Moon long-distance path planning method fusing terrain, illumination and abundance features

Abstract

The application belongs to the technical field of path planning, and discloses a lunar long-distance path planning method fusing terrain, illumination and abundance features. First, a safety map is obtained by comprehensively analyzing DEM images, average illumination rate images and rock abundance images using safety evaluation rules. Then, the safety map is constructed into a tile pyramid, and is divided into tiles according to layers and stored in HDFS. Then, the global planning path is obtained by performing distributed calculation on each layer of tiles from top to bottom using a Spark distributed calculation engine. Finally, the global planning path is combined with DOM images to implement fine periodic planning.

Description

Technical Field

[0001] This invention belongs to the technical field of path planning, specifically relating to a lunar long-distance path planning method that integrates terrain, illumination, and abundance features. Background Technology

[0002] The Moon is a crucial target for human deep space exploration and an ideal base and outpost for conducting extraterrestrial space exploration. As Earth's only natural satellite, the Moon's unique geological structure and surface characteristics record the early evolutionary history of the solar system. Sampling and analyzing lunar surface materials can provide insights into planetary formation processes and the evolutionary mechanisms of the Earth-Moon system. Furthermore, the permanently shadowed regions at the lunar poles may be rich in volatile substances such as water ice, providing crucial resource guarantees for future deep space exploration. Against this backdrop, lunar rovers, as core equipment in lunar exploration missions, are gradually evolving towards multi-functional integration, possessing comprehensive capabilities such as all-terrain traversal, resource exploration and utilization, manned exploration, and large-scale movement. However, the complex terrain and geomorphological environment of the lunar surface pose severe challenges to lunar rover exploration missions, making safe and efficient route planning a critical prerequisite for lunar rover mobile exploration.

[0003] Significant progress has been made in lunar path planning research, but several shortcomings remain. First, most path planning studies rely on Digital Elevation Models (DEMs) for path calculations. However, the resolution of full-lunar DEMs is relatively low, failing to fully reflect the complex terrain features of the lunar surface, such as the distribution of small craters and rocks. These unconsidered environmental factors may pose a threat to the safety of lunar rovers. Second, as the lunar exploration range expands from small local areas to tens or even hundreds of kilometers, the computational complexity of path planning increases exponentially, making it difficult for single-machine computing power to efficiently handle large-scale path planning in complex environments. Furthermore, current path planning methods consider relatively singular environmental factors, mostly optimizing solely based on terrain, without comprehensively considering other important factors that may affect path safety and energy efficiency, such as illumination and rock abundance. Summary of the Invention

[0004] This invention proposes a lunar long-distance path planning method that integrates terrain, illumination, and abundance features. It solves the technical problems of current path planning methods that consider only one environmental factor, and most of them are optimized based only on terrain factors, without comprehensively considering other important factors that may affect path safety and energy efficiency, such as illumination and rock abundance.

[0005] This invention can be achieved through the following technical solutions:

[0006] A lunar long-distance path planning method integrating terrain, illumination, and abundance features is proposed. First, a safety map is obtained by comprehensively analyzing DEM imagery, average illumination imagery, and rock abundance imagery using safety evaluation rules. Then, the safety map is constructed into a tile pyramid, divided into layers and stored in HDFS. Next, the Spark distributed computing engine is used to perform distributed computing on each layer of tiles from top to bottom to obtain the global planned path. Finally, the global planned path is refined and periodically planned by combining DOM imagery.

[0007] Furthermore, when using the safety evaluation rules for comprehensive analysis, the mobility analysis is first performed on the three images: DEM image, average illuminance image, and rock abundance image, generating four mobility maps. Then, the intersection of these four mobility maps is performed to obtain a single mobility map. Finally, this mobility map is convolved to obtain a safety map.

[0008] The accessibility map has only two values, 0 and 1. 0 represents that the point is an obstacle and cannot be passed; 1 represents that the point can be passed.

[0009] Furthermore, the slope angle θ and roughness δ of each pixel in the DEM image, the average illuminance μ of each pixel in the average illuminance image, and the rock abundance of each pixel in the rock abundance image were calculated respectively. Generate four accessibility maps according to the following conditions;

[0010] When the slope angle θ is greater than the first threshold, it is considered impassable and its value is 0; when it is less than the first threshold, its value is 1, in order to generate a slope accessibility map M. slope ;

[0011] When the roughness δ is greater than the second threshold, it is considered impassable and its value is 0; when it is less than the second threshold, its value is 1, in order to generate a roughness accessibility map M. rough ;

[0012] When the average illuminance μ is less than the first threshold, it is considered impassable and its value is 0; when it is greater than the first threshold, its value is 1, in order to generate an illuminance accessibility map M. illumin ;

[0013] When the rock abundance φ is greater than the first threshold, it is considered impassable and its value is 0; when it is less than the first threshold, its value is 1, in order to generate a rock abundance accessibility map M. rock .

