Robot addressing method based on three-dimensional point cloud semantic segmentation and dynamic topology reconstruction

Abstract

The application discloses a robot addressing method based on three-dimensional point cloud semantic segmentation and dynamic topology reconstruction, and particularly relates to the technical field of robot addressing, wherein elements such as walls, furniture and doors are extracted through online semantic segmentation, a continuous probability attribution field is constructed by taking the wall distribution and passable areas as constraints, and the space attribution is probabilized; a hierarchical topology graph is dynamically constructed, stable self-positioning is realized by utilizing attribution template probability matching and active exploration; a soft boundary transition strategy is introduced during path planning, address switching is implemented through lag confirmation in combination with door or passage crossing events during navigation, and the edge passable probability and the attribution field are updated in real time, so that the stable continuity of the address output of the robot in a dynamic indoor environment is realized.

Description

Technical Field

[0001] This invention relates to the field of robot addressing technology, and more specifically, to a robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction. Background Technology

[0002] In addressing methods based on lidar point cloud semantic segmentation and topology reconstruction, the continuity of physical space and the discrete determinism of addresses constitute a fundamental contradiction. LiDAR can only see geometric point clouds with smooth transitions; open office areas and corridor intersections lack clear wall boundaries, making it impossible for semantic segmentation to naturally delineate room boundaries. When subsequently relying on region segmentation to extract topology nodes, a slightly strict threshold will produce a large number of fragmented nodes, causing the address system to collapse; a slightly loose threshold will merge areas with different functions, failing to distinguish between lounges and interior corners near windows. The most direct impact occurs during localization. When the robot stops at the meeting room door, the point cloud simultaneously contains features of both the corridor and the interior. Hard attribution will cause address labels to jump back and forth between the corridor and the meeting room, disrupting the determinism of path planning and task handover. Therefore, the sharp contradiction between continuous spatial semantics and the requirements of discrete addressing forces the topology graph to accommodate probabilistic attribution or hierarchical addresses, which precisely conflicts with the deterministic assumptions upon which classical graph search relies. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a robot addressing method based on three-dimensional point cloud semantic segmentation and dynamic topology reconstruction to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction includes the following steps: Acquire real-time point clouds from LiDAR and perform online semantic segmentation to extract structural elements and connectivity features; Multi-frame semantic point clouds are stitched together into a local sub-map. Within the sub-map, with wall distribution and passable areas as constraints, an attribution probability distribution is generated for each spatial location to form a probability attribution field. In the probability attribution field, a hierarchical topology graph is dynamically constructed based on the peak region of the attribution probability and the interconnection relationship of the connecting elements. The nodes of the topology graph correspond to spatial units with probability boundaries. The nodes maintain address labels and attribution template features, and the edges record the passage probability and gate state of the connecting nodes. When the robot is self-localizing and addressing, it performs probability matching between the current local semantic point cloud and the belonging template of each node, outputs the probability distribution of the current location address, and uses the address with the highest probability as the localization result. If the highest probability is lower than a set threshold, it triggers active boundary exploration to refine the belonging field. After receiving the target address, a path search is performed on the hierarchical topology map to plan a path that includes a deterministic backbone node sequence and a transition strategy for the boundary of an ambiguous region. During navigation, the probability of belonging to the current location is continuously calculated. When the probability of belonging to adjacent nodes alternates in leading, the address switch is confirmed by delaying the connection event of crossing the gate or channel in order to suppress address jitter.

[0005] In a preferred embodiment, acquiring real-time point clouds from a lidar radar and performing online semantic segmentation to extract structural elements and connectivity features specifically includes: Motion distortion compensation for point clouds is performed using inertial measurement units. The compensated point cloud is then divided into cubic units. The mean coordinates and mean reflection intensity within non-empty units are used as inputs. A sparse convolutional neural network is employed to predict semantic categories unit by unit, resulting in preliminary semantic labels for walls, floors, doors, and furniture.

[0006] In a preferred embodiment, the doors in the connection elements are extracted in the following manner: Units marked as gates are aggregated into gate point cloud clusters according to spatial connectivity. After fitting the principal plane, the number of points in the cluster whose distance from the principal plane is less than a preset distance threshold is counted as the number of opening points. The ratio of the number of opening points to the baseline number of points on the same plane in the closed state of the gate obtained offline is used as the degree of opening.

[0007] In a preferred embodiment, the passage confidence of the gate is updated using an iterative smoothing method, wherein the passage confidence of the current frame is a weighted sum of the confidence of the previous frame and the incremental value determined according to the opening degree.

