A high-fidelity topological map construction method and system for mobile robots

Abstract

The application discloses a kind of high fidelity topological map construction method and system for mobile robot, method includes the following steps: S1, to the original occupancy grid map is executed grid map multistage smoothing preprocessing and distance field construction, to obtain the Euclidean distance field of free space;S2, based on Euclidean distance field, in combination with physical safety constraint, free space is filtered, and morphological refinement is executed to the available space after filtering, to extract single-pixel width Voronoi skeleton;S3, to Voronoi skeleton is executed topological skeleton pruning based on branch length, and all pixel points of skeleton after pruning are semantized using neighbor counting method, obtain the intermediate topological skeleton with graph theory semantic features, pixel point on intermediate topological skeleton is divided into end point, edge pixel and intersection;S4, to the intersection in intermediate topological skeleton is executed clustering and merging based on Euclidean distance, and to the network after clustering and merging is executed end depth filtering, obtain high fidelity topological map.

Description

Technical Field

[0001] This invention relates to the field of map building technology for robot navigation, and in particular to a method and system for building high-fidelity topology maps for mobile robots. Background Technology

[0002] Topological maps are lightweight map models that abstract complex physical environments into discrete connected graphs composed of nodes and edges. They express spatial geometric features with extremely low data dimensionality and are widely used in fields such as autonomous navigation, path planning, and environmental cognition for mobile robots. The construction of topological maps is usually based on occupancy grid maps. By extracting the environmental skeleton and semanticizing it into the form of nodes and edges, large-scale two-dimensional grid data is reduced to a simple topological representation.

[0003] In the prior art, patent CN111323037A discloses a novel Voronoi path planning algorithm for skeleton extraction of mobile robots. This algorithm first binarizes a three-color grid map, then performs erosion and dilation processing sequentially, and then extracts the skeleton from the processed binary map based on the eight-neighbor labeling method to obtain a Voronoi skeleton map. Finally, it searches for global paths on the skeleton map and smooths the paths using improved cubic spline interpolation. Although this scheme simplifies the map to some extent through erosion and dilation, it only uses erosion and dilation as auxiliary steps after binarization and does not design a dedicated multi-level morphological filtering mechanism for the jagged noise of walls unique to LiDAR mapping. At the same time, this scheme does not introduce a safe passage width constraint coupled with the actual physical size of the robot during the skeleton extraction process, nor does it perform aggregation processing for redundant intersection nodes caused by discrete grids on the skeleton. Therefore, this prior art is essentially a skeleton extraction scheme oriented towards path planning, rather than a scheme oriented towards global topology modeling, and its output skeleton may still have problems such as proximity to obstacles, intersection splitting, and redundant short branches.

[0004] The patent with publication number CN115143980A discloses an incremental topology map construction method based on pruned Voronoi graphs. This method maps the boundary points of drivable areas to a global grid map, constructs a traditional Voronoi graph with the boundary points as the base points, determines the incremental generation area through the maximum empty circle property, and removes Voronoi edges located in the occupied space by expanding the boundary points. Then, the pruned Voronoi graph is represented by an undirected graph and invalid short edges with a length less than the sum of the robot length and the length threshold are deleted. Finally, isolated edges are removed based on the breadth-first search algorithm to obtain a fully connected topology map. While this scheme achieves incremental construction and preliminary pruning of the topology map, its Voronoi diagram is generated analytically using obstacle point sets as base points. This heavily relies on the positioning accuracy of boundary points and lacks specific smoothing and suppression methods for wall jagged noise and isolated spikes generated by LiDAR scanning. The noise of the base points themselves is directly projected as noise in the Voronoi edges, which cannot be completely eliminated through post-pruning. Furthermore, this scheme only prunes short edges by adding a length threshold and a single connectivity relationship, and achieves secondary invalid edge removal by fitting short edges with consistent orientation. It does not set up a spatial distance-based clustering and merging mechanism for multiple adjacent key intersections generated by grid discretization in wide areas such as corridor intersections and open spaces. This makes it difficult to merge multiple nodes split from the same physical intersection in the topology map into a single central node, resulting in redundancy of topology nodes in intersection areas.

[0005] As can be seen from the above existing technologies, in terms of fully automatic extraction of a topological skeleton from a noisy original grid map that meets both robot physical access constraints and has high semantic consistency, there is no comprehensive solution in the industry that can simultaneously take into account noise suppression, path traversability, topological stability and network structure simplicity. Summary of the Invention

[0006] Purpose of the invention: In order to overcome the shortcomings of the existing technology, the present invention provides a high-fidelity topology map construction method and system for mobile robots, which solves the problems of topology distortion caused by mapping noise, lack of constraints on the actual passage capability of robots, generation of a large number of redundant topology nodes in wide areas, and difficulty in eliminating invalid short branches in the existing topology map construction methods.