[0014] Furthermore, according to the following formula, the slope accessibility map M is... slope Roughness and accessibility map M rough Light Accessibility Map M illuminRock abundance accessibility map M rock Perform intersection operations to obtain a traversability map M. traverse ;

[0015] M traverse =M slope ∩M rough ∩M illumin ∩M rock

[0016] According to the following formula, the drivability map M traverse The security map M is obtained by performing convolution calculations. safety ,

[0017]

[0018] Among them, M traverse (i,j) represents the drivability map M. traverse The pixels (i,j) on the graph are given by K(u,v), where K(u,v) represents the element in the convolution kernel K, and M represents the pixel (i,j). safety (i,j) represents the convolution result of kernel K at pixel (i,j).

[0019] Furthermore, the A* algorithm is improved in terms of both cost function and data structure, and then the improved A* algorithm is used for path planning on each tile layer.

[0020] Furthermore, the cost function F(N) of the improved A* algorithm is calculated using the following equation.

[0021] F(N)=G(N)+(1.5-w×M safety (N))×H(N)

[0022] Where F(N) represents the total cost from the starting point to the ending point, G(N) represents the actual cost from the starting point to the current node N, H(N) represents the estimated cost from the current node N to the ending point, w represents the static weight, and M... safety (N) represents the traversability of the current point N in the security map.

[0023] Furthermore, in the improved A* algorithm, the data structure of the open table is implemented using a min-heap and a hash table, while the data structure of the close table is implemented using a hash table.

[0024] Furthermore, refined periodic planning includes the following steps:

[0025] S1. The Bresenham algorithm is used to simplify the path nodes on the global planning path to obtain a simplified path;

[0026] S2. For the simplified path, take every two adjacent path nodes as a cycle, determine the latitude and longitude range of each path segment, and extract the high-resolution DOM image and security map of the corresponding area.

[0027] S3. The CenterNet network model based on deep learning is used to detect small craters in the extracted high-resolution DOM images.

[0028] S4. The detected small meteorite craters are used as obstacles and superimposed on the extracted safety map. Then, the improved A* algorithm is used to re-plan the path to achieve periodic planning.

[0029] Furthermore, when using the Bresenham algorithm to simplify the path nodes on the global planning path, starting from the starting point, each path node in the path is traversed sequentially. It is determined whether the connection between the current starting point and the traversed node passes through or is close to an obstacle point. If not, the traversal continues; if it exists, the previous path node of the traversed node is added to the key point list, and the previous path node is set as the starting point of the next round of traversal. The above process is repeated until the destination is reached, thus achieving path simplification.

[0030] The beneficial technical effects of this invention are as follows:

[0031] 1. Taking into account factors such as lunar terrain slope, roughness, average illumination, rock abundance, and the distance and density of obstacles around image pixels, a set of safety evaluation rules based on the lunar environment was constructed. The safety map generated based on these rules can effectively quantify the safety of various regions on the lunar surface and provide support for path planning.

[0032] 2. An improved A* algorithm was designed. This method improves the cost function and data structure based on the traditional A* algorithm, reducing the time and space complexity and significantly improving the safety and efficiency of path planning. Furthermore, a distributed path planning method based on a security map tile pyramid (DPPS-STP) is proposed, which reduces computation time while ensuring path safety, showing a significant advantage, especially in large-scale path planning. Experimental results show that in long-distance path planning tasks, compared with the OC-WHT-A* algorithm in a single-machine environment, the DPPS-STP algorithm using the OC-WHT-A* algorithm of this invention has a computational speedup of 11.53, significantly improving the speed of path planning.

[0033] 3. The overall planned path is optimized based on the Bresenham algorithm to reduce the path length and turning angle. Combined with high-resolution DOM imagery, CenterNet is used to detect small crater obstacles to perform more refined periodic path planning, thereby further improving the safety of path planning. Attached Figure Description

[0034] Figure 1 This is a structural diagram of the method proposed in this invention;

[0035] Figure 2 This is a schematic diagram of the structure of the convolution kernel K of the present invention;

[0036] Figure 3 This is a flowchart of the security map generation process of the present invention;

[0037] Figure 4 This is a schematic diagram of the weighted A* algorithm for open and closed lists based on hash table data structure according to the present invention;

[0038] Figure 5 This is a schematic diagram illustrating the impact of different weights on path selection according to the present invention;

[0039] Figure 6 This is a schematic diagram of the iterative calculation process of the tile pyramid model of the present invention;

[0040] Figure 7 This is a schematic diagram of the simplified path based on the Bresenham algorithm of the present invention;

[0041] Figure 8 This is a schematic diagram of the CenterNet crater detection method of the present invention;

[0042] Figure 9 This invention provides a path comparison of different A* algorithms in a single-machine environment.

[0043] Figure 10 A comparison of path planning time for single-machine and distributed A* algorithms in three regions of this invention;

[0044] Figure 11 This invention provides a comparison of single-machine and distributed path planning in three regions. Detailed Implementation

[0045] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings and preferred embodiments.