[0008] In a preferred embodiment, the ground plane in the structural element is extracted in the following manner: Starting with the lowest unit, select units whose semantic label is "ground" and whose local normal vector forms an angle less than a preset angle as seeds. Expand and merge these seeds into adjacent units that meet the same conditions until no new units can be incorporated. The resulting connected region is then considered a passable area.

[0009] In a preferred embodiment, the walls in the structural elements and the open passages in the connecting elements are extracted in the following manner: The wall unit is projected onto the horizontal plane, and the wall line segment is obtained by fitting the projection point set with a polyline. Identify gaps in the wall line enclosure outline that are not occupied by wall units on the horizontal plane. When the gap width is greater than the preset width and there is no unit marked as a door at the gap position, mark the gap as an open passage.

[0010] In a preferred embodiment, stitching together multiple frames of semantic point clouds into a local sub-image and forming a probability attribution field includes: A local coordinate system is established with the robot's starting position as the origin. Odometry is used to transform multiple consecutive frames of semantic point clouds into the local coordinate system. The accumulated point clouds within a preset time period form a local sub-graph. The wall points are projected as occupied grids, and the ground points are projected as passable grids. Candidate functional areas are obtained and the core points of each area are determined through connectivity analysis. Starting from all core points simultaneously, multi-source expansion is performed on the passable grid, and the number of grid steps from each grid to each core point is calculated. For each grid cell, the probability of that grid cell being assigned to each candidate region is calculated based on the number of steps taken, using an exponential decay method that increases with the number of steps. The probability of wall grid cells being assigned to a region is set to zero, thus forming a probability assignment field.

[0011] In a preferred embodiment, dynamically constructing a hierarchical topology graph in the probability attribution field includes: Grids with an attribution probability higher than the first probability threshold are merged into reliable core regions, and each reliable core region is a topology node. Each node maintains its spatial extent outline, a normalized vector of internal furniture category distribution as a homepage template feature, and an address label.

[0012] In a preferred embodiment, probability matching is performed during robot self-localization and addressing, including: Accumulate furniture semantic points within a preset spatial range centered on the robot's current position within a preset time period before the current moment, count the number of each type of furniture and normalize them to obtain a query vector; if the number of points is insufficient, skip template matching. Extract the probability of the current grid cell belonging to each node from the probability attribution field; For each node, the similarity between the query vector and the node's belonging template is calculated, and the result is weighted and summed with the belonging probability to obtain the overall matching probability.

[0013] In a preferred embodiment, after receiving the target address, a path search is performed, including: The target address is matched with the topology node label to determine the target node. When multiple nodes are matched, the node with the highest probability of belonging to the current location and not lower than the fourth probability threshold is selected. Obtain at least one node with the highest probability of belonging to the current location as a candidate starting node; Set the passage cost of the topological edge to monotonically increase as the passage probability decreases; Starting from each candidate starting node and ending at the target node, the sequence of nodes with the minimum cost is searched to form the backbone node sequence.

[0014] The technical effects and advantages of this invention are as follows: 1. This invention solves the problem that traditional area segmentation easily produces fragmented nodes and address labels jump back and forth during robot localization in continuous spaces such as open office areas and corridor intersections due to the lack of clear physical boundaries. It extracts elements such as walls, furniture and doors through online semantic segmentation, and constructs a continuous probability attribution field with wall distribution and passable areas as constraints, thus probabilizing spatial attribution. It dynamically constructs a hierarchical topology map and uses attribution template probability matching and active exploration to achieve stable self-localization. A soft boundary transition strategy is introduced during path planning. In navigation, it combines door or passage events to implement delayed confirmation address switching and updates edge passability probability and attribution field in real time. This achieves stable and continuous address output by the robot in dynamic indoor environments, suppresses address jitter, and ensures addressing determinism and environmental adaptability.

[0015] 2. This invention obtains the probability distribution of spatial location belonging to each functional area, rather than a discrete, either-or classification. This eliminates the problem of fragmented nodes and area merging caused by the selection of segmentation thresholds. During self-localization and addressing, it uses the local furniture composition and node classification template for probability matching, outputs the probability distribution of the address, and uses the address with the highest probability as the positioning result. When the matching is ambiguous, it actively explores and refines the classification field, effectively avoiding the address label from jumping back and forth between the corridor and the meeting room due to hard classification in transitional areas such as the door of the meeting room, thus ensuring the stability of the positioning output. Attached Figure Description

