Plane 2D point cloud map updating method based on dynamic value evaluation

Abstract

The invention discloses a plane 2D point cloud map updating method based on dynamic value evaluation, which belongs to the technical field of intelligent navigation, and comprises the following steps: acquiring sensor data of a current frame, and matching the sensor data with a currently maintained global map to obtain a matching result; based on the matching result, updating and evaluating the target unit in the map, and fusing an instant cleaning criterion based on space consistency and a long-term value criterion based on historical state information; according to a result of the update evaluation, determining an update action on the target unit, the update action at least comprising a deletion operation; an update action is performed to update the global map. According to the method, the accurate map updating decision is realized through intelligent fusion of instant and long-term double criteria, and the global scoring and fault-tolerant control mechanism is combined, so that the long-term consistency and high precision of the map are ensured, the operation efficiency and stability of the system are remarkably improved, and reliable and adaptive navigation support is provided for the robot in a dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent navigation technology, and in particular to a method for updating a planar 2D point cloud map based on dynamic value assessment. Background Technology

[0002] In robotics, 2D SLAM (Simultaneous Localization and Mapping) based on LiDAR is crucial for achieving autonomous robot navigation. Traditional SLAM algorithms (such as Gmapping and Cartographer) typically use the constructed environmental map as a static map for subsequent localization and path planning after initial mapping. However, real-world environments are often dynamically changing, and static maps have significant drawbacks in this scenario: First, they have poor environmental adaptability. When the map undergoes persistent changes (such as furniture displacement or the addition of obstacles), the static map cannot reflect the latest environment, leading to localization drift or planning failure. Second, there is the "ghosting" problem, where dynamic objects that briefly appear during mapping or operation (such as pedestrians or temporary stacks of objects) are scanned and fixed in the map, creating false occupancy information that does not match reality, thus interfering with the robot's normal operation in the long term.

[0003] The aforementioned problems are particularly prominent in AGV positioning applications. To address environmental changes, the industry has proposed various map maintenance and update solutions, but all have certain limitations, mainly including:

[0004] 1. Manual Marking and Update Solution: Mapping is completed once during the AGV deployment phase. Subsequent monitoring relies on remote monitoring by maintenance personnel. When environmental changes are detected, "dirty areas" are manually selected within the changed region, triggering a specific re-optimization process (such as Cartographer's "constraint-only" re-optimization). Alternatively, the corresponding raster values ​​on the old map can be directly replaced with new scan data, or occupied rasters meeting preset ranges and thresholds can be forcibly reset to an unknown state. This solution is highly dependent on manual intervention, inefficient, and struggles to respond to changes in real time, lacking sufficient intelligence.

[0005] 2. Dynamic Point Cloud Removal Scheme: This scheme attempts to distinguish between dynamic and static objects before data fusion. By comparing current laser scan data with existing maps or historical scan data, it uses methods such as geometric consistency and temporal analysis to identify and remove point clouds belonging to dynamic objects (such as moving pedestrians and vehicles), incorporating only static point clouds into the map update. However, this scheme is highly dependent on the accuracy of dynamic point identification. In complex AGV operation scenarios, sensor noise, slight environmental disturbances, and the ambiguity of the edges between dynamic objects and static structures can easily lead to misjudgments, potentially incorrectly removing point clouds that belong to static map structures. This erodes the effectiveness of the map and ultimately impairs positioning accuracy and system robustness.

[0006] 3. Incremental Map Update Scheme: This scheme employs a probabilistic fusion approach. For example, in a probabilistic raster map, based on Bayesian update rules, new data from each scan is fused with the existing map with fixed weights. The probability value of repeatedly observed and occupied rasters increases, while the probability value of observed empty rasters decreases. This scheme aims to allow the map to slowly adapt to long-term environmental changes through continuous observation, and to allow the "ghosting" effects of transient dynamic objects to be naturally "erased" by subsequent observations. However, its core flaw lies in the lack of intelligent judgment in the update mechanism. It treats all observation data "equally" using fixed update weights and rules, failing to effectively distinguish between "real, persistent environmental structural changes" and "temporary, dynamic object interference." This results in a sluggish response and delayed updates when facing real environmental changes, and in areas where dynamic objects frequently appear, it may introduce noise due to over-fusion.