[0046] like Figure 1 As shown, this invention provides a lunar long-distance path planning method that integrates terrain, illumination, and abundance features. First, a safety map is obtained by comprehensively analyzing DEM images, average illumination images, and rock abundance images using safety evaluation rules. Then, the safety map is constructed into a tile pyramid, divided into tiles by layer and stored in HDFS. Next, the Spark distributed computing engine is used to perform distributed computing on each layer of tiles from top to bottom to obtain the global planned path. Finally, combined with DOM images, the global planned path is subjected to fine-grained periodic planning.

[0047] Specifically as follows:

[0048] Step 1: Construct a security map

[0049] like Figure 3 As shown, the safety map is generated by comprehensively analyzing DEM imagery, average illumination imagery, and rock abundance imagery using safety evaluation rules. First, considering lunar slope, roughness, average illumination, and rock abundance, accessibility analysis is performed on each of these three images, generating four accessibility maps corresponding to these factors. Then, the intersection of these four accessibility maps is performed to obtain a single accessibility map. Finally, this accessibility map is convolved to obtain the safety map. Notably, these accessibility maps only have two values: 0 and 1. 0 represents a point as an obstacle and is impassable; 1 represents a point that is passable.

[0050] (1) Slope angle

[0051] The slope angle is a crucial parameter of lunar surface topography, directly reflecting its steepness. Excessive slope not only reduces the stability of the lunar rover but may also lead to slippage or rollover risks. This study uses the 8-neighborhood method to calculate the slope, and the formula for calculating the slope angle θ is as follows:

[0052]

[0053] Among them, f x and f y H represents the gradient of the DEM image in the horizontal and vertical directions, respectively. i This represents the elevation value of the i-th of the surrounding 9 grid cells, where Cellsize is the size of the image grid.

[0054] Slope Accessibility Map M slope It is generated under the following conditions: when the slope angle θ is greater than the threshold, it is considered impassable and its value is 0; when it is less than the threshold, its value is 1.

[0055] (2) Roughness

[0056] Roughness represents the degree to which elevation points within a region deviate from the fitted plane, reflecting the terrain's undulations. Excessive roughness can affect the safety and stability of the lunar rover. The formula for calculating roughness δ is as follows:

[0057]

[0058] in This represents the average of all elevation values ​​in the above 9 grids.

[0059] A roughness-based accessibility map M is generated based on the following conditions. roughWhen the roughness δ is greater than the threshold, it is marked as 0, indicating that it is impassable; otherwise, it is marked as 1.

[0060] (3) Average light illuminance

[0061] Sunlight is essential for lunar rovers to maintain warmth and receive sufficient solar energy. This invention uses lunar annual average illuminance data. In actual missions, sunlight varies over time. Due to the long duration of long-distance exploration missions, the average illuminance μ can reflect the actual lighting conditions during the mission to a certain extent. It can be obtained based on image analysis of the average illuminance. By limiting the average illuminance, areas with poor lighting conditions on the moon can be effectively avoided.

[0062] If an area remains in shadow for an extended period, and its average illuminance is below a threshold, the illuminance accessibility map M... illumin Then mark this region's value as 0; otherwise, mark it as 1.

[0063] (4) Rock abundance

[0064] Rocks are one of the main features of the lunar surface and can pose a potential hazard to landers and rovers. Rock abundance Defined as the proportion of the total area covered by rocks with a diameter greater than or equal to a specific value (usually 10 cm) within a certain region, it can be obtained based on rock abundance image analysis.

[0065] When the abundance is greater than the threshold, the rock abundance accessibility map M rock This region is marked as 0, indicating that it is impassable; if the abundance is less than the threshold, it is marked as 1.

[0066] Based on the mobility analysis of the above four factors, four corresponding mobility maps were obtained.

[0067] By analyzing four accessibility maps M slope M rough M illumin and M rock Perform intersection operations to obtain the final drivability map M. traverse That is, only pixels that simultaneously meet all four conditions for passage are marked as map M. traverse The passable points are listed below.

[0068] M traverse =M slope ∩M rough ∩M illumin ∩M rock #(5)

[0069] Based on the current dynamic parameters of the Yutu lunar rover, the maximum slope threshold is set to 20°, and the roughness threshold is set to Cellsize / 5. Referring to recommendations from similar missions, the rock abundance threshold is limited to 7%. After multiple experimental tests, the average illuminance threshold for the Oceanus Procellarum region and the Chang'e-4 landing area is limited to 44%, and the average illuminance threshold for the Antarctic region is limited to 7%.

[0070] Due to the accessibility map M traverse The pixel values ​​only reflect whether the current grid cell is an obstacle point, and cannot fully reflect the density of obstacle points in the surrounding area. Furthermore, the traditional A* algorithm typically uses path length as an evaluation metric during path planning, without considering the distance between the planned path and obstacle points. This results in paths that are too close to obstacles, increasing the risk to the lunar rover. Therefore, this invention addresses this issue in the generated drivability map M. traverse Based on this, a 7x7 convolution kernel K is defined to calculate the security of pixels. Points closer to the center of the convolution kernel have higher weights. The convolution kernel K is as follows: Figure 2 .

