ·(kn k (5)

[0022] Where, δ γ γ0 is the azimuth search step size, k is the search traversal value, and γ0 is the initial azimuth value.

[0023] For each azimuth angle γ k Position sampling is performed, with cross-searches conducted in both the azimuth and normal directions.

[0024]

[0025] Where δ is the distance search step size; thus, the pose search network S is constructed;

[0026] Step 4: Reconstruct the digital elevation map into a 3D point cloud. The digital elevation map is in the form of (u, v, d), where the first two dimensions are pixels and the last dimension is height. Map the pixels to 3D coordinates using the following formula.

[0027]

[0028] Where λ is the scaling factor of the elevation map; u and v are the horizontal and vertical pixels, respectively, and d is the pixel elevation;

[0029] Then, the land Figure 3Projecting the 3D point cloud onto each node s of the pose search network S i,j,k Above, the projection transformation is

[0030] p uv =KTP (8)

[0031] Where, p uv Here, K represents the pixel coordinates of the projection point, T represents the homogeneous transformation matrix of the camera coordinate system relative to the world system, and P represents the map point cloud.

[0032] After obtaining the projected scatter plot, the contour curves of the projected plot are extracted using an edge extraction algorithm. The same operation is performed on all nodes of the search network S to obtain the contour set. Extract the contour curve from the camera image, denoted as e. img A contour e is represented as

[0033]

[0034] Among them, (u i ,v i ) represents the pixel coordinates of the i-th node in the contour, and n represents the total number of nodes.

[0035] Step 5: Construct an azimuth pyramid to eliminate the influence of rover roll and pitch on matching. Expand the dimension of the pose search network.

[0036]

[0037] Where, x i ,y j ,γ k The construction method is the same as described in step three, n l Let α be the sampling radius for the roll angle. l It is the sampled value of the l-th roll angle, constructed as

[0038] α l =α0+δ α ·(ln l (11)

[0039] Where α0 is the initial value of the roll angle, δ α The roll angle is the search step size. The contour set is obtained by projecting S' using the method in step four. and image outline e img Because the probe's pitch changes the longitudinal position of the curve without altering its shape, for each projected profile e, it is longitudinally translated so that its first node aligns with e. img Overlap, that is:

[0040]

[0041] Among them, v i and v′ i These are the vertical pixel coordinates of the i-th node of the contour before and after translation, v0 and v1, respectively. These are the vertical pixel coordinates of the first node of the projected contour and the image contour, respectively. Let e' be the vertically aligned projected contour. The vertically aligned contour eliminates the influence of the inspector's pitch on the matching.

[0042] Connect e' with e img A similarity matching algorithm is used to obtain the similarity network for each pose node.

[0043]

[0044] Where, d i,j,k,l For outline e i,j,k,l The corresponding similarity.

[0045] Eliminate rolling dimensions by selecting the optimal value.

[0046]

[0047] Where, d′ i,j,k This represents the similarity of each three-degree-of-freedom pose node.

[0048] The index of the pose node that yields the best match is:

[0049]

[0050] The optimal pose is

[0051]

[0052] The above optimization method can be used to improve the accuracy and robustness of the positioning of extraterrestrial astronauts.

[0053] Beneficial effects:

[0054] 1. The present invention discloses a method for locating extraterrestrial celestial bodies based on terrain contour matching. By reducing the dimensionality of a three-dimensional digital elevation map through contour projection and matching two-dimensional features, the global location of the celestial body is obtained. Compared with other global positioning methods, the positioning accuracy is high and the requirements for sensor configuration are low.

[0055] 2. This invention discloses a method for locating extraterrestrial celestial bodies based on terrain contour matching. It employs a "relative + absolute" positioning method. Relative positioning outputs real-time, continuous locator positioning via visual odometry; absolute positioning derives the locator's absolute position through template matching of mountain contours, calibrating the locator's cumulative error. This invention effectively improves the accuracy and robustness of the positioning system while ensuring real-time positioning.

[0056] 3. This invention discloses a method for locating extraterrestrial celestial bodies based on terrain contour matching. It employs contour alignment and the construction of an "attitude pyramid" to select each node s. i,j,k The optimal attitude angle is found to eliminate the influence of roll and pitch angles on profile curve matching. Attached Figure Description

[0057] Figure 1 A schematic diagram of an extraterrestrial rover positioning method based on terrain contour matching disclosed in this invention.

[0058] Figure 2 The lunar raster map conversion process in this invention example; wherein Figure a is a digital elevation map; and Figure b is a raster map.