[0007] In summary, existing map maintenance solutions primarily rely on fixed thresholds, simple rules, or manual judgment, resulting in limited intelligence and an inability to adaptively and accurately identify and handle environmental changes of varying natures. This leads to AGVs frequently encountering problems such as outdated maps, positioning failures, or unreasonable path planning during actual operation, hindering their long-term reliable operation in dynamic scenarios. Therefore, there is an urgent need for a technical solution capable of intelligently sensing the nature of environmental changes and updating maps accordingly. Summary of the Invention

[0008] The purpose of this invention is to provide a planar 2D point cloud map update method based on dynamic value assessment, so as to solve the problem of insufficient long-term operational reliability and dynamic adaptability caused by the lack of intelligent environmental perception and adaptive capabilities in traditional AGV map maintenance methods.

[0009] To achieve the above objectives, this application adopts the following technical solution:

[0010] This application discloses a planar 2D point cloud map update method based on dynamic value assessment, which includes the following steps:

[0011] Obtain sensor data for the current frame and match the sensor data with the currently maintained global map to obtain the matching result;

[0012] Based on the matching results, the target units in the map are updated and evaluated, wherein the update evaluation integrates the immediate cleanup criterion based on spatial consistency and the long-term value criterion based on historical state information.

[0013] Based on the results of the update evaluation, an update action for the target unit is determined, and the update action includes at least a deletion operation;

[0014] Perform the update action to update the global map.

[0015] Preferably, the instant cleanup criterion based on spatial consistency includes:

[0016] Based on the matching results, a matching score is calculated between the target unit and the matched sensor data.

[0017] When the matching score is lower than the first threshold, a continuity metric is calculated based on the matching status of each adjacent unit of the target unit.

[0018] Based on the continuity metric, the confidence level of the target unit is attenuated;

[0019] If the confidence level after decay is lower than the second threshold, a decision tendency to delete the target unit is generated.

[0020] Preferably, the calculation of the continuity measure includes:

[0021] Based on the matching results, determine whether each adjacent unit of the target unit matches the current sensor data, and count the number of matching adjacent units;

[0022] Divide the number by the total number of adjacent units to obtain the basic continuity score;

[0023] The basic continuity score is multiplied by a boundary protection coefficient based on the target cell location to obtain the final continuity metric.

[0024] Specifically, for target units located at the edge or corner of the map, the boundary protection coefficient is greater than 1; for target units at other locations, the boundary protection coefficient is equal to 1.

[0025] Preferably, the long-term value criterion based on historical state information is implemented through a reinforcement learning model, which is configured to perform the following steps:

[0026] The update decision process for the target unit is modeled as a Markov decision process, wherein the state includes at least the current matching degree, current confidence degree, historical update count, and historical variance of the observations of the target unit.

[0027] Based on the Bellman optimal equation, the state value function value corresponding to different update actions performed on the target unit is calculated, which serves as the basis for evaluating its long-term expected value.

[0028] Preferably, the reward function in the reinforcement learning model is set in the following manner:

[0029] When the update action is deletion, the output instant reward value is negatively correlated with the current confidence level of the target unit, positively correlated with the current mismatch level, and positively correlated with the historical observation variance.

[0030] Preferably, the update assessment step is implemented through a decision fusion mechanism, which integrates the decision tendency generated by the immediate cleanup criteria with the expected value assessed by the long-term value criteria to determine the final update action.

[0031] Preferably, the method further includes a map version replacement decision step:

[0032] Maintain and update multiple map versions, including the currently used map and at least one backup map;

[0033] A multi-dimensional scoring function is used to comprehensively score the multiple map versions, and the multi-dimensional scoring function is based at least on the map. Figure 1 Construction of consistency score, positioning accuracy score and coverage integrity score;

[0034] Based on the scoring results, it is determined whether to replace the currently used map with the backup map that has the highest overall score; wherein, the determination introduces a lag mechanism, and the replacement operation is only performed when the overall score of the backup map continuously exceeds the overall score of the currently used map and the difference is greater than a set replacement threshold.

[0035] Preferably, the global map consists of multiple sub-maps.

[0036] Preferably, the target unit is a sub-map, or the target unit is a grid unit that constitutes a sub-map.

[0037] Preferably, the sensor data is two-dimensional point cloud data obtained by lidar scanning.

[0038] Preferably, a data preprocessing step is included before matching the sensor data with the currently maintained global map:

[0039] The sensor data is converted from the sensor coordinate system to the global coordinate system used by the global map, based on the sensor installation parameters and the carrier pose.