[0071] For the accessibility map M traverse For each pixel on the map, perform the following operations (6) and (7). If the pixel is an obstacle (value 0), skip the operation. Finally, a security map M with a value range of 0 to 1 can be obtained. safety The larger the value, the higher the security of the point and the fewer obstacles around it; the smaller the value, the lower the security of the point and the more obstacles around it. When the value is 0, the point is an obstacle and cannot be passed. The security map M is calculated using convolution. safety The pixel values ​​can effectively reflect the density of obstacle points in the grid and its surrounding area.

[0072]

[0073] Among them, M traverse (i,j) represents map M traverse The pixels (i,j) on the graph are given by K(u,v), where K(u,v) represents the element in the convolution kernel K, and M represents the pixel (i,j). Safety (i,j) represents the convolution result of kernel K at pixel (i,j), S K This represents the sum of the weights of the convolution kernel K.

[0074] Step 2: Generate a global planning path using a distributed computing strategy.

[0075] S21. Improved A* Algorithm (OC-WHT-A*)

[0076] The A* algorithm determines node priority by calculating the evaluation function F(N) = G(N) + H(N) for each node, where G(N) is the actual cost from the starting point to the current node N, and H(N) is the heuristically estimated cost from the current node to the target node. Hong et al. proposed an improved A* algorithm (OC-RA-A*Algorithm) with an open-closed list random access data structure. This algorithm uses a min-heap and a two-dimensional array to implement the open list, reducing the time complexity of determining whether an adjacent node is in the open list from O(n) to O(1). However, this method requires pre-allocating a large amount of memory for the two-dimensional array when processing large images. Furthermore, the nodes actually accessed by the A* algorithm are usually concentrated near the path from the starting point to the destination, meaning that even with a very large two-dimensional array, only a small portion of the space is actually used, resulting in a significant waste of memory resources.

[0077] This invention makes two improvements to the OC-RA-A* algorithm: First, it improves the data structure of the open and close tables in the algorithm; second, it incorporates the safety of path points into the calculation of the cost estimation in the A* algorithm.

[0078] First, by improving the data structure, such as Figure 4 As shown, while maintaining the average time complexity of finding a node in the open table, the space complexity of the algorithm is significantly reduced. Figure 4 As shown, in the improved A* algorithm (OC-WHT-A*Algorithm) based on the hash table data structure for open and closed lists, the data structure of the open list is implemented using a min-heap and a hash table, while the data structure of the close list is implemented using a hash table. Compared with OC-RA-A*, the average time complexity of OC-HT-A* in determining whether the adjacent nodes of the current node are in the open list is also O(1), but the space complexity is reduced from O(m... 2 The time complexity is reduced to O(n), where m is the image size and n is the number of nodes visited during the pathfinding process. The improved A* algorithm, with its enhanced data structure, effectively reduces space complexity while maintaining search efficiency.

[0079] Second, to address the problem of paths being too close to obstacles in the traditional A* algorithm, the safety of path nodes is incorporated into the cost estimation calculation based on the above improvements. The lower the safety of path node N, the higher the estimated cost of that point. This allows the A* algorithm to prioritize nodes with higher safety when searching for paths, avoiding areas with more obstacles and improving the overall safety of the path.

[0080] The cost function calculation formula for the improved A* algorithm (OC-WHT-A*) is as follows:

[0081] F(N)=G(N)+(1.5-w×M Safety (N))×H(N)#(8)

[0082] F(N) represents the total cost from the starting point to the ending point, G(N) represents the actual cost from the starting point to the current node N, H(N) represents the estimated cost from the current point N to the ending point, w is the static weight, and M... Safety H(N) represents the descentability of the current node N. When the descentability of a path node is high, the coefficient before H(N) is smaller, and the total cost F(N) of the path node is also smaller. Furthermore, different values ​​of w result in different path selection tendencies in the algorithm. Larger w leads to a slower algorithm, which tends to seek safer paths; smaller w leads to a faster algorithm, which tends to prioritize finding a path with a beginning and end point. The influence of weight w on the path is as follows: Figure 5 As shown.

[0083] S22, Global Planning Path Generation

[0084] This invention proposes a distributed path planning strategy DPPS-STP based on a security map tile pyramid, implemented using the distributed computing engine Spark. The core process is as follows: First, top-level pyramid tile data is read from HDFS, a tile RDD is constructed, and relevant tiles are filtered based on the input start and end points. Then, distributed path planning is performed on the filtered tiles to generate a path RDD for that layer. For non-bottom-level tiles, the start and end points of the next-level path planning are inferred using the generated path RDD, and path planning continues until all bottom-level tiles are processed. Finally, the path results are written to HDFS. The path planning on each tile layer is performed using the improved A* algorithm described below.

