Positioning method and apparatus, electronic device, and storage medium

Abstract

Embodiments of the present disclosure provide a positioning method, device, electronic equipment and storage medium. The positioning method comprises: based on the scan point data obtained by the robot in the first target scene at the current position, using the first semantic label corresponding to each scan point in the scan point data identified by the semantic segmentation model, wherein the first semantic label is used to indicate the probability of the corresponding scan point being at least one obstacle type; based on the scan point data and the first semantic label, the matching result obtained by matching the first grid map of the first target scene is determined, and the pose of the robot on the first grid map is determined, wherein the first grid map at least includes: the second semantic label used to indicate the occupied grid in the first grid map, wherein the second semantic label is used to indicate the probability of the corresponding grid being at least one obstacle type.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a positioning method, apparatus, electronic device, and storage medium. Background Technology

[0002] In the field of industrial and commercial mobile robots, 2D LiDAR is commonly used for obstacle perception. 2D LiDAR utilizes a single-line laser to perform a horizontal omnidirectional scan within a fixed two-dimensional plane, acquiring two-dimensional cross-sectional distance information of the environment. Due to its inherent advantages of low cost, small data volume, and high real-time performance, 2D LiDAR has been widely adopted. However, the imaging principle of 2D LiDAR dictates that it can only perceive geometric structural features within a single horizontal plane, severely lacking spatial information in the height dimension. In traditional 2D LiDAR-based localization schemes, the environment is typically abstracted as a binary grid map, where each grid cell only has two states: idle and occupied (i.e., only distinguishing between empty space and obstacles). When a robot is in complex scenarios with numerous highly repetitive and symmetrical structures, such as warehouses, supermarkets, or corridors, the geometric shapes of obstacles at different locations on the 2D cross-section are highly similar. Geometric matching algorithms are prone to multiple localization solutions and position ambiguity during relocalization, failing to uniquely determine the robot's current pose.

[0003] Therefore, how to accurately determine the pose of a robot using 2D LiDAR is an urgent problem to be solved. Summary of the Invention

[0004] This disclosure provides positioning methods, apparatus, electronic devices, and storage media.

[0005] According to a first aspect of the embodiments of this disclosure, a positioning method is proposed, the method comprising: Based on the scan point data obtained by the robot at its current position in the first target scene, a semantic segmentation model is used to identify the first semantic label corresponding to each scan point in the scan point data, wherein the first semantic label is used to indicate the probability that the corresponding scan point is at least one type of obstacle; Based on the scan point data and the first semantic label, the matching result obtained by matching the first grid map of the first target scene determines the pose of the robot on the first grid map, wherein the first grid map includes at least: a second semantic label for indicating the occupied grid in the first grid map, wherein the second semantic label is used to indicate the probability that the corresponding grid is at least one obstacle type.

[0006] In some embodiments, determining the robot's pose on the first grid map based on the matching result obtained by matching the scan point data and the first semantic label against the first grid map of the first target scene includes: Determine the first geometric feature quantity corresponding to the scan point data, and based on the comparison results between the first geometric feature quantity and the second geometric feature quantity corresponding to at least one observation coordinate in the first grid map, determine at least one candidate positioning coordinate from multiple observation points; At different poses on each candidate positioning coordinate, the scanning point corresponding to the scanning point data is projected onto the first grid map, and the joint likelihood field result of the scanning point corresponding to the scanning point data and the first grid map is confirmed respectively. The joint likelihood field result is determined based on the first geometric likelihood field result of the scanning point and the obstacles in the first grid map, and the first semantic consistency comparison result of comparing the first semantic label of the scanning point and the second semantic label of the grid corresponding to the scanning point. Based on the joint likelihood field results at different poses at each candidate positioning coordinate, the preliminary matching pose of each candidate positioning coordinate is determined. For all preliminary matching poses, geometric registration is performed based on the scanned point data and the obstacles in the first grid map to determine at least one registration matching pose; The robot's pose on the first grid map is determined based on the joint positioning confidence corresponding to at least one registration-matched pose.

[0007] In some embodiments, the joint likelihood field result includes a weighted sum of the first geometric likelihood field result and the first semantic consistency comparison result.

[0008] In some embodiments, the joint positioning confidence is determined based on at least one of the following: the second geometric likelihood field result corresponding to the registration-matched pose, the geometric distance evaluation result of the scan point, and the second semantic consistency comparison result; wherein: The second geometric likelihood field includes: the likelihood field of all scan points and obstacles in the first grid map when the robot projects the scan points corresponding to the scan point data onto the first grid map under the registration and matching pose; The evaluation results of the geometric distance of the scanning points include: the evaluation results of the average distance of all scanning points to the nearest obstacle when the robot projects the scanning points corresponding to the scanning point data onto the first grid map under the registration and matching pose; The second semantic consistency comparison result includes: the comparison result of the first semantic label of the scan point and the second semantic label of the grid corresponding to the scan point when the robot projects the scan point data onto the first grid map under the registration and matching pose.

[0009] In some embodiments, the joint location confidence determination includes the geometric mean of the product of the second geometric likelihood field result, the geometric distance evaluation result of the scan point, and the second semantic consistency comparison result.

[0010] In some embodiments, the step of performing geometric registration based on the scan point data and obstacles in the first grid map for all preliminary matched poses to determine at least one registration matching pose includes: All the preliminary matching poses are used as the iterative initial poses for geometric registration of the scanning points with obstacles in the first grid map. Registration is performed using the two-dimensional generalized iterative nearest point 2DGICP to determine at least one registration matching pose.

[0011] In some embodiments, determining the robot's pose on the first grid map based on the joint positioning confidence corresponding to at least one registration-matched pose includes at least one of the following: In the first mode, the first registered and matched pose with a joint positioning confidence level higher than the confidence threshold is determined as the pose of the robot on the first grid map; In the second mode, the registration and matching pose with the highest joint positioning confidence and a joint positioning confidence higher than the confidence threshold is determined as the pose of the robot on the first grid map.

[0012] In some embodiments, the first geometric feature quantity includes: a statistical value of the measured distance of all scan points in the scan point data; The second geometric feature quantity includes: the statistical value of the distance measured at the theoretical scanning point of the robot under the observation coordinates.

[0013] In some embodiments, the method further includes: Acquire the robot's scan point data within the second target scene; A second grid map of the second target scene is established based on the scan points in the scan point data; Label the type of obstacle occupying each grid cell in the second grid map; The obstacle type of the grid where the scan point is located is used as the obstacle type of the scan point; The semantic segmentation model is trained based on the scan point data and the obstacle type at each scan point.

[0014] In some embodiments, the method further includes: Based on the scan data obtained from the robot's global scan of the first target scene, it is determined that the first grid map is included. A semantic segmentation model is used to identify the third semantic label for each scan point in the first raster map; Based on the third semantic label of the scan point in each occupied grid in the first grid map, a second semantic label is determined for each occupied grid; the third semantic label is used to indicate the probability that the corresponding scan point is at least one type of obstacle.

[0015] In some embodiments, determining the second semantic label corresponding to each occupied raster based on the third semantic label of the scan point in each occupied raster in the first raster map includes one of the following: A Bayesian filter is used to filter the third semantic label of all scan points in the occupied grid to obtain the second semantic label of the occupied grid. The third semantic labels of all scan points in the occupied grid are accumulated, and the second semantic label of the occupied grid is determined based on the accumulated value.

[0016] In some embodiments, the scan point data includes: scan point reflection intensity data and scan point distance data.

[0017] According to a second aspect of the present disclosure, a positioning device is provided, the device comprising: a processing module, the processing module being configured to: Based on the scan point data obtained by the robot at its current position in the first target scene, a semantic segmentation model is used to identify the first semantic label corresponding to each scan point in the scan point data, wherein the first semantic label is used to indicate the probability that the corresponding scan point is at least one type of obstacle; Based on the scan point data and the first semantic label, the matching result obtained by matching the first grid map of the first target scene determines the pose of the robot on the first grid map, wherein the first grid map includes at least: a second semantic label for indicating the occupied grid in the first grid map, wherein the second semantic label is used to indicate the probability that the corresponding grid is at least one obstacle type.

[0018] In some embodiments, the processing module is specifically used for: Determine the first geometric feature quantity corresponding to the scan point data, and based on the comparison results between the first geometric feature quantity and the second geometric feature quantity corresponding to at least one observation coordinate in the first grid map, determine at least one candidate positioning coordinate from multiple observation points; At different poses on each candidate positioning coordinate, the scanning point corresponding to the scanning point data is projected onto the first grid map, and the joint likelihood field result of the scanning point corresponding to the scanning point data and the first grid map is confirmed respectively. The joint likelihood field result is determined based on the first geometric likelihood field result of the scanning point and the obstacles in the first grid map, and the first semantic consistency comparison result of comparing the first semantic label of the scanning point and the second semantic label of the grid corresponding to the scanning point. Based on the joint likelihood field results at different poses at each candidate positioning coordinate, the preliminary matching pose of each candidate positioning coordinate is determined. For all preliminary matching poses, geometric registration is performed based on the scanned point data and the obstacles in the first grid map to determine at least one registration matching pose; The robot's pose on the first grid map is determined based on the joint positioning confidence corresponding to at least one registration-matched pose.

[0019] In some embodiments, the joint likelihood field result includes a weighted sum of the first geometric likelihood field result and the first semantic consistency comparison result.