[0007] Technical solution: To achieve the above objectives, the present invention provides a high-fidelity topology map construction method for mobile robots, the method comprising the following steps:

[0008] S1. Perform multi-level smoothing preprocessing and distance field construction on the original occupied grid map to obtain the Euclidean distance field of free space; the Euclidean distance field is numerically equivalent to the straight-line physical distance from the corresponding free space pixel to the nearest obstacle pixel.

[0009] S2. Based on the Euclidean distance field, the free space is filtered in combination with physical security constraints, and morphological thinning is performed on the filtered available space to extract a single-pixel wide Voronoi skeleton.

[0010] S3. Perform topological skeleton pruning based on branch length on the Voronoi skeleton, and use the neighbor counting method to perform semantic processing on all pixels of the pruned skeleton to obtain an intermediate topological skeleton with graph theory semantic features. The pixels on the intermediate topological skeleton are distinguished as endpoints, edge pixels and intersections.

[0011] S4. Perform Euclidean distance-based clustering and merging on the intersections in the intermediate topology skeleton, and perform end-point depth filtering on the clustered and merged network to obtain a high-fidelity topology map.

[0012] Further, step S1 specifically includes:

[0013] S11. Perform binarization mapping on the original occupied grid map, set the pixels representing free space to passable state, and merge the obstacle pixels and the undetected unknown area pixels into an impassable state to obtain a binarized map.

[0014] S12. Perform morphological opening and morphological closing operations sequentially on the binarized map to fill in depressions on the wall surface smaller than a specific value and smooth out sharp protrusions to obtain a smooth map; the structuring elements used in the morphological opening and morphological closing operations can be square, circular or cross-shaped structuring elements.

[0015] S13. Perform map edge clipping on the smoothed map, clipping away the pixel area of ​​a preset width of the outermost edge of the map to obtain an effective free space map; the preset width is preferably 5 pixels;

[0016] S14. Perform Euclidean distance transformation on the effective free space graph, calculate the straight-line physical distance from each free space pixel to the nearest obstacle pixel, and obtain the Euclidean distance field.

[0017] Further, step S2 specifically includes:

[0018] S21. Based on the Euclidean distance field, set a physical safety threshold, traverse all pixels in the Euclidean distance field, retain pixels with distance values ​​greater than or equal to the physical safety threshold, and remove narrow regions with distance values ​​less than the physical safety threshold to obtain a safe free space map; the physical safety threshold can be a fixed value, such as being determined based on the robot's minimum safe passage width.

[0019] S22. Perform a morphological thinning algorithm on the safe free space graph, iteratively peeling away the outer boundary pixels of the safe free space graph layer by layer until only a connected skeleton with a width of 1 pixel remains in the entire space. The connected skeleton is the Voronoi skeleton. In this step, the morphological thinning algorithm can be the Zhang-Suen thinning algorithm or the Guo-Hall thinning algorithm.

[0020] Furthermore, step S3 specifically includes:

[0021] S31. Calculate the branch length of the Voronoi skeleton. The pixel length from the end of each branch to its main connection point is taken as the branch length. If the branch length is less than a preset length threshold, the branch is identified as a noise branch caused by environmental distortion and is cut off to obtain a clean skeleton.

[0022] S32. For each skeleton pixel on the cleaned skeleton, scan the number of skeleton pixels in its eight neighboring regions, and semantically label the skeleton pixels according to the following rules: when the number of skeleton pixels is 1, mark the skeleton pixel as an endpoint; when the number of skeleton pixels is 2, mark the skeleton pixel as an edge pixel; when the number of skeleton pixels is greater than or equal to 3, mark the skeleton pixel as an intersection point; the marked endpoints, edge pixels, and intersection points together constitute the intermediate topological skeleton. In this step, the eight neighboring regions refer to the eight pixels remaining after removing the center pixel in a 3×3 neighborhood centered on the current skeleton pixel.

[0023] Furthermore, step S4 specifically includes:

[0024] S41. Obtain all intersections in the intermediate topological skeleton, and search for multiple intersections whose Euclidean distance is less than or equal to the preset spatial clustering threshold as the merging condition. Group the multiple intersections found into the same cluster, and calculate the average value of the coordinates of all intersections in the cluster as the coordinates of the synthesized intersection. Replace all intersections in the original cluster with the synthesized intersection to obtain the aggregated skeleton.