[0085] Leveraging the properties of the pyramid model, the starting and ending point sets for each tile are calculated from coarse to fine. First, coarse-grained path planning is performed starting from the top tile layer, and the starting and ending point sets for the next tile layer are calculated based on the path results. Then, path planning is performed on the next tile layer, using a divide-and-conquer approach to accelerate the sub-path planning task on each tile. This process continues until the path to the bottom tile layer is calculated, ultimately yielding the overall path planning result.

[0086] The distributed computing engine Spark is used to read the tile data blocks stored in HDFS layer by layer from top to bottom according to the pyramid hierarchy. Spark tile RDDs are constructed for each layer of the pyramid, and then distributed batch processing is performed on the tile RDDs. During the batch processing, path planning is performed for each tile and path vertex RDDs are generated. The above operations are iterated in Spark until the bottom-level path vertex RDD is obtained. Finally, the path vertex RDDs are checked to see if they are passable. For vertices that are not passable in the path, a global lunar grid DEM path search is performed again to ensure the accessibility of the searched path.

[0087] In a master-slave cluster, Spark divides the path planning task into multiple subtasks and assigns them to slave nodes for execution. The slave nodes retrieve tile data from HDFS, execute the subtasks assigned by the master node, and finally write the planning results back to HDFS.

[0088] A schematic diagram of the iterative path planning process of a three-level pyramid is shown below. Figure 6 As shown in the diagram, the red dots represent the start and end points of the overall path for each layer of the tile pyramid, while the green dots represent the points mapped from the upper-layer path points to the boundaries of the lower-layer tiles, i.e., the start and end points of the sub-path planning tasks on each tile within each layer. First, based on the start and end point coordinates of the top layer (Layer-2), a coarse path for the top layer of the pyramid is calculated. Then, the path is mapped to the next layer (Layer-1) by multiplying the path point coordinates by 2 and identifying points whose coordinates fall on the boundaries of the lower-layer tiles. Based on the path point coordinates on the boundaries, tiles containing these coordinates are filtered out. The path result is divided into multiple segments, and the sub-path result is calculated on each tile. This process is executed in parallel using Spark. The above operations are repeated until the path result on the bottom layer (Layer-0) is obtained, which is the final result.

[0089] Step 3: Detailed Periodic Planning

[0090] The resolution of full-moon DEM imagery is relatively low (20 meters / pixel). While it can provide terrain information over a large area, it lacks sufficient detail in local areas (especially regions with a high density of small impact craters). This means that global path planning based on full-moon DEM imagery may fail to identify and avoid some small obstacles, increasing the rovers' operational risks. To address this issue, this invention introduces high-resolution DOM imagery (7 meters / pixel) to provide more refined terrain information. The refined periodic planning proposed in this invention mainly involves the following steps:

[0091] (1) Node simplification of global planning path: The path nodes are simplified based on the Bresenham algorithm, and the simplified path nodes are used as the input for subsequent fine planning.

[0092] Because the 8-direction A* path planning algorithm produces a path with many inflection points and large turning angles, it is not smooth enough. This invention optimizes the above path based on the Bresenham algorithm, removes redundant path nodes, reduces the overall path turning angle, and finally obtains a set of optimized path nodes with no obstacles between each pair, which can be used for more refined periodic planning.

[0093] The core idea of ​​the Bresenham algorithm is to use the slope of a straight line to progressively draw pixels, while simultaneously selecting the discrete pixel closest to the target line by incrementally calculating decision variables. During the drawing process, the algorithm first determines the principal direction of the line (i.e., the x-axis when the absolute value of the slope is less than 1, and the y-axis when it is greater than 1), and then progresses incrementally along this principal direction. Next, by calculating the vertical distance error from the current pixel to the line, the decision variables are updated to determine the position of the next pixel.

[0094] The core process of path optimization based on the Bresenham algorithm is as follows:

[0095] Starting from the origin, each path node in the path is traversed sequentially. It is determined whether the line connecting the current origin and the traversed node passes through or is close to an obstacle point; that is, whether there is a dangerous node on that line. If not, the traversal continues. If a dangerous node exists, the previous path node of the traversed node is added to the key point list, and that previous path node is set as the starting point for the next round of traversal. This process is repeated until the destination is reached, ultimately generating a simplified path. Because only path nodes that do not contain dangerous nodes are added to the key point list, the simplified path maintains the same safety as the original path.