[0020] In some embodiments, the joint positioning confidence is determined based on at least one of the following: the second geometric likelihood field result corresponding to the registration-matched pose, the geometric distance evaluation result of the scan point, and the second semantic consistency comparison result; wherein: The second geometric likelihood field includes: the likelihood field of all scan points and obstacles in the first grid map when the robot projects the scan points corresponding to the scan point data onto the first grid map under the registration and matching pose; The evaluation results of the geometric distance of the scanning points include: the evaluation results of the average distance of all scanning points to the nearest obstacle when the robot projects the scanning points corresponding to the scanning point data onto the first grid map under the registration and matching pose; The second semantic consistency comparison result includes: the comparison result of the first semantic label of the scan point and the second semantic label of the grid corresponding to the scan point when the robot projects the scan point data onto the first grid map under the registration and matching pose.

[0021] In some embodiments, the joint location confidence determination includes the geometric mean of the product of the second geometric likelihood field result, the geometric distance evaluation result of the scan point, and the second semantic consistency comparison result.

[0022] In some embodiments, the processing module is specifically used for: All the preliminary matching poses are used as the iterative initial poses for geometric registration of the scanning points with obstacles in the first grid map. Registration is performed using the two-dimensional generalized iterative nearest point 2DGICP to determine at least one registration matching pose.

[0023] In some embodiments, the processing module is specifically used for at least one of the following: In the first mode, the first registered and matched pose with a joint positioning confidence level higher than the confidence threshold is determined as the pose of the robot on the first grid map; In the second mode, the registration and matching pose with the highest joint positioning confidence and a joint positioning confidence higher than the confidence threshold is determined as the pose of the robot on the first grid map.

[0024] In some embodiments, the first geometric feature quantity includes: a statistical value of the measured distance of all scan points in the scan point data; The second geometric feature quantity includes: the statistical value of the distance measured at the theoretical scanning point of the robot under the observation coordinates.

[0025] In some embodiments, the processing module is further configured to: Acquire the robot's scan point data within the second target scene; A second grid map of the second target scene is established based on the scan points in the scan point data; Label the type of obstacle occupying each grid cell in the second grid map; The obstacle type of the grid where the scan point is located is used as the obstacle type of the scan point; The semantic segmentation model is trained based on the scan point data and the obstacle type at each scan point.

[0026] In some embodiments, the processing module is further configured to: Based on the scan data obtained from the robot's global scan of the first target scene, it is determined that the first grid map is included. A semantic segmentation model is used to identify the third semantic label for each scan point in the first raster map; Based on the third semantic label of the scan point in each occupied grid in the first grid map, a second semantic label is determined for each occupied grid; the third semantic label is used to indicate the probability that the corresponding scan point is at least one type of obstacle.

[0027] In some embodiments, the processing module is specifically used for one of the following: A Bayesian filter is used to filter the third semantic label of all scan points in the occupied grid to obtain the second semantic label of the occupied grid. The third semantic labels of all scan points in the occupied grid are accumulated, and the second semantic label of the occupied grid is determined based on the accumulated value.

[0028] In some embodiments, the scan point data includes: scan point reflection intensity data and scan point distance data.

[0029] According to a third aspect of the present disclosure, an electronic device is provided, the electronic device comprising: one or more processors; wherein the processors are configured to invoke instructions to cause the electronic device to perform the positioning method described in the first aspect.

[0030] According to a fourth aspect of the present disclosure, a storage medium is provided that stores instructions that, when executed on an electronic device, cause the electronic device to perform the positioning method described in the first aspect.

[0031] The localization method according to embodiments of this disclosure includes: based on scan point data acquired by a robot at its current position within a first target scene, using a semantic segmentation model to identify a first semantic label corresponding to each scan point in the scan point data, wherein the first semantic label is used to indicate the probability that the corresponding scan point is at least one obstacle type; based on the scan point data and the first semantic label, matching results obtained by matching a first grid map of the first target scene to determine the robot's pose on the first grid map, wherein the first grid map includes at least: a second semantic label used to indicate occupied grids in the first grid map, wherein the second semantic label is used to indicate the probability that the corresponding grid is at least one obstacle type. Thus, by semantically labeling the scan point data using a semantic segmentation model and constructing a grid map containing semantic information, the robot can effectively utilize the geometric and semantic features of the environment for matching during robot localization. In complex scenes with highly repetitive and highly symmetrical structures, this method can reduce multiple localization solutions and positional ambiguity, thereby improving the pose determination accuracy and robustness of the robot in such environments. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating a positioning method according to an exemplary embodiment. Figure 1 ; Figure 2 This is a flowchart illustrating a positioning method according to an exemplary embodiment. Figure 2 ; Figure 3 This is a flowchart illustrating a positioning method according to an exemplary embodiment. Figure 3 ; Figure 4 This is a flowchart illustrating a positioning method according to an exemplary embodiment. Figure 4 ; Figure 5 This is a schematic diagram of the structure of an electronic device according to an exemplary embodiment. Detailed Implementation

[0033] To make the technical solution and beneficial effects of the present invention more apparent and understandable, a detailed description is provided below by listing specific embodiments. The accompanying drawings are not necessarily drawn to scale, and local features may be enlarged or reduced to more clearly show the details of the local features; unless otherwise defined, the technical and scientific terms used herein have the same meanings as those in the technical field to which this application pertains.

[0034] This disclosure is not exhaustive, but merely illustrative of some embodiments, and is not intended to limit the scope of protection of this disclosure. Unless otherwise specified, each step in a particular embodiment can be implemented as an independent embodiment, and the steps can be arbitrarily combined. For example, a solution after removing some steps in a particular embodiment can also be implemented as an independent embodiment, and the order of the steps in a particular embodiment can be arbitrarily interchanged. Furthermore, the optional implementation methods in a particular embodiment can be arbitrarily combined; moreover, the embodiments can be arbitrarily combined, for example, some or all steps of different embodiments can be arbitrarily combined, and a particular embodiment can be arbitrarily combined with the optional implementation methods of other embodiments.

[0035] In each of the disclosed embodiments, unless otherwise specified or in case of logical conflict, the terminology and / or descriptions of the embodiments are consistent and can be referenced by each other. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0036] The terminology used in the embodiments of this disclosure is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure.

[0037] In this disclosure, unless otherwise stated, elements expressed in the singular form, such as "a," "an," "the," "the," "the," "the," "the," "the," "this," etc., can mean "one and only one," or "one or more," "at least one," etc. For example, when using articles such as "a," "an," "the," etc. in translation, the noun following the article can be understood as either a singular or a plural expression.

[0038] In the embodiments of this disclosure, "multiple" can be two or more.

[0039] In some embodiments, the terms “at least one of”, “one or more”, “a plurality of”, “multiple”, etc., may be used interchangeably.

[0040] In some embodiments, the notation "at least one of A and B", "A and / or B", "A in one case, B in another", "A in one case, B in another", etc., may include the following technical solutions depending on the situation: in some embodiments, A (A is executed regardless of B); in some embodiments, B (B is executed regardless of A); in some embodiments, execution is selected from A and B (A and B are selectively executed); in some embodiments, both A and B are executed. The same applies when there are more branches such as A, B, C, etc.

[0041] In some embodiments, the notation "A or B" may include the following technical solutions, depending on the situation: in some embodiments, A (execution of A regardless of B); in some embodiments, B (execution of B regardless of A); in some embodiments, selective execution from A and B (A and B are selectively executed). The same applies when there are more branches such as A, B, and C.

[0042] The prefixes "first," "second," etc., used in the embodiments of this disclosure are merely for distinguishing different descriptive objects and do not impose restrictions on the position, order, priority, value, or content of the descriptive objects. The description of the descriptive objects should be found in the claims or the context of the embodiments, and the use of prefixes should not constitute unnecessary restrictions. For example, if the descriptive object is a "field," the ordinal numbers preceding "field" in "first field" and "second field" do not restrict the position or order of the "fields." "First" and "second" do not restrict whether the "fields" they modify are in the same message, nor do they restrict the order of "first field" and "second field." Similarly, if the descriptive object is a "level," the ordinal numbers preceding "level" in "first level" and "second level" do not restrict the priority between "levels." Furthermore, the value of the descriptive object is not limited by ordinal numbers and can be one or more. For example, in "first device," the value of "device" can be one or more. Furthermore, the objects modified by different prefixes can be the same or different. For example, if the object being described is "device", then "first device" and "second device" can be the same device or different devices, and their types can be the same or different. Similarly, if the object being described is "information", then "first information" and "second information" can be the same information or different information, and their content can be the same or different.

[0043] In some embodiments, “including A,” “containing A,” “for indicating A,” and “carrying A” can be interpreted as directly carrying A or indirectly indicating A.

[0044] In some embodiments, terms such as “…”, “determine…”, “in the case of…”, “when…”, “when…”, “if…”, etc. can be used interchangeably.

[0045] In some embodiments, the terms “greater than,” “greater than or equal to,” “not less than,” “more than,” “more than or equal to,” “not less than,” “higher than,” “higher than or equal to,” “not lower than,” and “above” can be used interchangeably, as can the terms “less than,” “less than or equal to,” “not greater than,” “less than,” “less than or equal to,” “not more than,” “lower than,” “lower than or equal to,” “not higher than,” and “below”.

[0046] In some embodiments, devices, etc., can be interpreted as physical or virtual, and their names are not limited to the names recorded in the embodiments. Terms such as “device”, “equipment”, “circuit”, “network element”, “node”, “function”, “unit”, “section”, “system”, “network”, “chip”, “chip system”, “entity”, and “subject” can be used interchangeably.