[0025] S42. Traverse all endpoints in the aggregated skeleton, calculate the pixel distance between each endpoint and its nearest synthetic intersection, and if the pixel distance is less than a preset end filtering threshold, then determine the endpoint and its connected branches as redundant stubs and remove them to obtain the topology map.

[0026] The preset spatial clustering threshold can be a fixed value, such as 15 pixels. Preferably, the preset spatial clustering threshold is calculated using a weighted average strategy based on the distance field. The specific implementation of the weighted average strategy is as follows: extract the absolute distance value of each candidate intersection point in the Euclidean distance field as its weight coefficient. The larger the absolute distance value, the farther it is from the surrounding obstacles, and the larger the corresponding weight coefficient. Divide the calculated weighted coordinate sum by the weight sum to obtain the new coordinates of the center intersection point. This method uses the free space width as a weighting coefficient for weighted calculation, so that the synthesized center intersection point can adaptively deviate to a more open area of ​​physical space, reducing the physical risk of collision when the mobile robot turns or merges at intersections.

[0027] The preset end-filter threshold can be a fixed value, such as 20 pixels.

[0028] Furthermore, the physical security threshold is determined adaptively as follows:

[0029] Obtain the outer envelope radius r of the mobile robot, the map resolution σ, and the safety margin coefficient k according to the formula. The physical security threshold, calculated in pixels, is obtained. ; and in step S21, using the physical security threshold The Euclidean distance field is filtered.

[0030] The safety margin coefficient k is preferably in the range of 1.1 to 1.5, and is used to compensate for robot positioning errors and dynamic disturbances; for different types of robot platforms, the outer envelope radius r can be obtained by measuring the robot specification parameters.

[0031] Furthermore, the preset spatial clustering threshold in step S41 is determined in the following adaptive manner:

[0032] Obtain the average distance field value within a preset neighborhood range around each intersection point in the intermediate topological skeleton. With the average distance field value The product of the product with the preset scaling factor λ is used as the preset spatial clustering threshold, that is, the preset spatial clustering threshold is equal to... .

[0033] The preset proportional coefficient λ is preferably in the range of 1.0 to 2.0, so that the preset spatial clustering threshold takes a smaller value in the narrower area of ​​the corridor and a larger value in the open and wide area.

[0034] Furthermore, the method also includes a topology node semantic enhancement step S5 performed after step S4:

[0035] S51. Obtain the semantic perception results output by the sensing device mounted on the mobile robot. The semantic perception results include, but are not limited to, at least one of room category labels, doorplate labels, and fixed facility categories.

[0036] S52. For each synthetic intersection and endpoint in the topology map, query the semantic label that matches its spatial location in the semantic perception result according to its coordinates, and bind the queried semantic label as an additional attribute to the corresponding synthetic intersection or endpoint to obtain a semantically enhanced topology map.

[0037] The sensing device may include one or more of a visual camera, a depth camera, a QR code recognition module, or an RFID reading module; the matching of the semantic tags may be based on nearest neighbor search or a preset matching radius.

[0038] A high-fidelity topology map construction system for mobile robots includes:

[0039] The preprocessing and distance field construction module is used to perform multi-level smoothing preprocessing and distance field construction on the original occupied grid map to obtain the Euclidean distance field of free space; the Euclidean distance field is numerically equivalent to the straight-line physical distance from the corresponding free space pixel to the nearest obstacle pixel.

[0040] The skeleton extraction module is used to filter the free space based on the Euclidean distance field and combined with physical safety constraints, and to perform morphological thinning on the filtered available space to extract a single-pixel wide Voronoi skeleton.

[0041] The pruning and semantic processing module is used to perform topological skeleton pruning based on branch length on the Voronoi skeleton, and to perform semantic processing on all pixels of the pruned skeleton using the neighbor counting method to obtain an intermediate topological skeleton with graph theory semantic features. The pixels on the intermediate topological skeleton are distinguished as endpoints, edge pixels and intersections.

[0042] The clustering and merging module is used to perform Euclidean distance-based clustering and merging on the intersection points in the intermediate topology skeleton, and to perform end-point depth filtering on the clustered and merged network to obtain a high-fidelity topology map.

[0043] Beneficial Effects: The high-fidelity topology map construction method and system for mobile robots of the present invention have the following beneficial effects:

[0044] (1) Improving the structural stability and environmental representation accuracy of the topological network: By introducing multi-stage morphological processing before skeleton extraction and combining it with the aggregation optimization method of key nodes after skeleton generation, this invention effectively reduces problems such as skeleton spikes, abnormal bifurcations, and duplicate nodes caused by mapping noise in traditional methods. After the above processing, the generated topological skeleton is more continuous and concise in its overall structure, and can stably reflect the spatial relationships of the main corridors and intersection areas in the environment, avoiding local disconnections and structural distortions, thereby improving the accuracy of the macroscopic topological network in representing the real environment.