[0016] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example: Figure 1 The present invention provides a robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction, comprising the following steps: Acquire real-time point clouds from LiDAR and perform online semantic segmentation to extract structural elements such as walls, furniture, and floors, as well as connecting elements such as doors and passageways; Multi-frame semantic point clouds are stitched together into a local sub-map. Within the sub-map, with wall distribution and passable areas as constraints, a probability distribution of the attribution to each functional area is generated for each spatial location, forming a continuous probability attribution field. In the probability attribution field, a hierarchical topology graph is dynamically constructed based on the peak region of the attribution probability and the interconnection relationship of the connecting elements. The nodes of the topology graph correspond to spatial units with probability boundaries. The nodes maintain address labels and attribution template features, and the edges record the passage probability and gate state of the connecting nodes. When the robot is self-localizing and addressing, it performs probability matching between the current local semantic point cloud and the belonging template of each node, outputs the probability distribution of the current location address, and uses the address with the highest probability as the localization result. If the highest probability is lower than a set threshold, it triggers active boundary exploration to refine the belonging field. After receiving the target address, a path search considering the uncertainty of node affiliation is performed on the hierarchical topology map to plan a path that includes a deterministic backbone node sequence and a transition strategy for the boundary of the ambiguous affiliation region. During navigation, the attribution probability of the current location is continuously calculated. When the attribution probabilities of adjacent nodes alternately take the lead, the address switch is confirmed by delaying the connection event of crossing the gate or channel to suppress address jitter. At the same time, the passage probability of the edge and the attribution field are updated by real-time observation to maintain the dynamic consistency of the environment.

[0019] In this embodiment of the invention, real-time point clouds from a lidar sensor are acquired and online semantic segmentation is performed to extract structural elements such as walls, furniture, and floors, as well as connecting elements such as doors and passageways. Specifically, this includes: Using the angular velocity and acceleration output by the inertial measurement unit, linear interpolation compensation is performed on the pose deviation of different scanned points within the same frame of point cloud caused by robot motion, eliminating motion distortion. The space containing the compensated point cloud is divided into cubic units with a side length d=5cm. Only units containing at least one scanned point are retained. The average three-dimensional coordinates and average reflection intensity of all points in each non-empty unit are calculated to constitute the 4-dimensional input features (x, y, z, intensity) of that unit.

[0020] In one embodiment disclosed in this application, a lightweight 3D semantic segmentation network based on sparse convolution is used to predict the semantic category of the input features unit by unit. The network is based on the MinkUNet architecture, with the number of channels in each layer of the backbone network compressed to half of the original. The side length of the input cube unit is set to 5cm. In an indoor environment, approximately 20,000 points of LiDAR data per frame are segmented into approximately 5,000–8,000 non-empty units. The network's inference time per frame on an embedded GPU is approximately 18ms. The network outputs the probability of each unit belonging to one of five categories: wall, floor, door, furniture, and others. The category with the highest probability is taken as the initial semantic label. For example, if a unit outputs a probability of [0.85, 0.05, 0.03, 0.04, 0.03], corresponding to the highest probability of "wall," then the unit is labeled as "wall."

[0021] The initially labeled cells are aggregated into independent point cloud clusters for each gate based on spatial connectivity. Each gate's point cloud cluster contains all points scanned within that gate in the current frame. For each gate's point cloud cluster, its principal plane is fitted using a random sampling consensus algorithm, with a distance threshold of 2 cm and 100 iterations, to obtain the plane equation and normal vector. Points within the cluster whose vertical distance to the principal plane is less than the distance threshold are counted. Number of points For example, when a standard interior door is fully closed, after offline point cloud acquisition and processing, the number of points falling on the main plane in the closed state is... The calibration is set at 2150 points. During a real-time scan, the door was in a half-open state, and only 860 points within the main plane met the distance threshold. Therefore, the current opening degree... .

[0022] Update the passability confidence of the gate as a connection element using an iterative smoothing method. : The smoothing coefficient is among them. Take 0.9, For a piecewise linear function, when hour ,when hour ,when hour t represents the frame number, and the first time this gate is detected, it will be... Initialize to Using the example above, the following was observed the instant the door opened. , Initial confidence level The door was observed to be fully open for 10 consecutive frames, each frame... , After iteration A gradual increase from 0.4 and a stable value above 0.99 indicates that the door is continuously passable.