[0047] When a robot is in a complex environment with a large number of highly repetitive and symmetrical structures, such as a warehouse, supermarket, or corridor, the positioning scheme based on 2D LiDAR can only perceive the geometric features within a single horizontal plane. Furthermore, the environment is abstracted into a binary grid map. As a result, the geometric matching algorithm is prone to multiple positioning solutions and position ambiguity during relocalization, making it impossible to uniquely determine the robot's current pose.

[0048] This disclosure proposes a positioning method, such as... Figure 1 As shown, the specific implementation steps of the positioning method include: Step 101: Based on the scan point data obtained by the robot at its current position in the first target scene, a semantic segmentation model is used to identify the first semantic label corresponding to each scan point in the scan point data, wherein the first semantic label is used to indicate the probability that the corresponding scan point is at least one type of obstacle; Step 102: Based on the scan point data and the first semantic label, the robot's pose on the first grid map is determined by matching the first grid map of the first target scene. The first grid map includes at least a second semantic label for indicating the occupied grid in the first grid map, wherein the second semantic label is used to indicate the probability that the corresponding grid is at least one obstacle type.

[0049] Here, the localization method can be executed by the robot, such as by the robot's internal processor. Alternatively, the localization method can be executed by the robot in conjunction with a server (such as a cloud server), for example, the robot acquires scanning data, and the server processes and analyzes the scanning data.

[0050] Scan point data can be point cloud data acquired by a robot scanning its surrounding environment using its onboard sensors (such as 2D LiDAR). Scan points can be reflections of the laser beam from a 2D LiDAR. These scan points constitute a point cloud.

[0051] In some embodiments, the scan point data includes: scan point reflection intensity data and scan point distance data.

[0052] Scan point reflection intensity data can be the intensity information of the reflected signal measured by sensors such as LiDAR when emitting a laser beam and receiving the reflected signal. Reflection intensity data is typically related to factors such as the surface material, color, roughness of the scanned object, and the incident angle of the laser beam. By utilizing scan point reflection intensity data, additional features beyond geometry can be provided to semantic segmentation models, helping them better identify and distinguish different types of obstacles, such as glass versus walls, or metal versus wood, even if they may exhibit similar geometric contours. Scan point distance data can be the distance information from each scan point measured by the sensor to the sensor itself. Scan point distance data directly reflects the spatial position and geometry of the scanned object. Using scan point distance data, a two-dimensional point cloud of the environment can be constructed, which can then be used to generate raster maps, extract geometric features, and perform geometric registration.

[0053] A semantic segmentation model can be a type of machine learning model. It is used to classify each scan point in the input scan point data and output the semantic category to which that scan point belongs.

[0054] In one possible implementation, the semantic segmentation model is trained to identify the probability that a scanned point belongs to different obstacle types. Here, obstacle types can include: chairs, doors, elevators, people, pillars, sofas, tables, shelves, trash cans, and / or walls, etc.

[0055] The first semantic label is the classification result output by the semantic segmentation model after processing each scan point in the scan point data. The first semantic label is used to indicate the probability that the corresponding scan point belongs to at least one obstacle type. For example, the first semantic label indicates the probability that the corresponding scan point is a wall, shelf, pedestrian, or other obstacle.

[0056] The first grid map is a map used to represent the first target scene in which the robot is located. The first grid map can consist of a series of grids, each grid representing an area in the first target scene.

[0057] Here, the first grid map may contain occupied / vacant information, and may also contain semantic information (i.e., second semantic labels) for each occupied grid.

[0058] The second semantic label is the semantic information associated with each occupied grid cell in the first grid map. The second semantic label indicates the probability that the corresponding grid cell belongs to at least one obstacle type. For example, the second semantic label could indicate the probability that the corresponding grid cell is a wall area, a shelf area, or another obstacle area. A grid cell can contain one or more scan points, meaning that the area covered by a grid cell can contain multiple scan points.

[0059] In one possible implementation, the first grid map is predetermined. For example, it can be determined by the robot performing a global scan of the first target scene.

[0060] In one possible implementation, the second semantic label of the occupying grid can be determined based on the third semantic label of each scan point in the grid. Here, the occupying grid can be a grid with scan points during the global scan. For example, when the robot performs a global scan of the first target scene, it can determine the third semantic label of each scan point through a semantic segmentation model, and then determine the second semantic label of the occupying grid based on the third semantic label of each scan point in the grid. Here, the semantic segmentation model can run on the robot or on the server, and the operating entity of the semantic segmentation model can be determined based on the computing power and load of the robot or the server.

[0061] Pose can be a combination of a robot's position and orientation in its environment. For example, position can be represented by two-dimensional coordinates (e.g., x, y), and orientation can be represented by an angle (e.g., yaw angle θ). Determining the robot's pose is fundamental to achieving autonomous navigation and task execution.

[0062] Specifically, the robot can use its onboard 2D LiDAR and other sensors to perform horizontal scanning of its surroundings, thereby acquiring a series of scan points. The plotted data of these scan points can include distance information for each point, as well as information on the reflectance intensity of each point. This distance data is used to construct the geometric contours of the environment.

[0063] The semantic segmentation model can be a pre-trained classifier, such as a model built based on machine learning algorithms like Support Vector Machine (SVM), Random Forest, and / or Single-Branch Semantic Stereo Network (S3Net). The semantic segmentation model can be trained to predict the obstacle category of a scanned point (i.e., a laser reflection point) based on its plotted data (such as distance and reflection intensity). The semantic segmentation model outputs a first semantic label for each scanned point, indicating the probability that the corresponding scanned point is at least one obstacle type. For example, the first semantic label can be a probability distribution vector representing the probability that the scanned point is a wall, shelf, pillar, or other obstacle.

[0064] The robot can match the first grid map of the first target scene with the scan point data of the current position and the first semantic label to determine the robot's pose on the first grid map.

[0065] In one possible implementation, the robot can first perform geometric matching independently, comparing the geometric information of the scanned point data with the geometric structure of the first raster map to obtain preliminary matching results. Then, the first semantic label is used as auxiliary information to filter or verify the preliminary matching results. For example, a simple consistency measure between the semantic labels of the scanned points and the semantic labels of the map raster can be calculated to exclude matches that are geometrically similar but semantically inconsistent.

[0066] In this embodiment, unless otherwise specified, both geometric comparison (matching) and semantic comparison (matching) can be performed only for the static environment. For example, the scanned data can be compared with the geometric and semantic information of static obstacles. This reduces interference from the dynamic environment (such as walking objects) on the localization process.

[0067] Thus, by semantically labeling the scanned point data using a semantic segmentation model and constructing a grid map containing semantic information, the robot can effectively utilize the geometric and semantic features of the environment for matching during localization. In complex scenarios with highly repetitive and symmetrical structures, this reduces multiple localization solutions and positional ambiguity, thereby improving the robot's pose determination accuracy and robustness in such environments.

[0068] In some embodiments, determining the robot's pose on the first grid map based on the matching result obtained by matching the scan point data and the first semantic label against the first grid map of the first target scene includes: Determine the first geometric feature quantity corresponding to the scan point data, and based on the comparison results between the first geometric feature quantity and the second geometric feature quantity corresponding to at least one observation coordinate in the first grid map, determine at least one candidate positioning coordinate from multiple observation points; At different poses on each candidate positioning coordinate, the scanning point corresponding to the scanning point data is projected onto the first grid map, and the joint likelihood field result of the scanning point corresponding to the scanning point data and the first grid map is confirmed respectively. The joint likelihood field result is determined based on the first geometric likelihood field result of the scanning point and the obstacles in the first grid map, and the first semantic consistency comparison result of comparing the first semantic label of the scanning point and the second semantic label of the grid corresponding to the scanning point. Based on the joint likelihood field results at different poses at each candidate positioning coordinate, the preliminary matching pose of each candidate positioning coordinate is determined. For all preliminary matching poses, geometric registration is performed based on the scanned point data and the obstacles in the first grid map to determine at least one registration matching pose; The robot's pose on the first grid map is determined based on the joint positioning confidence corresponding to at least one registration-matched pose.

[0069] Here, the first geometric feature represents the geometric properties of the current scan point data.

[0070] For example, the first geometric feature quantity can be a statistical value of the measured distance of the scan points (such as mean, variance, histogram distribution, etc.), or a key geometric shape feature extracted from the scan point data. The measured distance of the scan points can include the distance from the scan point to the laser source. The statistical value of the measured distance can include the statistical value of the measured distance of all scan points corresponding to the scan point data. The statistical value of the measured distance can be determined based on factors such as the laser's time of flight.

[0071] The preset observation coordinates can be the coordinates of multiple global observation feature points set in the first grid map. These preset coordinates are used for initial filtering against the robot's current position. Multiple global observation feature points can be set to improve the success rate of the initial matching.

[0072] For the preset observation coordinates in the first grid map, theoretical scan point data obtained by the robot scanning at these coordinates can be simulated, and the corresponding second geometric feature quantity can be extracted from them. The second geometric feature quantity can also be used to characterize the geometric properties of the theoretical scan point data. The type of the second geometric feature quantity can be consistent with the first geometric feature quantity for effective comparison. By comparing the first and second geometric feature quantities, the geometric similarity between the current scan point data and different observation coordinates in the map can be evaluated. Based on this comparison result, observation coordinates with higher similarity are selected as candidate positioning coordinates.