[0059] Figure 3 An example of this invention is a location suitable for absolute positioning.

[0060] Figure 4 The point cloud map obtained in the example of this invention.

[0061] Figure 5 The present invention provides examples of camera image contours and optimal projection contours; wherein Figure a is the camera image contour and Figure b is the projection contour.

[0062] Figure 6 The matching degree distribution of the poses of each node in the absolute positioning in the example of this invention. Detailed Implementation

[0063] To better illustrate the purpose and advantages of this invention, the following example analysis of the navigation and positioning of a rover in a certain mission area on the lunar surface will provide a detailed explanation of this invention.

[0064] Example 1:

[0065] During a lunar rover mission, the lunar lander released a small lunar rover equipped only with an RGB camera. The camera had a field of view of 90×60 degrees, an imaging resolution of 1280×720 pixels, and an image frame rate of 1 FPS. The rover's mission was to traverse an area with known elevation maps, and during its journey, the rover needed to maintain precise positioning.

[0066] Step 1: Given the initial position of the rover on the map, extract optical flow feature points by running the optical flow front-end, solve for the inter-frame motion of the camera, and accumulate the inter-frame motion to give the camera pose relative to the initial point, expressed in Cartesian coordinates and Euler angles as follows:

[0067] pose = {x, y, z, γ, β, α}

[0068] Among them, the Euler angles γ, β, α are transposed to zyx, and their directions follow the right-hand rule.

[0069] Step Two: Process the digital elevation map by slicing it to obtain a two-dimensional raster map. This raster map reflects the distribution of craters in the task area, such as... Figure 2 As shown. Based on the rover pose and camera field of view obtained in step one, it is determined whether a crater exists within the camera's field of view as a reference. In the planar map, the rover's six-degree-of-freedom pose degenerates into a three-degree-of-freedom pose.

[0070] pose = {x, y, γ}

[0071] This pose uniquely determines a ray on the plane, which, when reflected on the map, is the ray at the center of the rover's camera field of view. Assuming the camera's line of sight is parallel to the rover's axis, and considering the camera's field of view angle fov, the camera's field of view can be represented as several rays covering the field of view. The expression for the i-th ray is:

[0072]

[0073] Where n is the total number of rays, and ray() is the mapping from orientation to ray. Collision detection is performed between the two rays and the black grid representing the crater in the grid map. The distance between the crater obstacle and the rovers is obtained by calculating the ray length. When a crater appears in the camera's field of view and the distance is less than a certain threshold, the absolute positioning thread is started. Figure 3 This demonstrates a scenario that meets the conditions for absolute positioning, with 3 sampling rays, a threshold range of 8-15m, and all three field lines intersecting the grid map at distances within the specified range.

[0074] Step 3: Using the initial position given by relative positioning as the center, generate a 3D pose search network, which can be represented as follows:

[0075]

[0076] Where, n i ,n j ,n k To determine the search radius for each dimension of the network, the azimuth angle is first sampled near its initial value.

[0077] γ k =γ0+δ γ ·(kn k )

[0078] Where, δ γ Let γ be the azimuth search step size, for each azimuth angle γ k Position sampling is performed, with cross-searches conducted in both the azimuth and normal directions.

[0079]

[0080] Where δ is the distance search step size. Thus, the pose search network P is constructed.

[0081] Step 4: Reconstruct the digital elevation map into a 3D point cloud, such as... Figure 4 As shown. Then, the ground Figure 3 Projecting the 3D point cloud onto each node s of the pose search network S i,j,k Above. After obtaining the projected scatter plot, the terrain outline of the projected plot is extracted using an edge extraction algorithm, denoted as... At the same time, the mountain outlines in the camera images are extracted.

[0082] Step 5: By comparing the projected contour with the image contour one by one using a dynamic warping algorithm, the similarity between the two curves can be calculated. The pose node with the highest matching degree is selected as the globally optimal pose. Taking an absolute localization point in the task as an example... Figure 5 It shows the contour of the camera image at the point and the projected contour in the optimal pose, which has the highest similarity to the image contour. Figure 6 The distribution of matching degree of each node in the search network at this location is shown (the ground truth value is 8.46, -2.47). It can be seen that the algorithm can converge to the ground truth value, and the similarity distribution is smooth near the ground truth value, which proves the effectiveness and robustness of the algorithm.