[0023] Extracting connected components and passable regions on the ground: Starting with the lowest-height cells, select those semantically labeled as "ground" and whose local point cloud normal vectors form an angle less than a preset angle with the vertical direction. The unit is used as the ground seed. Starting from each ground seed, check its six directly adjacent units in the six directions. If the adjacent units also satisfy the ground label and the angle between their normal vectors is less than 1, then the next adjacent unit is considered. If a new ground seed cell is found, it is merged into the current ground connectivity domain, and the process continues to expand outwards from the newly added cell to its neighbors until no more cells satisfying the conditions can be incorporated. This process is repeated until all ground seed cells are assigned to a connectivity domain, and each connectivity domain constitutes an independent walkable area. For example, in a 30m... 2Within the room, this step yields a complete and interconnected indoor floor area, with edges ending at wall units and the central portion covering the entire standing area.

[0024] Extracting wall structure: Projecting units semantically labeled as walls onto a horizontal plane yields a 2D point set. The projected point set is sorted and deduplicated at 2cm intervals along the horizontal direction. Using an iterative endpoint fitting method, starting from one end of the point set, a straight line is formed connecting the first and last points. The vertical distance from each intermediate point to this line is calculated. If the maximum distance exceeds 0.1m, the point set is divided into two segments. This process is repeated for each segment until the maximum distance between any point in each segment and its endpoints does not exceed 0.1m. The line segments connecting the endpoints of each segment are then output as wall line segments.

[0025] Detecting open passages: On a horizontal plane, extract wall line segments are used to form closed or semi-closed contours. Gaps between these contours that are not occupied by wall units are identified. The minimum passage width of each gap is measured. If the width exceeds a preset width of 0.8m, the gap location is checked to see if it corresponds to a unit marked as a door in three-dimensional space. If no door unit exists, the gap is marked as an open passage.

[0026] Passability confidence Door areas with a value higher than 0.6 and the aforementioned open passageways are marked as connection elements; the unit clusters corresponding to the wall lines, passable areas, and furniture categories, along with the connection elements, are summarized and output as the structural elements and connection elements.

[0027] In this embodiment of the invention, multiple frames of semantic point clouds are stitched together to form a local sub-image. Within the sub-image, constrained by wall distribution and passable areas, a probability distribution of the attribution to each functional area is generated for each spatial location, forming a continuous probability attribution field. Specifically, this includes: A local coordinate system is established with the robot's starting position as the origin. The inter-frame pose transformation output by the front-end laser odometry is used to transform the semantic point cloud of each continuously acquired frame to the local coordinate system. All point clouds accumulated within a preset time of 5 seconds constitute a local sub-graph, and the semantic label and source frame number of each point are retained.

[0028] Project all points in the local subgraph that are semantically labeled as walls onto the horizontal plane to generate a two-dimensional occupied grid at a resolution of 5cm. Grids into which the projected points fall are marked as occupied. At the same time, project points that are semantically labeled as ground onto the same grid plane. Grids into which the projected points fall and which are not occupied by walls are marked as passable.

[0029] Within a passable grid, examine the connectivity of each grid cell with its four neighboring cells. All directly or indirectly connected grid cells constitute a connected free domain. Grid cells with an area less than 1m² are discarded. 2The fragmented domains and the remaining connected free domains are the candidate functional regions, and each candidate region is assigned a unique number.

[0030] For each candidate region, the coordinates of all accessible grid cells within it are statistically analyzed, and the arithmetic mean of the horizontal and vertical coordinates is calculated to obtain the geometric center coordinates of the region. The grid cell within the candidate region that has the closest Euclidean distance to the geometric center coordinates is selected as the core point of the region.

[0031] Subsequently, starting from the core points of all candidate regions, a multi-source breadth-first expansion is performed: a first-in-first-out queue is maintained, all core point grids are added to the queue and the region number to which each core point belongs is recorded, and an initial distance of 0 is set between each core point grid and the region corresponding to that region.

[0032] Each time a grid is taken from the queue, its four neighboring grids marked as passable and not occupied by walls are checked. If the distance of the neighboring grid has not been recorded by the region associated with the currently taken grid, its distance is set to the current grid distance plus 1, and the neighboring grid is added to the queue. At the same time, the region from which the distance comes from the core point of the expansion starts is recorded. If the distance of the neighboring grid has been recorded by other regions, the records of each region are retained, and it is not added to the queue again.

[0033] Repeat this process until the queue is empty. All grids reachable from a core point via accessible grids will be assigned an integer step distance corresponding to that region. For the... Candidate regions, grid The distance value recorded at this location is denoted as If a certain grid cannot be obtained from the area Once the core point is reached, then It is set to a fixed value greater than the maximum possible number of steps in the entire graph. .