[0073] In one possible implementation, the difference between the first geometric feature and the second geometric feature can be determined, and all preset observation coordinates can be sorted in ascending order of the difference between each preset observation coordinate. The top N preset observation coordinates can be selected as potential candidate positioning coordinates for relocation in the current batch, thereby achieving pruning and filtering of the global search space.

[0074] In some embodiments, the first geometric feature quantity includes: a statistical value of the measured distance of all scan points in the scan point data; The second geometric feature quantity includes: the statistical value of the distance measured at the theoretical scanning point of the robot under the observation coordinates.

[0075] In one possible implementation, the statistical value may include an average value. The statistical value of the measured distances of all scan points in the scan point data may include the average of the measured distances of all scan points in the scan point data. The statistical value of the measured distances of the robot's theoretical scan points may include the average of the measured distances of the robot's theoretical scan points. The difference between the first geometric feature and the second geometric feature may include the absolute difference between the average of the measured distances of all scan points in the scan point data and the average of the measured distances of the robot's theoretical scan points.

[0076] After determining the candidate positioning coordinates, different poses corresponding to each candidate positioning can be determined based on predetermined rules. For example, using the candidate positioning coordinates as the initial position, several orientation angles to be searched are divided within the omnidirectional angle range of 0 to 2π according to a preset angle search step size, thereby combining several initial poses to be evaluated.

[0077] At different poses on each candidate localization coordinate, the robot projects the scan points corresponding to the scan point data onto the first grid map, and confirms the joint likelihood field result of the scan points corresponding to the scan point data and the first grid map respectively.

[0078] The joint likelihood field result can be a comprehensive evaluation based on two main aspects: First, the first geometric likelihood field result between the scanned point and obstacles in the first grid map. This first geometric likelihood field result quantifies the geometric fit between the scanned point and obstacles on the map; for example, the closer the scanned point is to an obstacle in the map, the higher its geometric likelihood field result. Second, the first semantic consistency comparison result, which compares the first semantic label of the scanned point with the second semantic label of the corresponding grid. This result evaluates the consistency between the semantic information of the scanned point and the semantic information of the corresponding area on the map. For example, if the first semantic label of the scanned point defines the probability of different obstacle types at that scanned point, and the grid is also defined by the second semantic label, the higher the similarity between the first and second semantic labels, the higher the score of the first semantic consistency comparison result. By fusing the first geometric likelihood field result and the first semantic consistency comparison result, for example through weighted summation, a more robust and accurate joint likelihood field result can be obtained, thus providing a more comprehensive evaluation of the matching quality of the current pose.

[0079] In some embodiments, the joint likelihood field result includes a weighted sum of the first geometric likelihood field result and the first semantic consistency comparison result.

[0080] Weighted summation involves assigning different weights to the first geometric likelihood field result and the first semantic consistency comparison result when calculating the joint likelihood field result. These weights reflect the importance or reliability of the geometric and semantic information, respectively, under the current localization scenario or specific conditions. By introducing weighted summation, the system can dynamically adjust the contribution of the two types of information to the final matching result based on the actual situation.

[0081] For example, in environments rich in geometric features, higher weights can be assigned to the geometric likelihood field results; while in environments with more discriminative semantic information, the weights of semantic consistency comparison results can be increased. These weights can be preset fixed values, or they can be parameters that are adaptively adjusted based on factors such as environmental characteristics, sensor quality, or positioning stage.

[0082] The joint likelihood field result can be expressed by expression (1): (1); in, Indicates the joint likelihood field result, The weights of the first geometric likelihood field result are represented. This represents the result of the first geometric likelihood field. The weights representing the results of the first semantic consistency comparison are: This indicates the result of the first semantic consistency comparison.

[0083] For each candidate positioning coordinate, possible poses can be traversed, and the joint likelihood field result under each pose can be calculated. The pose that makes the joint likelihood field result reach the maximum value or meet the preset threshold can be selected as the initial matching pose of the candidate positioning coordinate.

[0084] For all preliminary matched poses, geometric registration can be performed based on the scanned point data and obstacles in the first grid map to determine at least one registered matching pose. Geometric registration can be a pose optimization process. Using the preliminary matched pose as an initial value, geometric registration continuously adjusts the robot's pose through an iterative algorithm (e.g., Iterative Closest Point (ICP) algorithm or Generalized Iterative Closest Point (GICP) algorithm) to minimize the geometric error between the scanned point data and obstacles in the map. In this way, subtle deviations in the preliminary matched pose can be eliminated, resulting in a more accurate registered matching pose.

[0085] After determining the registration and matching poses, the robot's pose on the first grid map can be determined based on the joint localization reliability corresponding to each registration and matching pose. Joint localization reliability is a comprehensive indicator used to quantify the reliability of each registration and matching pose. It considers multiple factors, such as the geometric matching quality of the registration and matching poses, the consistency of distances between scanned points and map obstacles, and the consistency of semantic information. By comparing the joint localization reliability of different registration and matching poses, the pose with the highest joint localization reliability that meets the confidence condition is selected as the robot's final pose, thus obtaining an accurate localization result.

[0086] Thus, by comparing the geometric features of the scanned point data and the map, multiple candidate localization coordinates can be quickly and effectively filtered out, significantly reducing the search space for subsequent fine-grained matching and improving localization efficiency. When evaluating candidate poses, a joint likelihood field result combining geometric likelihood field and semantic consistency comparison is introduced. This allows pose evaluation to not only rely on pure geometric matching but also utilize semantic information to distinguish geometrically similar but semantically different regions, enhancing the robustness and accuracy of pose discrimination and effectively solving the problem of ambiguous localization in scenes with indistinct features or repetitive textures. Fine-grained optimization of the initial matched poses through geometric registration further improves localization accuracy. Based on joint localization confidence, the most reliable final pose is selected from multiple registered and matched poses, effectively reducing erroneous localization and enabling the robot to achieve high-precision and high-reliability localization in complex environments.

[0087] In some embodiments, the joint positioning confidence is determined based on at least one of the following: the second geometric likelihood field result corresponding to the registration-matched pose, the geometric distance evaluation result of the scan point, and the second semantic consistency comparison result; wherein: The second geometric likelihood field includes: the likelihood field of all scan points and obstacles in the first grid map when the robot projects the scan points corresponding to the scan point data onto the first grid map under the registration and matching pose; The evaluation results of the geometric distance of the scanning points include: the evaluation results of the average distance of all scanning points to the nearest obstacle when the robot projects the scanning points corresponding to the scanning point data onto the first grid map under the registration and matching pose; The second semantic consistency comparison result includes: the comparison result of the first semantic label of the scan point and the second semantic label of the grid corresponding to the scan point when the robot projects the scan point data onto the first grid map under the registration and matching pose.

[0088] Joint pose reliability is used to comprehensively and meticulously evaluate candidate poses after geometric registration by taking into account data from multiple dimensions, thereby improving the accuracy and reliability of the final pose determination.

[0089] The second geometric likelihood field is used to quantify the degree of geometrical fit between the scan point data currently acquired by the robot and the obstacles in the first grid map under the registration and matching pose.

[0090] For example, assuming the robot is in a certain registration-matching pose, all scan points from the scan point data are projected onto a first grid map. The distance between each projected scan point and the nearest obstacle in the first grid map is calculated, and these distances are converted into likelihood values ​​using a predefined likelihood function (e.g., a Gaussian function). The likelihood values ​​of all scan points can be accumulated or averaged to obtain a comprehensive second geometric likelihood field result. The higher this result value, the better the geometric matching between the scan points and map obstacles in that registration-matching pose.

[0091] The geometric distance evaluation result of the scan point is a direct quantification of the degree of geometric matching between the scan point data and the obstacles in the first grid map under the registration and matching pose.

[0092] In one possible implementation, the geometric distance of the scan point includes the mean score of the distance between the projected scan point and the nearest obstacle.

[0093] For example, assuming the robot is in a certain registration-matching pose, all scan points in the scan point data are projected onto a first grid map. For each projected scan point, its Euclidean distance to the nearest obstacle in the first grid map is calculated. Then, the average of all these distances is taken to obtain the geometric distance evaluation result of the scan point. The smaller the average value, the smaller the geometric deviation between the scan point and the obstacle on the map, i.e., the higher the geometric matching degree.

[0094] The second semantic consistency comparison result is used to evaluate the consistency between the semantic information carried by the scan point data and the semantic information of the corresponding position in the first grid map under a specific registration matching pose.

[0095] For example, assuming the robot is in a certain registration-matching pose, all scan points in the scan point data are projected onto a first grid map. For each projected scan point, its corresponding first semantic label is obtained, and the second semantic label corresponding to the grid in the first grid map into which it falls is obtained. Then, the consistency between these two semantic labels is evaluated by comparing them. The comparison method may include, but is not limited to: calculating the matching degree of the semantic labels (e.g., a matching degree of 1 if both indicate the same obstacle type, otherwise 0), and calculating the similarity of semantic probability distributions (e.g., using KL divergence or JS divergence). The semantic consistency comparison results of all scan points can be accumulated or averaged to obtain a comprehensive second semantic consistency comparison result. The higher this result value, the better the consistency between the semantics of the scan points and the semantics of the map under that registration-matching pose.