[0045] (2) Introducing explicit safe passage constraints in the topology modeling stage: This invention takes into account the robot's actual passage requirements in the environmental modeling stage. Through a safe space filtering mechanism based on distance transformation, the topological skeleton is distributed only within areas that meet the safe passage width requirements. The resulting topological path can naturally be located in the middle area of ​​the corridor or passage, maintaining a reasonable distance from obstacles, effectively reducing the probability of the robot driving close to walls or obstacles during subsequent navigation, thereby improving the safety and stability when performing navigation tasks based on this topological map.

[0046] (3) Reduce the computational complexity of global path planning and improve engineering applicability: This invention transforms the original two-dimensional occupancy map, which was originally composed of a large number of grid cells, into a graph structure composed of a small number of topological nodes with clear semantics and their connection relationships through topological skeleton extraction and node aggregation. This topological representation significantly reduces the number of nodes that need to participate in the search during the planning process, enabling the global path planning algorithm to complete path calculation in a smaller search space. This is beneficial to improving planning efficiency, reducing computational resource consumption, and enhancing the system's deployment applicability on actual robot platforms. Attached Figure Description

[0047] Figure 1 A flowchart illustrating a method for constructing high-fidelity topology maps for mobile robots;

[0048] Figure 2 A complete flowchart of a method for constructing high-fidelity topology maps for mobile robots;

[0049] Figure 3 A schematic diagram of a high-fidelity topology map construction system for mobile robots. Detailed Implementation

[0050] The invention will now be further described with reference to the accompanying drawings.

[0051] like Figure 1 The high-fidelity topology map construction method for mobile robots shown includes the following steps:

[0052] S1. Perform multi-level smoothing preprocessing and distance field construction on the original occupied grid map to obtain the Euclidean distance field of free space; the Euclidean distance field is numerically equivalent to the straight-line physical distance from the corresponding free space pixel to the nearest obstacle pixel.

[0053] S2. Based on the Euclidean distance field, the free space is filtered in combination with physical security constraints, and morphological thinning is performed on the filtered available space to extract a single-pixel wide Voronoi skeleton.

[0054] S3. Perform topological skeleton pruning based on branch length on the Voronoi skeleton, and use the neighbor counting method to perform semantic processing on all pixels of the pruned skeleton to obtain an intermediate topological skeleton with graph theory semantic features. The pixels on the intermediate topological skeleton are distinguished as endpoints, edge pixels and intersections.

[0055] S4. Perform Euclidean distance-based clustering and merging on the intersections in the intermediate topology skeleton, and perform end-point depth filtering on the clustered and merged network to obtain a high-fidelity topology map.

[0056] Compared to existing Voronoi-based topology construction methods, the proposed topology map construction method achieves fully automated end-to-end construction from noisy original grid maps to high-fidelity topology maps through multi-level preprocessing of the original occupied grid map, secure filtering of the Euclidean distance field, semanticization of the Voronoi skeleton, and cascaded clustering and merging of intersection points. This method abandons the rigid geometric splicing approach, simultaneously addressing four key engineering requirements within the same cascaded pipeline: noise suppression, path traversability, topological stability, and network structural simplicity. This results in a continuous and concise topology map that stably reflects the spatial relationships of major corridors and intersection areas in the environment, avoiding local disconnections and structural distortions, thereby improving the accuracy of the macroscopic topology network's representation of the real environment.

[0057] Because the topology map is reduced from a large number of grid cells to a small number of topology nodes with clear semantics and their connections, the subsequent global path planning algorithm can complete path calculation in a smaller search space, reducing computational resource consumption and enhancing the system's applicability on actual robot platforms. In addition, the output structure of the method is highly compatible with the underlying mapping and preprocessing frameworks of mainstream robot navigation systems such as Nav2 and ROS. It does not require large-scale modification of the core architecture of existing navigation systems and can be integrated and deployed simply through interface calls and parameter configuration.

[0058] Preferably, step S1 specifically includes:

[0059] S11. Perform binarization mapping on the original occupied grid map, set the pixels representing free space to passable state, and merge the obstacle pixels and the undetected unknown area pixels into an impassable state to obtain a binarized map.

[0060] S12. Perform morphological opening and morphological closing operations sequentially on the binarized map to fill in depressions on the wall surface smaller than a specific value and smooth out sharp protrusions to obtain a smooth map; the structuring elements used in the morphological opening and morphological closing operations can be square, circular or cross-shaped structuring elements.