[0034] Based on this, the grid For the first The probability of belonging to a candidate region is defined as: ,in, The attenuation coefficient is set to 10 grid steps (corresponding to a physical distance of 0.5 m), controlling the rate at which the assignment probability decreases as the detour distance increases; the denominator is summed over all candidate regions. For grids occupied by walls, all assignment probabilities are set to 0.

[0035] This allocation method determines the affiliation of each grid cell based on the detour distance caused by walls, rather than a simple spatial Euclidean distance. The affiliation probability vector of each grid cell is ultimately stored as a probability affiliation field. For example, within the same room, the affiliation probability of a grid cell to that room is generally higher than 0.9; at the doorway between two areas, the affiliation probability shows a continuous gradual change, with grid cells immediately adjacent to the center line of the doorway potentially exhibiting probability distributions of 0.55 and 0.45, gradually shifting the advantage after crossing the center line, thus avoiding affiliation jumps caused by hard boundaries. This probability affiliation field provides continuous spatial affiliation information for the subsequent construction of the hierarchical topology map.

[0036] In this embodiment of the invention, in the probability attribution field, a hierarchical topology graph is dynamically constructed based on the peak region of the attribution probability and the connectivity of the connecting elements. The nodes of the topology graph correspond to spatial units with probability boundaries. Nodes maintain address labels and attribution template features, and edges record the passage probability and gate state of connected nodes. Specifically, this includes: Traverse all grids in the probability attribution field and initially mark grids with a maximum attribution probability exceeding the first probability threshold of 0.8 as core grids. Merge spatially adjacent core grids into connected groups, and remove those containing fewer than 100 grids (corresponding to a physical area of ​​less than 0.25m²). 2 The remaining connected groups are treated as a reliable core region, and each reliable core region is assigned a unique node identifier, which is a spatial unit node in the topology graph.

[0037] When two reliable core areas are not directly adjacent, but there are no wall grids on the shortest path between them, and the maximum value of the attribution probability in the grids traversed by the shortest path is within a preset range of 0.4 to 0.8, and the attribution probability advantage gradually transitions from one area to another, a transition zone is determined to exist between these two core areas. Identified connectivity elements within the transition zone are extracted: if there is a door within the transition zone, the passability confidence of that door is used as the basis for identification. The initial passage probability is used as the edge, and the state and opening degree of the door are recorded on the edge. If there are no doors in the transition zone but there are marked open passages, the initial passability of the edge is set to 1.0 and marked as a permanent opening. If there are neither doors nor open passages in the transition zone, the edge is not created temporarily and will be created later when connecting elements are detected in the transition zone during subsequent observations. The information recorded for an edge includes the identifiers of the two endpoints, the passability probability, the type of connecting element, and the timestamp of the last update of the edge.

[0038] After each node is established, the following information is extracted from its corresponding reliable core area and maintained: spatial range, represented by the sequence of convex hull vertices of all grids within the core area; attribution template features, which are obtained by counting the number of each type of furniture in the units with the semantic label "furniture" within the core area, and dividing the number of each type of furniture by the total number of furniture to obtain a normalized vector. For example, if there are 625 furniture elements in a core area, including 320 chairs, 210 tables, and 95 cabinets, then the attribution template feature vector is [0.51, 0.34, 0.15]; and address labels, which are initially automatically generated semantic placeholders in the form of "region_01". Subsequent labels are obtained by receiving manual annotations or matching with the attribution template features in a pre-built address fingerprint database. When the vector cosine similarity exceeds 0.85, the label is automatically replaced with the corresponding semantic address.

[0039] The topology graph is dynamically adjusted as the probability attribution field is updated, with the following specific rules: When a reliable core region appears in the probability attribution field that does not correspond to any existing node, a new node is created and its information is extracted in the manner described above; when the outline of a reliable core region corresponding to an existing node changes, causing the area change to exceed 20% of the original area, its spatial convex hull is recalculated and the attribution template feature vector is updated; when the gate access confidence corresponding to an edge... When the difference from the current recorded value exceeds 0.2, the passage probability and timestamp of the edge are refreshed with the new value; when an edge has not been detected in any frame of observation for more than 15 seconds, its passage probability is multiplied by 0.9. If it has not been observed for 30 consecutive seconds, the passage probability is set to 0 and it is marked as blocked, so as to keep the topology graph following the opening and closing of doors and environmental changes in real time.