[0096] By introducing a second geometric likelihood field result and a geometric distance evaluation result for the scan points, this application can more finely evaluate the accuracy of registration and matching poses at the geometric level. By introducing a second semantic consistency comparison result, semantic information is integrated into the final pose confidence assessment. Even if two registration and matching poses are geometrically similar, if they differ semantically from their corresponding regions in the first grid map, they can be distinguished using the semantic consistency comparison result. For example, in environments with similar corridor structures but different functional areas (such as office areas and warehouse areas), geometric features may be highly similar, but semantic labels (such as "desk" and "shelf") can provide unique distinguishing information. By comparing the first semantic label of the scan point with the second semantic label of the corresponding grid, a more semantically reasonable pose can be identified. This deep fusion of geometric and semantic information improves the robustness and accuracy of localization.

[0097] In some embodiments described above in this application, the joint positioning confidence of the robot can be determined based on at least one of the second geometric likelihood field result, the geometric distance evaluation result of the scan point, and the second semantic consistency comparison result. However, in practical applications, how to effectively fuse these confidence information from different sources to obtain a more robust and accurate comprehensive confidence is a problem that needs to be solved. Simply selecting one of them or performing a simple linear combination may not fully utilize the complementarity of the information, and may even introduce unnecessary noise or bias, thereby affecting the accuracy and stability of the final positioning.

[0098] In some embodiments, the joint location confidence determination includes the geometric mean of the product of the second geometric likelihood field result, the geometric distance evaluation result of the scan point, and the second semantic consistency comparison result.

[0099] Using the geometric mean to combine these different confidence indices ensures that when any component is zero, the overall joint localization confidence is also zero. This effectively eliminates erroneous localization results when a key indicator (e.g., geometric fit range) is completely inconsistent. The geometric mean is insensitive to extreme values ​​and better reflects the multiplicative relationship between components, rather than a simple additive one. This means it can more fairly weigh the contributions of each indicator, reducing the excessive dominance of abnormally high or low values ​​of any single indicator in the final confidence score, thus providing a more robust and balanced comprehensive assessment.

[0100] In some embodiments, the step of performing geometric registration based on the scan point data and obstacles in the first grid map for all preliminary matched poses to determine at least one registration matching pose includes: All the preliminary matching poses are used as the iterative initial poses for geometric registration of the scanning points with obstacles in the first grid map. Registration is performed using the two-dimensional generalized iterative nearest point 2DGICP to determine at least one registration matching pose.

[0101] Two-dimensional generalized iterative nearest point (2D-GICP) is an algorithm for point cloud registration that iteratively finds the correspondence between two point sets and optimizes the transformation matrix to minimize the distance between them.

[0102] In each iteration, the 2D-GICP algorithm calculates the correspondence between each scan point in the scan point data under the current pose and the obstacle in the first grid map, and updates the robot's pose based on these correspondences until convergence or the maximum number of iterations is reached. After iteratively optimizing each initial matching pose through the 2D-GICP algorithm, one or more refined pose results are obtained, which are the registration matching poses.

[0103] Specifically, the 2D-GICP registration operator can perform point-to-area nonlinear least squares optimization iteration between the current frame scan data and the geometric obstacles of the first grid map to register and match the pose.

[0104] In one possible implementation, if the 2D-GICP algorithm fails to converge during iteration, the corresponding initial matching pose can be discarded.

[0105] In some embodiments, determining the robot's pose on the first grid map based on the joint positioning confidence corresponding to at least one registration-matched pose includes at least one of the following: In the first mode, the first registered and matched pose with a joint positioning confidence level higher than the confidence threshold is determined as the pose of the robot on the first grid map; In the second mode, the registration and matching pose with the highest joint positioning confidence and a joint positioning confidence higher than the confidence threshold is determined as the pose of the robot on the first grid map.

[0106] Here, the first mode can include a single-output mode, where the robot can sequentially determine the joint localization confidence of each registration-matching pose. The robot can make real-time judgments, and once it finds that the joint localization confidence of a certain registration-matching pose exceeds a preset confidence threshold, it immediately selects it as the robot's final pose and stops evaluating subsequent poses. The first mode can shorten the localization decision time and is suitable for application scenarios with high real-time requirements.

[0107] The second mode can include a multi-output mode, where the robot can make a judgment based on the joint positioning reliability of all registered and matched poses. The robot can calculate the joint positioning reliability of all registered and matched poses, then select the registered and matched poses whose joint positioning reliability is higher than a preset reliability threshold. Based on this, it further selects the pose with the highest joint positioning reliability from these reliable poses and uses it as the robot's final pose on the first grid map. The second mode ensures that the selected pose not only meets basic reliability requirements but is also the best among all reliable options, thus providing the robot with the most accurate and stable localization results.

[0108] In some embodiments, the method further includes: Acquire the robot's scan point data within the second target scene; A second grid map of the second target scene is established based on the scan points in the scan point data; Label the type of obstacle occupying each grid cell in the second grid map; The obstacle type of the grid where the scan point is located is used as the obstacle type of the scan point; The semantic segmentation model is trained based on the scan point data and the obstacle type at each scan point.

[0109] The above embodiments propose a method for robot localization based on semantic information. This method utilizes a semantic segmentation model to identify semantic labels of scan points and combines this with semantic labels from a grid map for pose matching. However, in its implementation, the performance of the semantic segmentation model directly affects the accuracy of localization, and training this model requires a large amount of scan point data with accurate semantic labels. A lack of effective semantic data preparation and model training mechanisms may lead to poor performance of the semantic segmentation model, thereby affecting the accuracy and robustness of localization.

[0110] The localization method proposed in this disclosure also includes the step of training a semantic segmentation model, such as... Figure 2 As shown, it includes: Step 201: Obtain the robot's scan point data within the second target scene; Step 202: Establish a second grid map of the second target scene based on the scan points in the scan point data; Step 203: Label the obstacle type for each grid cell in the second grid map; Step 204: Use the obstacle type of the grid where the scan point is located as the obstacle type of the scan point; Step 205: Train the semantic segmentation model based on the scan point data and the obstacle type of each scan point.

[0111] Here, the first target scene and the second target scene can be the same or different. The number of second target scenes can be increased to increase the training data.

[0112] The robot acquires scan point data within the second target scene to provide raw data for subsequent map building and model training. This scan point data is typically acquired by sensors such as LiDAR mounted on the robot while scanning the second target scene. The information contained in the scan point data is as described in the above embodiments and will not be repeated here.

[0113] A second grid map of the second target scene is built based on the scan points in the scan point data. This step converts the original scan point data into a structured grid map representation. The second grid map is typically a two-dimensional discrete grid, with each grid cell representing a region in the scene. The construction process can employ an occupancy grid map method, determining whether a grid is occupied (i.e., whether an obstacle exists) by statistically analyzing the density of scan points within each grid or the penetration of sensor rays. For example, a probability update rule can be used to estimate the occupancy probability of each grid.

[0114] After the second grid map is established, the occupied grid cells identified as obstacles need to be manually or semi-automatically labeled semantically. Obstacle types can include, but are not limited to, walls, tables, chairs, shelves, doors, windows, pillars, etc. The labeling process can be carried out using a graphical user interface tool, where operators assign the corresponding obstacle category to each occupied grid cell based on the visual information of the grid map and their understanding of the scene.

[0115] The obstacle type of the raster cell containing the scanned point is used as the obstacle type for the scanned point. This step is used to derive the semantic label at the scanned point level from the raster-level semantic annotation, thus providing point-wise semantic ground truth for the training of the semantic segmentation model. For each scanned point in the second raster map, the raster cell in which the scanned point is located is first determined. Then, the obstacle type of that raster cell, which has already been labeled, is directly assigned to the scanned point as its semantic label. This method simplifies the complexity of point-by-point manual annotation and ensures consistency between the semantic labels of the scanned points and the semantics of the map.

[0116] The training process involves using scan point data as input and the corresponding obstacle type at each scan point as a supervision signal (ground truth label). The parameters of the semantic segmentation model are adjusted by optimizing the loss function (such as cross-entropy loss), enabling the model to learn feature representations of different obstacle types and ultimately accurately predict the semantic label of the input scan point data. After training, the semantic segmentation model can be used to identify the first semantic label during localization.

[0117] The positioning method proposed in this disclosure also includes the step of establishing a first grid map, such as... Figure 3 As shown, it specifically includes: Step 301: Determine whether the first grid map is included based on the scan data obtained from the robot's global scan of the first target scene; Step 302: Use a semantic segmentation model to identify the third semantic label for each scan point in the first raster map; Step 303: Based on the third semantic label of the scan point in each occupied grid in the first grid map, determine the second semantic label corresponding to each occupied grid; the third semantic label is used to indicate the probability that the corresponding scan point is at least one type of obstacle.

[0118] Here, a robot can be used to create the first grid map of the first target scene.

[0119] The robot acquires comprehensive scan data covering the entire target scene through full movement and data collection. This data typically includes point cloud data, scan point reflection intensity data, scan point distance data, and the robot's pose information. Subsequently, simultaneous localization and mapping (SLAM) algorithms, such as a LiDAR-based SLAM system, are used to process this global scan data to construct a first grid map. This first grid map can be represented in two dimensions, with each grid cell storing probability information about its occupied, idle, or unknown state, providing a basic environmental model for subsequent localization processes.

[0120] A pre-trained semantic segmentation model is used to identify the third semantic label for each scan point in the first grid map. The semantic segmentation model is trained by acquiring scan point data in the second target scene, constructing the second grid map, labeling obstacle types, and so on, based on the method described above. The semantic segmentation model can perform point-by-point analysis on the original scan points constituting the first grid map, outputting a third semantic label for each scan point. This label indicates the probability that the corresponding scan point belongs to at least one obstacle type. For example, the label can be a probability vector representing the likelihood that the point belongs to different obstacle categories such as walls or tables.