[0061] S13. Perform map edge clipping on the smoothed map, clipping away the pixel area of ​​a preset width of the outermost edge of the map to obtain an effective free space map; the preset width is preferably 5 pixels;

[0062] S14. Perform Euclidean distance transformation on the effective free space graph, calculate the straight-line physical distance from each free space pixel to the nearest obstacle pixel, and obtain the Euclidean distance field.

[0063] When performing morphological opening and closing operations, the preset kernel size r can be a fixed value. However, preferably, the preset kernel size r is dynamically allocated based on the physical resolution of the original occupied raster map and the local environmental feature complexity of the current processing area. The method for evaluating the local environmental feature complexity is as follows: a sliding window is used to statistically analyze the rate of change of the curvature of obstacle boundaries within the current processing area; for large-scale open areas with a small rate of curvature change, a larger kernel size r is allocated for smoothing to completely eliminate discrete noise; for narrow, confined spaces with a large rate of curvature change or dense obstacles, a smaller kernel size r is allocated to protect subtle environmental geometry from being erased. By adaptively adjusting the preset kernel size r, while filtering out jagged noise at the edges of open areas, the loss of minor passable features caused by over-smoothing of dense obstacle areas is effectively avoided, thus improving the accuracy of environmental topology representation.

[0064] Through steps S11-S14, the smooth map and effective free space map are obtained by performing opening and closing operations and edge clipping on the binarized map. Then, the Euclidean distance field is obtained by Euclidean distance transformation. This suppresses raster jagged noise at the source, making the free space contour more continuous and regular. It fundamentally avoids the participation of noisy structures in the subsequent skeleton extraction process, effectively suppressing abnormal bifurcations and redundant paths caused by uneven boundaries in traditional skeleton extraction methods.

[0065] Preferably, step S2 specifically includes:

[0066] S21. Based on the Euclidean distance field, set a physical safety threshold, traverse all pixels in the Euclidean distance field, retain pixels with distance values ​​greater than or equal to the physical safety threshold, and remove narrow regions with distance values ​​less than the physical safety threshold to obtain a safe free space map; the physical safety threshold can be a fixed value, such as being determined based on the robot's minimum safe passage width.

[0067] S22. Perform a morphological thinning algorithm on the safe free space graph, iteratively peeling away the outer boundary pixels of the safe free space graph layer by layer until only a connected skeleton with a width of 1 pixel remains in the entire space. The connected skeleton is the Voronoi skeleton. In this step, the morphological thinning algorithm can be the Zhang-Suen thinning algorithm or the Guo-Hall thinning algorithm.

[0068] The morphological thinning algorithm consists of two sub-iteration processes: In the first sub-iteration, a first set of preset 3×3 neighborhood connectivity is used to determine the template, and pixels on the upper, right, and lower right boundaries that satisfy the external contour conditions are marked and deleted in parallel; In the second sub-iteration, a second set of determination templates is used to mark and delete pixels on the lower, left, and upper left boundaries that satisfy the conditions in parallel; The two sub-iterations are executed alternately until no pixels are deleted in the image, ensuring that the thinning process always progresses uniformly inward from multiple boundary directions; By adopting an alternate-direction parallel thinning iteration strategy, the convergence speed of the thinning calculation is accelerated, while ensuring that the extracted Voronoi skeleton has strict topological invariance and high-precision spatial centering characteristics.

[0069] Through steps S21-S22, the safe free space map is obtained by comparing the Euclidean distance field with the physical safety threshold, and then the Voronoi skeleton is obtained through morphological refinement. The actual passage requirements of the robot are taken into account in the environmental modeling stage. By using the geometric principle that the distance transformation value is equivalent to half the width of the environmental corridor, narrow dead angles and gaps that cannot meet the robot's safe passage requirements are completely eliminated, so that the Voronoi skeleton is only distributed in the area that meets the safe passage width requirements. This solves the engineering defect of the existing technology where the topological path grid is connected but cannot actually be passed through.

[0070] Preferably, step S3 specifically includes:

[0071] S31. Calculate the branch length of the Voronoi skeleton. The branch length is the pixel length from the end of each branch to its main connection point. If the branch length is less than a preset length threshold, the branch is identified as a noise branch caused by environmental distortion and is pruned and stripped to obtain a clean skeleton. In this embodiment, the preset length threshold is preferably 20 pixels.