[0040] In this embodiment of the invention, during robot self-localization, the current local semantic point cloud is probabilistically matched with the attribution template of each node to output the probability distribution of the current location address. The address with the highest probability is used as the localization result. If the highest probability is lower than a set threshold, active boundary exploration is triggered to refine the attribution field, specifically including: Accumulate semantic point clouds within a preset spatial range (radius 2.5m) centered on the robot's current pose, within a preset time period of 0.5 seconds prior to the current moment. Select points whose semantic label is "furniture," and count them separately according to preset furniture categories such as chairs, tables, and cabinets, forming a counting vector of length k. Divide this vector by the sum of its components to obtain the normalized query distribution vector. This vector directly reflects the relative proportion of furniture types within a local area. If the total number of furniture points within the cumulative range is less than 20, template matching is temporarily skipped, and the process proceeds directly to the subsequent field affiliation probability determination stage.

[0041] Extract the ownership probability distribution corresponding to the grid cell where the robot is currently located from the probability ownership field. ,in Traverse all nodes in the current topology graph. If the topology graph is still empty or the current grid has not yet formed an effective distribution in the probability attribution field, then all nodes will be... Set to 0.

[0042] For each existing topology node , and its pre-stored attribution template feature vector and Calculate cosine similarity This template vector is also obtained by normalizing the furniture category proportion when the node is created, and... Same dimension. Node The overall matching probability is defined as The weight This allows the currently observed furniture composition to have a higher influence than historically accumulated location priors; if template matching is skipped due to insufficient furniture points, then... It depends entirely on the field ownership probability, that is .

[0043] Press all nodes Sort the data in descending order to obtain the probability distribution of the current location address. Take the address with the highest probability. The corresponding node address label is output as the location result. If This self-location process is now complete, and the address has been confirmed. If This indicates that the existing topological map has significant ambiguity in its analysis of the current location, thus triggering active boundary exploration: First, determine the target node for exploration: select the node with the highest current overall matching probability as the reference node, and denot its number as . In the probabilistic attribution field, taking the current grid as the center, examine the four neighboring grids marked as passable, and calculate the node pairs of each neighboring grid. The robot is moved along this direction, with the largest increment between the probability of belonging to the current grid and the probability of belonging to the current grid. If multiple directions have the same increment, the direction towards the unexplored area is selected first. The robot moves along this direction, continuously collecting new semantic point clouds during the movement, updating the probability belonging field and local query vector, and re-executing the matching calculation. Once the overall matching probability of any node rises above 0.6, exploration is immediately stopped and the node address is output. If exploration continues for more than 30 seconds without reaching the matching threshold, and the newly entered area does not form a belonging probability higher than 0.4 with any existing node in the probability belonging field, the area is added to the topology as a new node, assigned a temporary placeholder address, and this temporary address is used as the localization result to ensure the continuity of addressing output.

[0044] In this embodiment of the invention, after receiving the target address, a path search considering the uncertainty of node affiliation is performed on the hierarchical topology graph to plan a path that includes a deterministic backbone node sequence and a transition strategy for the boundary of ambiguously affiliated regions. Specifically, this includes: The target address string is precisely matched or synonymously mapped with the address labels of each node in the topology map to determine a target node. If multiple nodes are matched, the node with the highest probability of belonging to the grid where the robot's position is located in the current probability attribution field, and whose probability is not lower than the fourth probability threshold of 0.3, is selected as the target node. If all probabilities are lower than 0.3, an "unclear target address" message is returned. Simultaneously, the probability distribution of the robot's current position grid to all topology nodes is obtained from the self-localization module. The two nodes with the highest probabilities are selected as candidate starting nodes and retained regardless of their probability values ​​to accommodate scenarios with ambiguous localization.

[0045] For each edge in the topological graph, its probability of passage It is already dynamically maintained. To naturally bypass blocking or near-closed connections in path planning, the edge cost is defined as: This form ensures that the cost approaches 1 when the probability of passage approaches 1, and the cost increases sharply when the probability of passage approaches 0, thereby guiding the path to prioritize passing through unobstructed connections.

[0046] Using each candidate starting node as the origin and the target node as the destination, Dijkstra's algorithm is employed to search for a node sequence with the minimum total cost on the topology graph. When both candidate starting nodes can reach the target node, the sequence with the lower total cost is selected as the final path; if the total costs are equal, the sequence with the higher probability of the starting node belonging to the target node is selected. The resulting backbone node sequence is the deterministic topological path.