[0121] Specifically, for each occupied grid cell in the first grid map, all scan points falling within that grid area can be identified. Then, the third semantic labels of these scan points are aggregated to infer the overall semantic attributes of the grid. For example, the second semantic label of the occupied grid can be determined by averaging, weighted averaging, or majority voting of the third semantic labels of all scan points within the grid. This aggregation process effectively elevates point-level semantic information to the grid level, ensuring that each occupied grid cell in the first grid map carries a clear semantic attribute.

[0122] Thus, a first grid map is constructed based on global scan data, and a pre-trained semantic segmentation model is used to semantically label the scan points that make up this map. This aggregates the semantic information at the scan point level to the grid level, thereby determining a second semantic label for each occupied grid cell in the first grid map. This ensures the accuracy and consistency of the semantic information in the map, enabling effective and reliable matching of the scan point data and its first semantic label acquired by the robot in real time with the second semantic label in the map during subsequent localization.

[0123] In some embodiments, determining the second semantic label corresponding to each occupied raster based on the third semantic label of the scan point in each occupied raster in the first raster map includes one of the following: A Bayesian filter is used to filter the third semantic label of all scan points in the occupied grid to obtain the second semantic label of the occupied grid. The third semantic labels of all scan points in the occupied grid are accumulated, and the second semantic label of the occupied grid is determined based on the accumulated value.

[0124] Bayesian filtering is a recursive state estimation algorithm based on probability theory, capable of effectively fusing uncertain information from multiple observation sources. The third semantic label of each scan point can be considered an observation of the true semantic state of the occupied grid. Bayesian filtering maintains a posterior probability distribution of possible semantic labels for the occupied grid and updates this distribution based on the third semantic label of each newly observed scan point. For example, if the third semantic label represents a probability vector of different obstacle types, Bayesian filtering iteratively updates the probability of the occupied grid belonging to each obstacle type based on these probability vectors and a pre-defined observation model, thus obtaining a second semantic label that integrates information from all scan points.

[0125] By employing Bayesian filtering, the third semantic labels of multiple scan points can be fused in a probabilistic manner, thereby obtaining a more accurate and robust second semantic label of the occupied grid, effectively reducing the impact of measurement errors of a single scan point on the grid semantic information.

[0126] The summation method determines the overall semantics of a grid by directly summing the semantic information of all scan points within the grid. For example, if the third semantic label represents the probability values ​​of different obstacle types, the probability values ​​of all scan points within the grid for a specific obstacle type can be summed. After summation, the magnitude of the summed value can be used to determine which obstacle type the grid is most likely to belong to, or the summed value can be normalized and used as the second semantic label for the grid.

[0127] By accumulating the third semantic labels of the scan points, semantic information can be aggregated simply and effectively, highlighting the tendency of the main obstacle types within the grid, thereby providing the grid with more representative second semantic labels.

[0128] In this embodiment of the disclosure, Figure 1 , Figure 2 and Figure 3 The steps shown can be performed individually or in combination. For example, Figure 1 The steps shown can be combined Figure 2 Implement the steps shown; Figure 1 The steps shown can be combined Figure 3 Implement the steps shown; Figure 1 The steps shown can be combined Figure 2 The steps shown are Figure 3The steps shown are to be implemented, and will not be repeated here.

[0129] A specific example is provided in conjunction with the above embodiments.

[0130] Here, as Figure 4 As shown, the following detailed explanations will be provided in the order of model training, first grid map establishment, and robot localization: Step 401: Construct a 2D LiDAR semantic segmentation dataset and train a 2D LiDAR semantic segmentation model based on the constructed dataset.

[0131] Specifically: 1) Select various indoor environments (second target scenarios) such as shopping malls, supermarkets, and warehouses to collect data (scanning data). The robot runs a Simultaneous Localization and Mapping (SLAM) algorithm library, such as Cartographer, in the Robot Operating System (ROS) environment to build the map. Data collection is carried out by keyboard remote control. The data includes 2D LiDAR scan data (distance and intensity data) and scene map.

[0132] 2) Common objects in indoor environments were labeled: chairs, doors, elevators, people, pillars, sofas, tables, shelves, trash cans, and walls. These object categories were further subdivided based on material, into three independent subsets for training (70% of the data), validation during training (10%), and final testing (20%). To avoid environmental bias across the entire dataset, the data was split from each scene using the above ratios, and then the scene data were combined using a joint harvester to create the final dataset.

[0133] 3) Data annotation uses an open-source semi-automatic semantic annotation framework to annotate the point cloud, and the LabelMe tool is used to annotate the generated raster map (second raster map). Then, the LiDAR scan data is matched with this labeled map to annotate each scan point.

[0134] 4) The semantic segmentation model can adopt the open-source S3-Net network architecture, using a variational autoencoder as the backbone network. The network input data consists of 2D LiDAR distance and intensity data. The LiDAR intensity data depends on the distance, the material properties of the object, and the angle at which the LiDAR strikes the material surface.

[0135] Step 402: Simultaneously run the trained semantic segmentation model and 2D SLAM algorithm (such as gmaping or cartographer) to build a semantic map of the scene.

[0136] Specifically: 1) The raw LaserScan data (scanning data from a global scan of the first target scene) is input into a 2D SLAM algorithm (such as Gmapping or Cartographer). The algorithm uses radar point clouds between consecutive frames for scan matching and combines odometry data for front-end registration and back-end map optimization. Its core task is to calculate the robot's current global pose in real time and maintain a dynamically updated geometric occupancy grid map (the first grid map).

[0137] 2) The same frame of LaserScan data (usually converted into a one-dimensional / two-dimensional sequence tensor or distance image representation, i.e., scan data) is input into the trained semantic segmentation model. The semantic segmentation model classifies each data point in the radar point cloud of that frame point by point and outputs a semantic label sequence that corresponds one-to-one with the original point cloud (e.g., 1=wall, 2=human leg, 3=sofa, etc.) (i.e., the third semantic label).

[0138] 3) The map is expanded into a multi-layer grid map, where each grid not only records the occupancy probability but also maintains a probability vector for the object category. Bayesian filtering or a counting accumulation mechanism is used to update the corresponding grids with the projected global semantic points. For example, if a grid at a certain global coordinate is repeatedly segmented into a "fixed wall" by radar points, the confidence level of that grid in the "wall" channel continuously increases. Moving targets such as humans are removed as their confidence level increases, ultimately resulting in a 2D semantic grid map that combines geometric boundaries and attribute labels (secondary semantic labels).

[0139] Step 403: Based on the constructed semantic grid map and real-time scanning data, relocalization is performed to obtain the robot's coordinates.

[0140] Specifically: 1) Real-time data acquisition and parallel extraction of geometric-semantic dual-track features.

[0141] The 2D lidar controlling the robot collects a single frame of raw laser point cloud data of the current environment, which is recorded as the current frame laser data.

[0142] Geometric feature extraction module: Input the current frame laser data into the geometric feature extraction module to calculate the global geometric feature quantity (first geometric feature quantity) of the current frame laser data. The global geometric feature quantity includes the average measurement distance of the laser point cloud of the current frame.

[0143] Semantic feature extraction track: Simultaneously, the current frame laser data is input into a pre-trained semantic segmentation model. The semantic segmentation model performs point-by-point classification on each laser data point in the current frame laser point cloud, and outputs a local semantic label sequence (first semantic label) that corresponds one-to-one with the current frame laser data, such as... (Labels include: walls, pillars, sofas, pedestrians, etc.)

[0144] 2) Loading of multi-channel preset semantic maps and construction of offline distance fields.

[0145] A pre-loaded two-dimensional prior semantic map (first grid map) is used, which includes geometric occupancy information channels and multi-layer object attribute semantic channels (second semantic labels). The grids corresponding to static environment labels in the two-dimensional prior semantic map are extracted, and distance transformations are performed independently for each static environment label. Multiple independent global distance reciprocal tables for static semantic categories are calculated to represent the nearest physical distance from any grid coordinate in the map to a specific static semantic obstacle.

[0146] 3) Coarse screening of candidate regions for large-scale relocation based on geometric features.

[0147] A number of global observation feature points (observation points) pre-sampled offline in the free area of ​​the prior semantic map are obtained. Each global observation feature point contains its two-dimensional coordinates in the map and the corresponding theoretical average radar distance feature sequence. Using a multi-threaded parallel computing architecture, all the global observation feature points are traversed, and the absolute difference between the average measured distance of the current frame laser data and the theoretical average radar distance feature sequence of each global observation feature point is calculated. All global observation feature points are sorted in ascending order of the absolute difference, and the top N (N is a positive number greater than or equal to 1) global observation feature points are selected as potential candidate regions for relocation in the current batch, thereby realizing pruning and filtering of the global search space.

[0148] 4) Omnidirectional pose search based on multi-channel semantic likelihood field.

[0149] For each potential candidate region selected in 3), the two-dimensional coordinates (candidate positioning coordinates) are used as the initial position. Within the omnidirectional angle range of 0~2π, several orientation angles to be searched are divided according to the preset angle search step size, thereby forming several initial poses to be evaluated. Using a multi-threaded parallel computing architecture, the current frame laser data and the corresponding local semantic label sequence are projected into the global map coordinate system according to each initial pose to be evaluated. The geometric-semantic joint likelihood field score (joint likelihood field result) under each initial pose to be evaluated is calculated: First, the geometric matching score (first geometric likelihood field result) is calculated by querying the corresponding global distance inverse table. Then, the semantic matching score (first semantic consistency comparison result) is calculated by comparing the degree of consistency between the projected laser point semantic label and the corresponding grid label in the two-dimensional prior semantic map. The geometric matching score and the semantic matching score are weighted and summed to obtain the joint likelihood field score. The calculation expression can be represented by expression (1).