[0096] A schematic diagram illustrating the path simplification based on the Bresenham algorithm is shown below. Figure 7 As shown in the diagram, the black area represents the obstacle area, the dashed line represents the original path, and the solid line represents the simplified path. Taking this diagram as an example, starting from the starting point n0, the Bresenham algorithm is used to determine if there is a dangerous node on the line connecting n0 and n1. If no dangerous node is found, the traversal continues. When traversing to point n3, it is found that the line connecting n0 and n3 passes through an obstacle point. Therefore, the previous node n2 of n3 is added to the key point list, and n2 is set as the starting point for the next round of traversal. Then, starting from point n2, the traversal proceeds sequentially. Similarly, when traversing to point n5, it is found that there is an obstacle point on the line connecting n2 and n5. Therefore, the previous node n4 of n5 is added to the key point list. The above process is repeated until the end point n is reached. k The path simplification process is now complete. The key point list now contains n0, n2, n4, and n. kThese points constitute the simplified path. Compared to the previous path, the simplified path shortens the path length and reduces the turning angle.

[0097] (2) Determining the scope of the periodic path planning: Taking each two adjacent nodes in the simplified path as a period, determine the latitude and longitude range of each path segment, and extract the high-resolution DOM image of the corresponding area by calculating the geographic coordinates between nodes.

[0098] (3) Small crater detection: The CenterNet deep learning-based network model was used to detect small craters in the extracted high-resolution DOM images. CenterNet is an efficient target detection algorithm that can accurately identify crater targets in the images.

[0099] CenterNet is a deep learning-based object detection algorithm, representing one of the anchor-free methods. Its core idea is to simplify object detection into a keypoint detection problem. Traditional object detection methods typically rely on predefined candidate boxes (anchors), predicting the object's position and size through regression analysis on these boxes. However, this method is not only computationally complex but also easily affected by parameter settings and box placement. In contrast, CenterNet abandons the anchor mechanism, directly detecting the center point of each object in the image and inferring the object's size, pose, and other attributes based on this. This significantly improves both speed and accuracy while maintaining detection speed.

[0100] This invention uses CenterNet to detect meteorite craters. By superimposing features from different levels, it generates more representative and robust hierarchical features, improving the estimation ability of crater centers, especially for small craters. The smallest crater detected using this method is 500m. Compared with the Robbins crater database, this method achieves a recall of 73.66% and a precision of 78.27% for crater detection. The algorithm is publicly available at https: / / github.com / ShuoweiZhang / crater_detection.

[0101] (4) Safety map construction: The detected small meteorite craters are used as obstacles and superimposed on the safety map mentioned above. The safety map takes into account factors such as terrain slope and meteorite crater distribution to provide support for local path planning. The OC-WHT-A* algorithm mentioned above is used again to generate local paths that avoid meteorite craters, so as to achieve fine periodic planning.

[0102] Figure 8The image shows a comparison of path planning results before and after adding crater obstacles, with red circles representing detected crater obstacles. It can be seen that most small craters were successfully identified. The blue path represents the original path, i.e., the path planning result without adding crater obstacles, while the green path represents the result of periodic path planning after adding crater obstacles. It can be seen that the original path still passes through obvious obstacle areas (such as craters) on high-resolution DOM imagery. However, by using CenterNet to detect small craters and adding them to the safety map, the path can avoid obstacles that cannot be identified by DEM imagery alone, improving the safety of path planning.

[0103] To demonstrate the effectiveness of the technology proposed in this invention, we conducted the following experiments:

[0104] 1. Evaluation Indicators

[0105] This study uses time cost to measure the performance of path planning algorithms. Time cost is defined as the total running time of path planning. The time cost of distributed algorithms includes cluster startup time and parallel computing time. The calculation formula is as follows:

[0106] T = T start +T compute #(9)

[0107] Furthermore, the quality of the path results is evaluated using path length, the number of hazardous nodes on the path, and the total turning angle. Path length refers to the actual distance traveled along the path from the starting point to the ending point. Hazardous nodes are defined as path nodes containing obstacles within a 40m radius. The total turning angle is obtained by calculating the angle between the direction vectors of every three consecutive points on the path and then summing them up.

[0108] 2. Experimental Environment

[0109] All experiments were conducted on three Linux virtual machines running CentOS 7, each with 6 cores and 16GB of RAM, built on a computer with an Intel(R) Core(TM) i9-10920X CPU @ 3.50GHz and 64GB RAM.

[0110] First, a 4000x4000 region was selected in the Oceanus Procellarum region of the Moon for comparative experiments under a single-machine environment. The improved algorithm was compared with the algorithm proposed by Hong et al., as well as the traditional A* algorithm and the OC-HT-A* algorithm (unweighted w) and OC-WHT-A* algorithm (weighted w) proposed in this invention. The results are shown in Table 2. Figure 9 As shown, in Figure 9In the above, (a) the path comparison of different A* algorithms for short distances, (b) the path comparison of different A* algorithms for medium distances, and (c) the path comparison of different A* algorithms for long distances.

[0111] Table 2 Comparison of short, medium, and long-distance path planning results of different A* algorithms in a single-machine environment.