[0047] For each pair of adjacent nodes in the backbone node sequence, determine their connectivity characteristics in the probability attribution field: If two nodes are directly adjacent on a horizontal grid and connected by a single connection element (door or open channel), and there are no continuous transition grids with attribution probabilities between 0.4 and 0.8 in between, then this segment is a deterministic boundary and no additional transition strategy is needed. If there is a continuous region between two nodes where the attribution probability gradually transitions from the current node's dominance to the adjacent node's dominance, i.e., a transition zone, then extract the equivalent grids with an attribution probability of 0.5 from this transition zone and connect them to form a soft boundary line. The corresponding boundary transition strategy is as follows: After the robot enters the transition zone along the path, if the attribution probability of the current grid to the next node exceeds the attribution probability to the current node for three consecutive frames, and the robot has already crossed the soft boundary, then the address output is switched to the next node; if there is a door in the transition zone and the door is detected as open at the moment of crossing, then the address switch is immediately confirmed; if the door becomes closed during the crossing process, then the movement is paused and the path is replanned based on the updated edge passage probability.

[0048] The planning output consists of two parts: the first part is the sequence of backbone nodes and the order of movement of nodes along the topological edges; the second part is the soft boundary position, switching conditions and real-time status of connection elements provided for each connection segment that requires a transition strategy, which is used to drive the robot to smoothly transition between deterministic nodes during actual navigation and avoid address output jitter due to ambiguity of ownership.

[0049] In this embodiment of the invention, the attribution probability of the current location is continuously calculated during navigation. When the attribution probabilities of adjacent nodes alternately lead, address switching is confirmed by delaying the connection event of a passage gate or channel to suppress address jitter. At the same time, the passage probability of the edge and the attribution field are updated by real-time observation to maintain dynamic consistency of the environment. Specifically, this includes: After acquiring a new semantic point cloud for each frame, the probability of the grid cell belonging to the robot on the current path is extracted from the probability attribution field based on the robot's current pose. and When the difference between the two values ​​is less than 0.1 and they alternately lead each other within three consecutive frames, the system is determined to be in a state of ambiguity regarding ownership. At this point, the output address label is not switched. Next, the type of connection element linking the two nodes is determined: if it is a gate, the main plane fitted by that gate is used as the crossing interface; if it is an open channel, the ownership probability field corresponding to that channel is used. The iso-raster lines serve as a virtual traversal interface.

[0050] The criterion for determining a crossing event is whether the robot has completely moved from one side of the interface to the other in space. Specifically, the signs of the robot's pose in the previous frame and the current frame relative to the normal vector direction of the crossing interface are recorded. When the sign changes and the vertical distance from the current frame pose to the interface is less than 0.3m, the robot is considered to be crossing the interface, the original address label is maintained, and a crossing marker is recorded. When the pose sign changes and continues to move away from the interface, causing the vertical distance from the robot to the interface to exceed 0.3m, and the sign no longer changes repeatedly, it is considered to have left the range of the connected element. After leaving, the attribution probability of 5 consecutive frames is collected. If in at least 4 of these 5 frames the attribution probability of the same node is higher than that of another node, and the average probability value of the leading node in the 5 frames is more than 0.05 higher than the average probability value of the other node, the address output is switched to the leading node, completing the hysteresis confirmation. This hysteresis mechanism only suppresses address switching during the crossing of the interface, effectively preventing address jitter caused by observation noise in doorways or passages.

[0051] Synchronous update of environmental representation: For gate connection elements, extract the current frame aperture degree. ,according to Update the access confidence and refresh the access probability of the corresponding edge; for the probability attribution field, project the wall points and ground points of the current frame onto the two-dimensional grid, update the occupied grid and accessible grid, and then, taking the core point of each node as the source, re-perform multi-source distance extension on the updated accessible grid, recalculate the attribution probability distribution of each grid, so that the attribution field is consistent with the latest observation.

[0052] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0053] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0054] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction, characterized in that: Includes the following steps: Acquire real-time point clouds from LiDAR and perform online semantic segmentation to extract structural elements and connectivity features; Multi-frame semantic point clouds are stitched together into a local sub-map. Within the sub-map, with wall distribution and passable areas as constraints, an attribution probability distribution is generated for each spatial location to form a probability attribution field. In the probability attribution field, a hierarchical topology graph is dynamically constructed based on the peak region of the attribution probability and the interconnection relationship of the connecting elements. The nodes of the topology graph correspond to spatial units with probability boundaries. The nodes maintain address labels and attribution template features, and the edges record the passage probability and gate state of the connecting nodes. When the robot is self-localizing and addressing, it performs probability matching between the current local semantic point cloud and the belonging template of each node, outputs the probability distribution of the current location address, and uses the address with the highest probability as the localization result. If the highest probability is lower than a set threshold, it triggers active boundary exploration to refine the belonging field. After receiving the target address, a path search is performed on the hierarchical topology map to plan a path that includes a deterministic backbone node sequence and a transition strategy for the boundary of an ambiguous region. During navigation, the probability of belonging to the current location is continuously calculated. When the probability of belonging to adjacent nodes alternates in leading, the address switch is confirmed by delaying the connection event of crossing the gate or channel in order to suppress address jitter.