[0150] The initial pose to be evaluated with the highest joint likelihood field score in each potential candidate region is selected as the best coarse matching pose (candidate localization coordinates) for that region.

[0151] 5) Local geometric fine registration based on generalized iterative nearest point.

[0152] Extract the top M (M is a positive number greater than or equal to 1) best coarse matching poses generated from each candidate region in step 4) as initial values ​​for iteration, and feed them into the two-dimensional generalized iterative nearest point (2D-GICP) registration operator; the 2D-GICP registration operator performs point-to-area nonlinear least squares optimization iteration between the current frame laser data and the geometric obstacles of the two-dimensional prior semantic map to calculate the global candidate poses (registration matching poses) after fine registration.

[0153] 6) Integrating semantic consistency in joint confidence assessment and multi-indicator deduplication.

[0154] For each finely registered global candidate pose, the laser point cloud of the current frame is globally projected, and three core verification metrics are calculated: the geometric likelihood field base score (second geometric likelihood field result), the mean distance score from the laser point to the nearest obstacle (geometric distance evaluation result of the scan point), and the static semantic consistency score (second semantic consistency comparison result). The static semantic consistency score is determined by statistically analyzing the proportion of laser points identified as static environment labels after projection that are completely consistent with the corresponding map grid semantic labels. The geometric likelihood field base score, mean distance score, and static semantic consistency score are multiplied as a product term, and the final joint localization reliability is calculated using the geometric mean formula. The legality of each global candidate pose is checked, and poses that are out of bounds, fall in unknown areas, or are less than a preset safety threshold in physical distance from obstacles are eliminated. For legal poses, nearest neighbor clustering suppression based on spatial distance and angle is performed to eliminate redundant relocalization candidate points.

[0155] 7) Adaptive early stopping determination and optimal global pose output.

[0156] The system determines whether the current relocalization task is in single-candidate output mode: If it is in single-candidate output mode, and the highest joint positioning confidence calculated in the current batch is greater than the preset high confidence upper limit threshold, the calculation of subsequent unexecuted batches is immediately interrupted, and the global candidate pose (registration matching pose) corresponding to the highest confidence is directly output as the optimal global pose to the robot's navigation system to complete the relocalization; if the early stopping condition is not met or it is in multi-candidate output mode, all batches are traversed and executed, and the pose with the highest joint positioning confidence and greater than the preset relocalization bottom line threshold is finally output as the final optimal global pose; if the confidence of all candidate poses is lower than the bottom line threshold, the relocalization is determined to have failed.

[0157] This disclosure also provides a positioning device, the device comprising: a processing module, the processing module being used for: Based on the scan point data obtained by the robot at its current position in the first target scene, a semantic segmentation model is used to identify the first semantic label corresponding to each scan point in the scan point data, wherein the first semantic label is used to indicate the probability that the corresponding scan point is at least one type of obstacle; Based on the scan point data and the first semantic label, the matching result obtained by matching the first grid map of the first target scene determines the pose of the robot on the first grid map, wherein the first grid map includes at least: a second semantic label for indicating the occupied grid in the first grid map, wherein the second semantic label is used to indicate the probability that the corresponding grid is at least one obstacle type.

[0158] In some embodiments, the processing module is specifically used for: Determine the first geometric feature quantity corresponding to the scan point data, and based on the comparison results between the first geometric feature quantity and the second geometric feature quantity corresponding to at least one observation coordinate in the first grid map, determine at least one candidate positioning coordinate from multiple observation points; At different poses on each candidate positioning coordinate, the scanning point corresponding to the scanning point data is projected onto the first grid map, and the joint likelihood field result of the scanning point corresponding to the scanning point data and the first grid map is confirmed respectively. The joint likelihood field result is determined based on the first geometric likelihood field result of the scanning point and the obstacles in the first grid map, and the first semantic consistency comparison result of comparing the first semantic label of the scanning point and the second semantic label of the grid corresponding to the scanning point. Based on the joint likelihood field results at different poses at each candidate positioning coordinate, the preliminary matching pose of each candidate positioning coordinate is determined. For all preliminary matching poses, geometric registration is performed based on the scanned point data and the obstacles in the first grid map to determine at least one registration matching pose; The robot's pose on the first grid map is determined based on the joint positioning confidence corresponding to at least one registration-matched pose.

[0159] In some embodiments, the joint likelihood field result includes a weighted sum of the first geometric likelihood field result and the first semantic consistency comparison result.

[0160] In some embodiments, the joint positioning confidence is determined based on at least one of the following: the second geometric likelihood field result corresponding to the registration-matched pose, the geometric distance evaluation result of the scan point, and the second semantic consistency comparison result; wherein: The second geometric likelihood field includes: the likelihood field of all scan points and obstacles in the first grid map when the robot projects the scan points corresponding to the scan point data onto the first grid map under the registration and matching pose; The evaluation results of the geometric distance of the scanning points include: the evaluation results of the average distance of all scanning points to the nearest obstacle when the robot projects the scanning points corresponding to the scanning point data onto the first grid map under the registration and matching pose; The second semantic consistency comparison result includes: the comparison result of the first semantic label of the scan point and the second semantic label of the grid corresponding to the scan point when the robot projects the scan point data onto the first grid map under the registration and matching pose.

[0161] In some embodiments, the joint location confidence determination includes the geometric mean of the product of the second geometric likelihood field result, the geometric distance evaluation result of the scan point, and the second semantic consistency comparison result.

[0162] In some embodiments, the processing module is specifically used for: All the preliminary matching poses are used as the iterative initial poses for geometric registration of the scanning points with obstacles in the first grid map. Registration is performed using the two-dimensional generalized iterative nearest point 2DGICP to determine at least one registration matching pose.

[0163] In some embodiments, the processing module is specifically used for at least one of the following: In the first mode, the first registered and matched pose with a joint positioning confidence level higher than the confidence threshold is determined as the pose of the robot on the first grid map; In the second mode, the registration and matching pose with the highest joint positioning confidence and a joint positioning confidence higher than the confidence threshold is determined as the pose of the robot on the first grid map.

[0164] In some embodiments, the first geometric feature quantity includes: a statistical value of the measured distance of all scan points in the scan point data; The second geometric feature quantity includes: the statistical value of the distance measured at the theoretical scanning point of the robot under the observation coordinates.

[0165] In some embodiments, the processing module is further configured to: Acquire the robot's scan point data within the second target scene; A second grid map of the second target scene is established based on the scan points in the scan point data; Label the type of obstacle occupying each grid cell in the second grid map; The obstacle type of the grid where the scan point is located is used as the obstacle type of the scan point; The semantic segmentation model is trained based on the scan point data and the obstacle type at each scan point.

[0166] In some embodiments, the processing module is further configured to: Based on the scan data obtained from the robot's global scan of the first target scene, it is determined that the first grid map is included. A semantic segmentation model is used to identify the third semantic label for each scan point in the first raster map; Based on the third semantic label of the scan point in each occupied grid in the first grid map, a second semantic label is determined for each occupied grid; the third semantic label is used to indicate the probability that the corresponding scan point is at least one type of obstacle.

[0167] In some embodiments, the processing module is specifically used for one of the following: A Bayesian filter is used to filter the third semantic label of all scan points in the occupied grid to obtain the second semantic label of the occupied grid. The third semantic labels of all scan points in the occupied grid are accumulated, and the second semantic label of the occupied grid is determined based on the accumulated value.

[0168] In some embodiments, the scan point data includes: scan point reflection intensity data and scan point distance data.

[0169] This disclosure also proposes an electronic device comprising: one or more processors; wherein the processors are configured to invoke instructions to cause the electronic device to perform the positioning method described in the first aspect.

[0170] This disclosure also proposes a storage medium storing instructions that, when executed on an electronic device, cause the electronic device to perform the positioning method described in the first aspect.

[0171] It should be understood that the division of units or modules in the above device is only a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, the units or modules in the device can be implemented by a processor calling software: for example, the device includes a processor connected to a memory containing instructions. The processor calls the instructions stored in the memory to implement any of the above methods or to implement the functions of the units or modules in the above device. The processor can be, for example, a general-purpose processor, such as a Central Processing Unit (CPU) or a microprocessor, and the memory can be internal or external to the device. Alternatively, the units or modules in the device can be implemented in the form of hardware circuits. The functionality of some or all of the units or modules can be achieved through the design of these hardware circuits, which can be understood as one or more processors. For example, in one implementation, the hardware circuit is an application-specific integrated circuit (ASIC), and the functionality of some or all of the units or modules is achieved through the design of the logical relationships between the components within the circuit. In another implementation, the hardware circuit can be implemented using a programmable logic device (PLD), such as a field-programmable gate array (FPGA), which can include a large number of logic gates. The connection relationships between the logic gates are configured through configuration files, thereby achieving the functionality of some or all of the units or modules. All units or modules of the above device can be implemented entirely through processor-called software, entirely through hardware circuits, or partially through processor-called software with the remaining parts implemented through hardware circuits.