[0112]

[0113] In short-distance scenarios, the traditional A* algorithm takes the longest time, at 81.348 seconds. OC-RA-A* and OC-HT-A* take almost the same time, differing by less than 0.1 seconds. This shows that OC-HT-A* reduces space complexity without increasing time complexity. The OC-WHT-A* algorithm is the fastest, improving upon OC-HT-A* by 6.93%. The time cost of the traditional A* algorithm is approximately 10.96 times that of OC-HT-A* and approximately 11.78 times that of OC-WHT-A*. Since OC-RA-A* and OC-HT-A* only improve the data structure, their cost function design is the same as the traditional A* algorithm. Therefore, the path planning results of the three algorithms are also the same, with the same number of dangerous nodes (70). However, OC-WHT-A* improves the cost function H(N) by introducing a weight w to avoid areas with dense obstacles, thus significantly reducing the number of dangerous nodes in the planned path to 0.

[0114] At medium distances, the traditional A* algorithm's time consumption increases significantly to 1418.512 seconds. OC-RA-A* and OC-HT-A* take similar times. The OC-WHT-A* algorithm is the fastest, showing a significant speed advantage over OC-HT-A*, improving speed by 48.74%. The time cost of the traditional A* algorithm is approximately 113.50 times that of OC-WHT-A*. It can be seen that as distance increases, the path search space expands, and the planning time of the traditional A* algorithm increases exponentially. OC-WHT-A* introduces a weight w, optimizing the path search space and making it more efficient for large-scale path planning. Furthermore, the path planned by OC-WHT-A* still ensures safety, with zero dangerous nodes, significantly outperforming the path plans obtained by the other three algorithms.

[0115] In long-distance scenarios, the traditional algorithm takes 10145.400 seconds, while OC-RA-A* and OC-HT-A* take similar times, with OC-HT-A* being slightly faster at 77.760 seconds. The OC-WHT-A* algorithm remains the fastest, taking only 36.160 seconds, a 53.49% improvement over OC-HT-A*. Furthermore, the path planned by the OC-WHT-A* algorithm maintains excellent safety, with zero dangerous nodes along its path. It can be seen that the traditional A* algorithm is difficult to apply to long-distance path planning, while the OC-WHT-A* algorithm maintains the fastest planning speed across short, medium, and long distances, and its performance advantage continues to increase with distance. While ensuring speed, the OC-WHT-A* algorithm also guarantees the safety of the path results, effectively avoiding obstacles in short, medium, and long-distance scenarios.

[0116] Furthermore, to verify the efficiency of the DPPS-STP distributed algorithm, this invention selected three regions—Ocean of Storms, the Chang'e 4 landing area, and the lunar south pole—for path planning experiments. The single-machine OC-WHT-A* algorithm, the DPPS-STP algorithm using OC-HT-A*, and the DPPS-STP algorithm using OC-WHT-A* were compared and tested. The results are shown in Tables 3 and 4. Figure 10-11 As shown, in Figure 11 In the figure, (a) shows the path comparison between the single-machine and distributed A* algorithms in the Storm Ocean region, (b) shows the path comparison between the single-machine and distributed A* algorithms in the CE-4 landing region, (c) shows the path comparison between the single-machine and distributed A* algorithms in the Antarctic region, (d) shows a magnified view of a local area of ​​figure (a), (e) shows a magnified view of a local area of ​​figure (b), and (f) shows a magnified view of a local area of ​​figure (c).

[0117] In terms of time efficiency, it can be seen that the distributed algorithm significantly shortens the path planning time. The DPPS-STP distributed algorithm using OC-HT-A* has a significantly shorter time than the single-machine OC-WHT-A* algorithm in all three regions, with a maximum reduction of 90.15%, demonstrating the efficiency advantage of the DPPS-STP distributed algorithm in path planning. Furthermore, in all three regions, the DPPS-STP distributed algorithm using OC-WHT-A* has a shorter time than the DPPS-STP distributed algorithm using OC-HT-A*, with an average time cost reduction of approximately 28.5%. This is because the DPPS-STP algorithm using OC-WHT-A* becomes faster in path planning calculations for a single tile due to the addition of weight w.

[0118] Regarding path length, the DPPS-STP distributed algorithm using OC-HT-A* yielded the shortest path lengths in all three regions. The path lengths planned by the single-machine OC-WHT-A* algorithm and the DPPS-STP distributed algorithm using OC-WHT-A* were approximately the same. Compared to the DPPS-STP distributed algorithm using OC-HT-A*, the OC-WHT-A* DPPS-STP distributed algorithm resulted in a path length increase of 30.749 km in the Oceanus Procellarum region, 23.96 km in the Chang'e 4 landing area, and 23.463 km in the Antarctic region. The maximum increase in path length across the three regions was less than 8%.