[0072] S32. For each skeleton pixel on the cleaned skeleton, scan the number of skeleton pixels in its eight neighboring regions, and semantically label the skeleton pixels according to the following rules: when the number of skeleton pixels is 1, mark the skeleton pixel as an endpoint; when the number of skeleton pixels is 2, mark the skeleton pixel as an edge pixel; when the number of skeleton pixels is greater than or equal to 3, mark the skeleton pixel as an intersection point; the marked endpoints, edge pixels, and intersection points together constitute the intermediate topological skeleton. In this step, the eight neighboring regions refer to the eight pixels remaining after removing the center pixel in a 3×3 neighborhood centered on the current skeleton pixel.

[0073] In steps S31-S32 above, redundant short branches in the Voronoi skeleton are pruned using the preset length threshold to obtain the purified skeleton. Pixels are then classified into endpoints, edge pixels, and intersections using a neighbor counting method. This process removes redundant short branches that have no practical value in navigation but increase the size and complexity of the topology network, thus purifying the topology network, reducing the search burden of the global path planning algorithm, and avoiding local deadlock problems caused by redundant branches.

[0074] Preferably, step S4 specifically includes:

[0075] S41. Obtain all intersections in the intermediate topological skeleton, and search for multiple intersections whose Euclidean distance is less than or equal to the preset spatial clustering threshold as the merging condition. Group the multiple intersections found into the same cluster, and calculate the average value of the coordinates of all intersections in the cluster as the coordinates of the synthesized intersection. Replace all intersections in the original cluster with the synthesized intersection to obtain the aggregated skeleton.

[0076] S42. Traverse all endpoints in the aggregated skeleton, calculate the pixel distance between each endpoint and its nearest synthetic intersection, and if the pixel distance is less than a preset end filtering threshold, then determine the endpoint and its connected branches as redundant stubs and remove them to obtain the topology map.

[0077] The preset spatial clustering threshold can be a fixed value, such as 15 pixels. Preferably, the preset spatial clustering threshold is calculated using a weighted average strategy based on the distance field. The specific implementation of the weighted average strategy is as follows: extract the absolute distance value of each candidate intersection point in the Euclidean distance field as its weight coefficient. The larger the absolute distance value, the farther it is from the surrounding obstacles, and the larger the corresponding weight coefficient. Divide the calculated weighted coordinate sum by the weight sum to obtain the new coordinates of the center intersection point. This method uses the free space width as a weighting coefficient for weighted calculation, so that the synthesized center intersection point can adaptively deviate to a more open area of ​​physical space, reducing the physical risk of collision when the mobile robot turns or merges at intersections.

[0078] The preset end-filter threshold can be a fixed value, such as 20 pixels.

[0079] In steps S41-S42 above, multiple neighboring intersections are combined into a unique composite intersection by using the preset spatial clustering threshold to obtain the aggregated skeleton. Then, the topology map is obtained by end-point depth filtering, which eliminates discrete grid redundancy and facilitates the efficient implementation of subsequent path planning and decision-making processes.

[0080] Preferably, the physical security threshold is determined in an adaptive manner as follows:

[0081] Obtain the outer envelope radius r of the mobile robot, the map resolution σ, and the safety margin coefficient k according to the formula. The physical security threshold, calculated in pixels, is obtained. ; and in step S21, using the physical security threshold The Euclidean distance field is filtered.

[0082] The safety margin coefficient k is preferably in the range of 1.1 to 1.5, and is used to compensate for robot positioning errors and dynamic disturbances; for different types of robot platforms, the outer envelope radius r can be obtained by measuring the robot specification parameters.

[0083] Using the above-mentioned adaptive physical safety threshold determination method, the filtering result of the Euclidean distance field is adaptively adjusted according to the changes in the robot platform and the resolution of the original occupied grid map by combining the outer envelope radius r, the map resolution σ, and the safety margin coefficient k.

[0084] Preferably, the preset spatial clustering threshold in step S41 is determined in the following adaptive manner:

[0085] Obtain the average distance field value within a preset neighborhood range around each intersection point in the intermediate topological skeleton. With the average distance field value The product of the product with the preset scaling factor λ is used as the preset spatial clustering threshold, that is, the preset spatial clustering threshold is equal to... .

[0086] The preset proportional coefficient λ is preferably in the range of 1.0 to 2.0, so that the preset spatial clustering threshold takes a smaller value in the narrower area of ​​the corridor and a larger value in the open and wide area.

[0087] An adaptive spatial clustering threshold is used, based on the average distance field value. The product of the pre-set proportional coefficient λ ensures that the generation of the synthetic intersection reflects the physical attributes of a single intersection under physical intersections of different widths.