[0083] During the mission, three locations were selected for absolute positioning, and the results before and after positioning are shown in Table 1. To simulate the initial position error, a 2m position error was set near the rover's starting position. During rover initialization, absolute positioning was run to adjust the rover's positioning error to within 1m. When the rover reached the middle area, the visual odometry accumulated positioning error. After absolute positioning, the positioning could still be adjusted back to within 1m accuracy, achieving an optimization rate of 80.6%. Upon reaching the vicinity of the target area, absolute positioning was run again to optimize the positioning to within 1.5m, still ensuring the positioning accuracy of the final target.

[0084] Table 1. Positioning before and after absolute positioning point optimization.

[0085]

[0086] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for locating extraterrestrial celestial bodies based on terrain contour matching, characterized in that: Includes the following steps, Step 1: By running the optical flow front-end, extract optical flow feature points, solve the inter-frame motion of the camera, and accumulate the inter-frame motion to give the rovers pose relative to the starting point, expressed in Cartesian coordinates and Euler angles as follows: in, Using rectangular coordinates, For Euler angles, the corresponding transposition is zyx, and the directions follow the right-hand rule; Step 2: Process the digital elevation map, obtaining a 2D raster map through slicing. This raster map reflects the distribution of craters in the mission area. Based on the camera's field of view and the rover's pose obtained in Step 1, determine whether there are craters within the camera's field of view as a reference. In the planar map, the rover's six-DOF pose degenerates into a three-DOF pose. according to Define a ray on the plane, representing the position of the rover and the direction of its camera field of view; the camera's line of sight is parallel to the rover's axis, taking into account the camera's field of view angle. Then, the camera's field of view is represented as several rays covering the field of view, and the expression for the i-th ray is: in, Let be the total number of rays, and ray() be the mapping from orientation to ray; Collision detection is performed between the ray and the black grid representing the crater in the grid map, and the ray length is calculated, which is the distance between the crater obstacle and the rovers; when a crater appears in the camera's field of view and the distance is less than a set threshold, step three is performed to start the absolute positioning thread; Step 3: Using the initial position given by relative positioning as the center, generate a 3D pose search network, represented as... in, , , Let be the search radius for each dimension of the network. For pose nodes; The azimuth angle is sampled near the initial value. in, Let k be the azimuth search step size, and k be the search traversal value. This is the initial value of the azimuth angle; For each azimuth angle Position sampling is performed, with cross-searches conducted in both the azimuth and normal directions. in, Let S be the distance search step size; this completes the pose search network S. Step 4: Reconstruct the digital elevation map into a 3D point cloud. The digital elevation map is as follows: In the form of a graph, the first two dimensions are pixels and the last dimension is height. The following formula maps pixels to three-dimensional coordinates. in, d represents the scaling factor of the elevation map; u and v represent the horizontal and vertical pixels, respectively, and d represents the pixel elevation. Project the 3D point cloud of the map onto each node of the pose search network S. Above, the projection transformation is in, Here, K represents the pixel coordinates of the projection point, T represents the homogeneous transformation matrix of the camera coordinate system relative to the world system, and P represents the map point cloud. After obtaining the projected scatter plot, the contours of the projected plot are extracted using an edge extraction algorithm; the same operation is performed on all nodes of the search network S to obtain the contour set. Extract the contours from the camera image, denoted as... A contour e is represented as in, Let be the pixel coordinates of the i-th node in the contour, and n be the total number of nodes; Step 5: Construct an azimuth pyramid to eliminate the effects of rover roll and pitch on the matching; Dimensional expansion of pose search network in, The construction method is the same as described in step three. The sampling radius is the roll angle. It is the sampled value of the l-th roll angle, constructed as in, This is the initial value of the roll angle. This refers to the search step size based on the roll angle; Will The contour set is obtained by projection using the method in step four. and image outline Because the probe's pitch changes the longitudinal position of the curve without altering its shape, for each projected profile... Translate it vertically so that its first node is aligned with... Overlap, that is: in, and These are the vertical pixel coordinates of the i-th node of the contour before and after translation. and Let be the vertical pixel coordinates of the first node of the projected contour and the image contour, respectively; let the vertically aligned projected contour be denoted as . The longitudinally aligned profile eliminates the influence of the rover's pitch on the matching. Will and A similarity matching algorithm is used to obtain the similarity network for each pose node. in, For outline Corresponding similarity; Eliminate rolling dimensions by selecting the optimal value. in, This represents the similarity of each three-degree-of-freedom pose node; The index of the pose node that yields the best match is: （15） The optimal pose is （16）。

Images