[0172] In this disclosure, the processor is a circuit with signal processing capabilities. In one implementation, the processor can be a circuit with instruction read and execute capabilities, such as a Central Processing Unit (CPU), a microprocessor, a graphics processing unit (GPU) (which can be understood as a type of microprocessor), or a digital signal processor (DSP). In another implementation, the processor can implement certain functions through the logical relationships of hardware circuits. The logical relationships of the aforementioned hardware circuits are fixed or reconfigurable. For example, the processor is a hardware circuit implemented using an application-specific integrated circuit (ASIC) or a programmable logic device (PLD), such as an FPGA. In a reconfigurable hardware circuit, the process of the processor loading a configuration document and configuring the hardware circuit can be understood as the process of the processor loading instructions to implement the functions of some or all of the above units or modules. In addition, it can also be a hardware circuit designed for artificial intelligence, which can be understood as an ASIC, such as a Neural Network Processing Unit (NPU), a Tensor Processing Unit (TPU), a Deep Learning Processing Unit (DPU), etc.

[0173] Figure 5 This is a schematic diagram of the structure of the electronic device 9100 provided in this embodiment. The electronic device 9100 can be a network device, a terminal (e.g., a user equipment), a chip, chip system, or processor that supports the network device in implementing any of the above methods, or a chip, chip system, or processor that supports the terminal in implementing any of the above information transmission methods. The electronic device 9100 can be used to implement the positioning method described in the above method embodiments; please refer to the description in the above method embodiments for details.

[0174] like Figure 5 As shown, the electronic device 9100 includes one or more processors 9101. The processor 9101 can be a general-purpose processor or a special-purpose processor, etc. The processor 9101 is used to invoke instructions to cause the electronic device 9100 to execute any of the above positioning methods.

[0175] In some embodiments, the electronic device 9100 further includes one or more memories 9102 for storing instructions. Optionally, all or part of the memories 9102 may also be located outside the electronic device 9100.

[0176] In some embodiments, the electronic device 9100 further includes one or more transceivers 9103. When the electronic device 9100 includes one or more transceivers 9103, the steps of sending, receiving and / or acquiring in the above method are performed by the transceivers 9103, and the other steps are performed by the processor 9101.

[0177] In some embodiments, the acquisition steps in the above method can also be executed by the processor 9101, for example, acquiring information from the memory 9102.

[0178] In some embodiments, a transceiver may include a receiver and a transmitter, which may be separate or integrated. Optionally, the terms transceiver, transceiver unit, transceiver, transceiver circuit, etc., may be used interchangeably; the terms transmitter, transmitting unit, transmitter, transmitting circuit, etc., may be used interchangeably; and the terms receiver, receiving unit, receiver, receiving circuit, etc., may be used interchangeably.

[0179] Optionally, the electronic device 9100 further includes one or more interface circuits 9104 connected to the memory 9102. The interface circuits 9104 can be used to receive signals from the memory 9102 or other devices, and can be used to send signals to the memory 9102 or other devices. For example, the interface circuits 9104 can read instructions stored in the memory 9102 and send the instructions to the processor 9101.

[0180] The electronic device 9100 described in the above embodiments may be a network device or a terminal, but the scope of the electronic device 9100 described in this disclosure is not limited thereto, and the structure of the electronic device 9100 may vary. Figure 5 The limitations. Electronic devices can be standalone devices or part of a larger device. For example, the electronic devices can be: (1) standalone integrated circuits (ICs), or chips, or chip systems or subsystems; (2) a collection of one or more ICs, optionally including storage components for storing data and programs; (3) ASICs, such as modems; (4) modules that can be embedded in other devices; (5) receivers, terminal devices, smart terminal devices, cellular phones, wireless devices, handheld devices, mobile units, vehicle-mounted devices, network devices, cloud devices, artificial intelligence devices, etc.; (6) others, etc.

[0181] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program commands. The aforementioned program can be stored in a storage medium, including various media capable of storing program code such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks or optical disks.

[0182] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several commands to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0183] It should be understood that the above embodiments are exemplary and are not intended to encompass all possible implementations included in the claims. Various modifications and changes can be made to the above embodiments without departing from the scope of this disclosure. Similarly, the various technical features of the above embodiments can be arbitrarily combined to form other embodiments of the present invention that may not be explicitly described. Therefore, the above embodiments only illustrate several implementations of the present invention and do not limit the scope of protection of this patent.

Claims

1. A positioning method, characterized in that, The method includes: Based on the scan point data obtained by the robot at its current position in the first target scene, a semantic segmentation model is used to identify the first semantic label corresponding to each scan point in the scan point data, wherein the first semantic label is used to indicate the probability that the corresponding scan point is at least one type of obstacle; Based on the scan point data and the first semantic label, the matching result obtained by matching the first grid map of the first target scene determines the pose of the robot on the first grid map, wherein the first grid map includes at least: a second semantic label for indicating the occupied grid in the first grid map, wherein the second semantic label is used to indicate the probability that the corresponding grid is at least one obstacle type.

2. The positioning method according to claim 1, characterized in that, Based on the scanned point data and the first semantic label, the robot's pose on the first grid map of the first target scene is determined by matching the matching result, including: Determine the first geometric feature quantity corresponding to the scan point data, and based on the comparison results between the first geometric feature quantity and the second geometric feature quantity corresponding to at least one observation coordinate in the first grid map, determine at least one candidate positioning coordinate from multiple observation points; At different poses on each candidate positioning coordinate, the scanning point corresponding to the scanning point data is projected onto the first grid map, and the joint likelihood field result of the scanning point corresponding to the scanning point data and the first grid map is confirmed respectively. The joint likelihood field result is determined based on the first geometric likelihood field result of the scanning point and the obstacle in the first grid map, and the first semantic consistency comparison result of comparing the first semantic label of the scanning point and the second semantic label of the grid corresponding to the scanning point. Based on the joint likelihood field results at different poses at each candidate positioning coordinate, the preliminary matching pose of each candidate positioning coordinate is determined. For all preliminary matching poses, geometric registration is performed based on the scanned point data and the obstacles in the first grid map to determine at least one registration matching pose; The robot's pose on the first grid map is determined based on the joint positioning confidence corresponding to at least one registration-matched pose.

3. The positioning method according to claim 2, characterized in that, The joint positioning confidence is determined based on at least one of the following: the second geometric likelihood field result corresponding to the registration and matching pose, the geometric distance evaluation result of the scan point, and the comparison result of the second semantic consistency; wherein: The second geometric likelihood field includes: the likelihood field of all scan points and obstacles in the first grid map when the robot projects the scan points corresponding to the scan point data onto the first grid map under the registration and matching pose; The evaluation results of the geometric distance of the scanning points include: the evaluation results of the average distance of all scanning points to the nearest obstacle when the robot projects the scanning points corresponding to the scanning point data onto the first grid map under the registration and matching pose; The second semantic consistency comparison result includes: the comparison result of the first semantic label of the scan point and the second semantic label of the grid corresponding to the scan point when the robot projects the scan point data onto the first grid map under the registration and matching pose.

4. The positioning method according to claim 2, characterized in that, Determining the robot's pose on the first grid map based on the joint positioning confidence corresponding to at least one registration-matched pose includes at least one of the following: In the first mode, the first registered and matched pose with a joint positioning confidence level higher than the confidence threshold is determined as the pose of the robot on the first grid map; In the second mode, the registration and matching pose with the highest joint positioning confidence and a joint positioning confidence higher than the confidence threshold is determined as the pose of the robot on the first grid map.

5. The positioning method according to any one of claims 1 to 4, characterized in that, The method further includes: Acquire the robot's scan point data within the second target scene; A second grid map of the second target scene is established based on the scan points in the scan point data; Label the type of obstacle occupying each grid cell in the second grid map; The obstacle type of the grid where the scan point is located is used as the obstacle type of the scan point; The semantic segmentation model is trained based on the scan point data and the obstacle type at each scan point.

6. The positioning method according to claim 5, characterized in that, The method further includes: Based on the scan data obtained from the robot's global scan of the first target scene, it is determined that the first grid map is included. A semantic segmentation model is used to identify the third semantic label for each scan point in the first raster map; Based on the third semantic label of the scan point in each occupied grid in the first grid map, a second semantic label is determined for each occupied grid; the third semantic label is used to indicate the probability that the corresponding scan point is at least one type of obstacle.

7. The positioning method according to claim 6, characterized in that, The step of determining the second semantic label corresponding to each occupied grid cell based on the third semantic label of the scan point in each occupied grid cell in the first grid map includes the following: A Bayesian filter is used to filter the third semantic label of all scan points in the occupied grid to obtain the second semantic label of the occupied grid. The third semantic labels of all scan points in the occupied grid are accumulated, and the second semantic label of the occupied grid is determined based on the accumulated value.

8. A positioning device, characterized in that, The device includes a processing module, the processing module being used for: Based on the scan point data obtained by the robot at its current position in the first target scene, a semantic segmentation model is used to identify the first semantic label corresponding to each scan point in the scan point data, wherein the first semantic label is used to indicate the probability that the corresponding scan point is at least one type of obstacle; Based on the scan point data and the first semantic label, the matching result obtained by matching the first grid map of the first target scene determines the pose of the robot on the first grid map, wherein the first grid map includes at least: a second semantic label for indicating the occupied grid in the first grid map, wherein the second semantic label is used to indicate the probability that the corresponding grid is at least one obstacle type.

9. An electronic device, characterized in that, The electronic device includes: One or more processors; The processor is used to invoke instructions to cause the electronic device to perform the positioning method according to any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores instructions that, when executed on an electronic device, cause the electronic device to perform the positioning method according to any one of claims 1 to 7.

Images