[0088] Preferably, the method further includes a topology node semantic enhancement step S5 performed after step S4:

[0089] S51. Obtain the semantic perception results output by the sensing device mounted on the mobile robot. The semantic perception results include, but are not limited to, at least one of room category labels, doorplate labels, and fixed facility categories.

[0090] S52. For each synthetic intersection and endpoint in the topology map, query the semantic label that matches its spatial location in the semantic perception result according to its coordinates, and bind the queried semantic label as an additional attribute to the corresponding synthetic intersection or endpoint to obtain a semantically enhanced topology map.

[0091] The sensing device may include one or more of a visual camera, a depth camera, a QR code recognition module, or an RFID reading module; the matching of the semantic tags may be based on nearest neighbor search or a preset matching radius.

[0092] The topology node semantic enhancement process S51-S52 is adopted. The semantically enhanced topology map is obtained by spatial matching of the semantic perception result and the topology map, which provides node support with semantic attributes for upper-level task planning.

[0093] like Figure 2 This is a flowchart of a complete high-fidelity topology map construction method for mobile robots in one embodiment.

[0094] This invention also provides a high-fidelity topology map construction system for mobile robots. The high-fidelity topology map construction system may include or be divided into one or more program modules. One or more program modules are stored in a storage medium and executed by one or more processors to complete this invention and implement the aforementioned high-fidelity topology map construction method. The program module referred to in this embodiment of the invention refers to a series of computer program instruction segments capable of performing a specific function, which is more suitable than the program itself for describing the execution process of the high-fidelity topology map construction method in the storage medium. The following description will specifically introduce the functions of each program module in this embodiment, such as... Figure 3 As shown, the high-fidelity topology map construction system includes:

[0095] The preprocessing and distance field construction module is used to perform multi-level smoothing preprocessing and distance field construction on the original occupied grid map to obtain the Euclidean distance field of free space; the Euclidean distance field is numerically equivalent to the straight-line physical distance from the corresponding free space pixel to the nearest obstacle pixel.

[0096] The skeleton extraction module is used to filter the free space based on the Euclidean distance field and combined with physical safety constraints, and to perform morphological thinning on the filtered available space to extract a single-pixel wide Voronoi skeleton.

[0097] The pruning and semantic processing module is used to perform topological skeleton pruning based on branch length on the Voronoi skeleton, and to perform semantic processing on all pixels of the pruned skeleton using the neighbor counting method to obtain an intermediate topological skeleton with graph theory semantic features. The pixels on the intermediate topological skeleton are distinguished as endpoints, edge pixels and intersections.

[0098] The clustering and merging module is used to perform Euclidean distance-based clustering and merging on the intersection points in the intermediate topology skeleton, and to perform end-point depth filtering on the clustered and merged network to obtain a high-fidelity topology map.

[0099] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing high-fidelity topology maps for mobile robots, characterized in that, The method includes the following steps: S1. Perform multi-level smoothing preprocessing and distance field construction on the original occupied raster map to obtain the Euclidean distance field in free space; S2. Based on the Euclidean distance field, the free space is filtered in combination with physical security constraints, and morphological thinning is performed on the filtered available space to extract a single-pixel wide Voronoi skeleton. S3. Perform topological skeleton pruning based on branch length on the Voronoi skeleton, and use the neighbor counting method to perform semantic processing on all pixels of the pruned skeleton to obtain an intermediate topological skeleton with graph theory semantic features. The pixels on the intermediate topological skeleton are distinguished as endpoints, edge pixels and intersections. S4. Perform Euclidean distance-based clustering and merging on the intersections in the intermediate topology skeleton, and perform end-point depth filtering on the clustered and merged network to obtain a high-fidelity topology map.

2. The high-fidelity topology map construction method for mobile robots according to claim 1, characterized in that, Step S1 specifically includes: S11. Perform binarization mapping on the original occupied grid map, set the pixels representing free space to passable state, and merge the obstacle pixels and the undetected unknown area pixels into an impassable state to obtain a binarized map. S12. Perform morphological opening and morphological closing operations on the binarized map in sequence to fill in the depressions on the wall surface that are smaller than a specific value and smooth out sharp protrusions to obtain a smooth map. S13. Perform map edge clipping on the smooth map, clipping away the pixel area of ​​the outermost edge of the map with a preset width to obtain an effective free space map; S14. Perform Euclidean distance transformation on the effective free space graph, calculate the straight-line physical distance from each free space pixel to the nearest obstacle pixel, and obtain the Euclidean distance field.