[0119] Regarding the number of dangerous nodes, while the DPPS-STP algorithm using OC-HT-A* plans a shorter path than the DPPS-STP algorithm using OC-WHT-A*, the number of dangerous nodes along its path is significantly increased. The DPPS-STP algorithm using OC-HT-A* resulted in 604 dangerous nodes in the Oceanus Procellarum region, and over 1000 dangerous nodes in both the Chang'e 4 landing area and the Antarctic region. In contrast, the DPPS-STP algorithm using OC-WHT-A* resulted in 0 dangerous nodes in both the Oceanus Procellarum region and the Chang'e 4 landing area, and only 40 dangerous nodes in the Antarctic region, indicating that the path effectively avoids areas surrounding obstacles, ensuring the safety of the path.

[0120] Table 3 Comparison of long-distance path planning time under single-machine and distributed methods in three study areas

[0121]

[0122] Table 4 Comparison of long-distance path planning time, path length, and number of dangerous nodes under single-machine and distributed methods in three study areas.

[0123]

[0124] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lunar long-distance path planning method integrating terrain, illumination, and abundance features, characterized in that: First, a security map is obtained by comprehensively analyzing DEM images, average illumination images, and rock abundance images using security evaluation rules. Then, the security map is constructed into a tile pyramid, divided into tiles by layer and stored in HDFS. Next, the Spark distributed computing engine is used to perform distributed computing on each layer of tiles from top to bottom to obtain a global planning path. Finally, combined with DOM images, a fine-grained periodic planning is performed on the global planning path. The A* algorithm is improved in terms of both cost function and data structure, and then the improved A* algorithm is used for path planning on each tile layer. In the improved A* algorithm, the data structure of the open table is implemented using a min-heap and a hash table, while the data structure of the close table is implemented using a hash table. Calculate the cost function of the improved A* algorithm using the following equation. , Where F(N) represents the total cost from the starting point to the ending point, G(N) represents the actual cost from the starting point to the current node N, H(N) represents the estimated cost from the current node N to the ending point, and w represents the static weights. Indicates the drivability of the current point N in the security map; The refined periodic planning includes the following steps: S1. The Bresenham algorithm is used to simplify the path nodes on the global planning path to obtain a simplified path; S2. For the simplified path, take every two adjacent path nodes as a cycle, determine the latitude and longitude range of each path segment, and extract the high-resolution DOM image and security map of the corresponding area. S3. The CenterNet network model based on deep learning is used to detect small craters in the extracted high-resolution DOM images. S4. The detected small meteorite craters are used as obstacles and superimposed on the extracted safety map. Then, the improved A* algorithm is used to re-plan the path to achieve periodic planning.

2. The lunar long-distance path planning method integrating terrain, illumination, and abundance features according to claim 1, characterized in that: When using the safety evaluation rules for comprehensive analysis, firstly, the mobility analysis of three images—DEM image, average illuminance image, and rock abundance image—is performed separately, generating four mobility maps. Then, the intersection of these four mobility maps is performed to obtain a single mobility map. Finally, this mobility map is convolved to obtain a safety map. The accessibility map has only two values, 0 and 1, where 0 represents that the point is an obstacle and cannot be passed. 1 indicates that this point is passable.

3. The lunar long-distance path planning method integrating terrain, illumination, and abundance features according to claim 2, characterized in that: Calculate the slope angle θ and roughness δ of each pixel in the DEM image, the average illuminance μ of each pixel in the average illuminance image, and the rock abundance φ of each pixel in the rock abundance image, and generate four accessibility maps according to the following conditions. When the slope angle θ is greater than the first threshold, it is considered impassable and its value is 0; when it is less than the first threshold, its value is 1, in order to generate a slope accessibility map. ; When the roughness δ is greater than the second threshold, it is considered impassable and its value is 0; when it is less than the second threshold, its value is 1, in order to generate a roughness accessibility map. ; When the average illuminance μ is less than the first threshold, it is considered impassable and its value is 0; when it is greater than the first threshold, its value is 1, in order to generate an illuminance accessibility map. ; When the rock abundance φ is greater than the first threshold, it is considered impassable and its value is 0; when it is less than the first threshold, its value is 1, in order to generate a rock abundance accessibility map. .

4. The lunar long-distance path planning method integrating terrain, illumination, and abundance features according to claim 3, characterized in that: The following formula is used to map slope accessibility. Roughness accessibility map Light accessibility map Rock abundance accessibility map Perform intersection operations to obtain a traversability map. ; The following formula is used to analyze the accessibility map. A security map is obtained by performing convolution calculations. , in, Map showing drivability pixels on , This represents the elements in the convolution kernel K. This indicates that the convolution kernel K is at pixel point The convolution result at the point.

5. The lunar long-distance path planning method integrating terrain, illumination, and abundance features according to claim 1, characterized in that: When using the Bresenham algorithm to simplify path nodes on a globally planned path, starting from the starting point, each path node in the path is traversed sequentially. It is determined whether the connection between the current starting point and the traversed node passes through or is close to an obstacle point. If not, the traversal continues; if it exists, the previous path node of the traversed node is added to the key point list, and the previous path node is set as the starting point of the next round of traversal. The above process is repeated until the destination is reached, thus achieving path simplification.

Images