2. The robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction according to claim 1, characterized in that: Acquire real-time point clouds from LiDAR and perform online semantic segmentation to extract structural elements and connectivity features, specifically including: Motion distortion compensation for point clouds is performed using inertial measurement units. The compensated point cloud is then divided into cubic units. The mean coordinates and mean reflection intensity within non-empty units are used as inputs. A sparse convolutional neural network is employed to predict semantic categories unit by unit, resulting in preliminary semantic labels for walls, floors, doors, and furniture.

3. The robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction according to claim 2, characterized in that: The gates in the connection elements are extracted in the following way: Units marked as gates are aggregated into gate point cloud clusters according to spatial connectivity. After fitting the principal plane, the number of points in the cluster whose distance from the principal plane is less than a preset distance threshold is counted as the number of opening points. The ratio of the number of opening points to the baseline number of points on the same plane in the closed state of the gate obtained offline is used as the degree of opening.

4. The robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction according to claim 3, characterized in that: The passage confidence of the gate is updated using an iterative smoothing method. The passage confidence of the current frame is a weighted sum of the confidence of the previous frame and the incremental value determined according to the opening degree.

5. The robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction according to claim 2, characterized in that: The ground plane in the structural element is extracted in the following way: Starting with the lowest unit, select units whose semantic label is "ground" and whose local normal vector forms an angle less than a preset angle as seeds. Expand and merge these seeds into adjacent units that meet the same conditions until no new units can be incorporated. The resulting connected region is then considered a passable area.

6. The robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction according to claim 2, characterized in that: The walls in the structural elements and the open passages in the connection elements are extracted in the following ways: The wall unit is projected onto the horizontal plane, and the wall line segment is obtained by fitting the projection point set with a polyline. Identify gaps in the wall line enclosure outline that are not occupied by wall units on the horizontal plane. When the gap width is greater than the preset width and there is no unit marked as a door at the gap position, mark the gap as an open passage.

7. The robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction according to claim 1, characterized in that: Multi-frame semantic point clouds are stitched together to form local sub-images and a probability attribution field, including: A local coordinate system is established with the robot's starting position as the origin. Odometry is used to transform multiple consecutive frames of semantic point clouds into the local coordinate system. The accumulated point clouds within a preset time period form a local sub-graph. The wall points are projected as occupied grids, and the ground points are projected as passable grids. Candidate functional areas are obtained and the core points of each area are determined through connectivity analysis. Starting from all core points simultaneously, multi-source expansion is performed on the passable grid, and the number of grid steps from each grid to each core point is calculated. For each grid cell, the probability of that grid cell being assigned to each candidate region is calculated based on the number of steps taken, using an exponential decay method that increases with the number of steps. The probability of wall grid cells being assigned to a region is set to zero, thus forming a probability assignment field.

8. The robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction according to claim 1, characterized in that: Dynamically constructing a hierarchical topology graph in the probability attribution field includes: Grids with an attribution probability higher than the first probability threshold are merged into reliable core regions, and each reliable core region is a topology node. Each node maintains its spatial extent outline, a normalized vector of internal furniture category distribution as a homepage template feature, and an address label.

9. The robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction according to claim 1, characterized in that: Probabilistic matching is performed during robot self-localization and addressing, including: Accumulate furniture semantic points within a preset spatial range centered on the robot's current position within a preset time period before the current moment, count the number of each type of furniture and normalize them to obtain a query vector; if the number of points is insufficient, skip template matching. Extract the probability of the current grid cell belonging to each node from the probability attribution field; For each node, the similarity between the query vector and the node's belonging template is calculated, and the result is weighted and summed with the belonging probability to obtain the overall matching probability.

10. The robot addressing method based on 3D point cloud semantic segmentation and dynamic topology reconstruction according to claim 1, characterized in that: After receiving the target address, perform a path search, including: The target address is matched with the topology node label to determine the target node. When multiple nodes are matched, the node with the highest probability of belonging to the current location and not lower than the fourth probability threshold is selected. Obtain at least one node with the highest probability of belonging to the current location as a candidate starting node; Set the passage cost of the topological edge to monotonically increase as the passage probability decreases; Starting from each candidate starting node and ending at the target node, the sequence of nodes with the minimum cost is searched to form the backbone node sequence.

Images