3. The high-fidelity topology map construction method for mobile robots according to claim 1, characterized in that, Step S2 specifically includes: S21. Based on the Euclidean distance field, set a physical safety threshold, traverse all pixels in the Euclidean distance field, retain pixels with distance values ​​greater than or equal to the physical safety threshold, and remove narrow regions with distance values ​​less than the physical safety threshold to obtain a safe free space map. S22. Perform a morphological thinning algorithm on the safe free space graph, and iteratively peel off the outer boundary pixels of the safe free space graph layer by layer until only a 1-pixel-wide connected skeleton remains in the entire space. The connected skeleton is the Voronoi skeleton.

4. The method for constructing a high-fidelity topology map for mobile robots according to claim 1, characterized in that, Step S3 specifically includes: S31. Calculate the branch length of the Voronoi skeleton. The pixel length from the end of each branch to its main connection point is taken as the branch length. If the branch length is less than a preset length threshold, the branch is identified as a noise branch caused by environmental distortion and is cut off to obtain a clean skeleton. S32. For each skeleton pixel on the purification skeleton, scan the number of skeleton pixels in its eight neighboring regions, and semantically label the skeleton pixels according to the following rules: when the number of skeleton pixels is 1, mark the skeleton pixel as an endpoint; when the number of skeleton pixels is 2, mark the skeleton pixel as an edge pixel; when the number of skeleton pixels is greater than or equal to 3, mark the skeleton pixel as an intersection point; the marked endpoints, edge pixels, and intersection points together constitute the intermediate topology skeleton.

5. The method for constructing a high-fidelity topology map for mobile robots according to claim 1, characterized in that, Step S4 specifically includes: S41. Obtain all intersections in the intermediate topological skeleton, and search for multiple intersections whose Euclidean distance is less than or equal to the preset spatial clustering threshold as the merging condition. Group the multiple intersections found into the same cluster, and calculate the average value of the coordinates of all intersections in the cluster as the coordinates of the synthesized intersection. Replace all intersections in the original cluster with the synthesized intersection to obtain the aggregated skeleton. S42. Traverse all endpoints in the aggregated skeleton, calculate the pixel distance between each endpoint and its nearest synthetic intersection, and if the pixel distance is less than a preset end filtering threshold, then determine the endpoint and its connected branches as redundant stubs and remove them to obtain the topology map.

6. The high-fidelity topology map construction method for mobile robots according to claim 3, characterized in that, The physical security threshold is determined adaptively as follows: Obtain the outer envelope radius r of the mobile robot, the map resolution σ, and the safety margin coefficient k according to the formula. The physical security threshold, calculated in pixels, is obtained. ; and in step S21, using the physical security threshold The Euclidean distance field is filtered.

7. The method for constructing a high-fidelity topology map for mobile robots according to claim 5, characterized in that, The preset spatial clustering threshold in step S41 is determined in the following adaptive manner: Obtain the average distance field value within a preset neighborhood range around each intersection point in the intermediate topological skeleton. With the average distance field value The product of the product with the preset scaling factor λ is used as the preset spatial clustering threshold, that is, the preset spatial clustering threshold is equal to... .

8. The method for constructing a high-fidelity topology map for mobile robots according to claim 1, characterized in that, The method further includes a topology node semantic enhancement step S5 performed after step S4: S51. Obtain the semantic perception results output by the sensing device mounted on the mobile robot. The semantic perception results include, but are not limited to, at least one of room category labels, doorplate labels, and fixed facility categories. S52. For each synthetic intersection and endpoint in the topology map, query the semantic label that matches its spatial location in the semantic perception result according to its coordinates, and bind the queried semantic label as an additional attribute to the corresponding synthetic intersection or endpoint to obtain a semantically enhanced topology map.

9. A high-fidelity topology map construction system for mobile robots, characterized in that, include: The preprocessing and distance field construction module is used to perform multi-level smoothing preprocessing and distance field construction on the original occupied raster map to obtain the Euclidean distance field in free space. The skeleton extraction module is used to filter the free space based on the Euclidean distance field and combined with physical safety constraints, and to perform morphological thinning on the filtered available space to extract a single-pixel wide Voronoi skeleton. The pruning and semantic processing module is used to perform topological skeleton pruning based on branch length on the Voronoi skeleton, and to perform semantic processing on all pixels of the pruned skeleton using the neighbor counting method to obtain an intermediate topological skeleton with graph theory semantic features. The pixels on the intermediate topological skeleton are distinguished as endpoints, edge pixels and intersections. The clustering and merging module is used to perform Euclidean distance-based clustering and merging on the intersection points in the intermediate topology skeleton, and to perform end-point depth filtering on the clustered and merged network to obtain a high-fidelity topology map.

Images