[0040] Preferably, the matching step employs a two-stage matching process:

[0041] In the first stage, a correlation scanning matching algorithm is used to perform a discrete search within a preset position and angle search range. The coarse matching pose is determined by calculating the correlation score between the two-dimensional point cloud data and the map grid.

[0042] In the second stage, using the coarse matching pose as the initial value, the iterative nearest point algorithm is used for precise registration to obtain the precise matching pose and the matching result.

[0043] Preferably, the matching result includes at least one of the following: matching rate, residual distance, matching calculation time, and number of matching iterations.

[0044] Preferably, the step of performing the update action is executed asynchronously by a background thread that is independent of the main thread that performs the positioning and navigation functions.

[0045] Preferably, the method further includes an update trigger control step, which prevents map updates from being triggered when any of the following constraints are met:

[0046] a. Time interval limit: The time interval between the current moment and the last successful update is less than the minimum time threshold;

[0047] b. Update frequency limit: Within the preset statistical time window, the number of updates has reached the maximum threshold.

[0048] c. Pose stability limit: Since the last update, the pose change of the carrier is less than the displacement threshold, and the time elapsed is less than the time threshold.

[0049] The present invention has the following beneficial effects:

[0050] This application achieves intelligent and accurate decision-making for map updates by integrating an immediate cleanup criterion based on spatial consistency with a long-term value criterion based on reinforcement learning, effectively distinguishing between dynamic interference and permanent environmental changes. By combining a global map version scoring and replacement mechanism with trigger control with hysteresis thresholds, the system significantly improves operational efficiency and stability while ensuring long-term map consistency and high positioning accuracy. With the addition of strict coordinate alignment, asynchronous execution, and rollback fault tolerance design, the application provides mobile robots with a reliable, adaptive, and long-term maintainable high-quality map in dynamic and complex environments, significantly enhancing the robustness and practicality of their autonomous navigation. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of a planar 2D point cloud map update method based on dynamic value assessment provided in an embodiment of this application;

[0053] Figure 2 This is a flowchart illustrating the implementation of the immediate cleanup guidelines provided in this application's embodiments;

[0054] Figure 3 This is a flowchart illustrating the calculation of the continuity metric provided in the embodiments of this application;

[0055] Figure 4 This is an execution flowchart of the reinforcement learning model provided in the embodiments of this application;

[0056] Figure 5 This is a flowchart illustrating the implementation of map version replacement decision-making provided in this application embodiment. Detailed Implementation

[0057] To make the technical solution of this application clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The terms "first," "second," etc., in the claims and specification of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate. This is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of units is not necessarily limited to those units, but may include other units not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0058] like Figure 1 As shown, this application provides a method for updating a planar 2D point cloud map based on dynamic value assessment, including the following steps:

[0059] S110. Obtain the sensor data of the current frame and match the sensor data with the currently maintained global map to obtain the matching result;

[0060] S120. Based on the matching results, the target units in the map are updated and evaluated. The update evaluation integrates the immediate cleanup criterion based on spatial consistency and the long-term value criterion based on historical state information.

[0061] S130. Based on the results of the update evaluation, determine the update actions for the target unit. The update actions shall include at least a deletion operation.

[0062] S140. Perform an update action to update the global map.

[0063] First, data is acquired using sensors. This stage primarily involves acquiring radar data. Specifically, raw LiDAR data is received through the `sensor::LidarFrameData_` structure, and a single-frame two-dimensional LiDAR point cloud (`LidarPointCloud2D`) is defined. , among which, the Point cloud data can be represented in polar coordinates in the radar coordinate system. , representing the distance and angle of the point, respectively, corresponding to the local coordinate point. and has a timestamp Simultaneously, the current vehicle pose in the global coordinate system is obtained. ,in, For the overall position of the vehicle body, This refers to the vehicle's yaw angle.

[0064] Furthermore, a data preprocessing step is included before matching the sensor data with the currently maintained global map:

[0065] The sensor data is converted from the sensor coordinate system to the global coordinate system used by the global map, based on the sensor installation parameters and the carrier pose.

[0066] Let the installation parameters of the lidar relative to the vehicle body be: , representing the installation offset in the x and y directions in the vehicle coordinate system and the installation rotation angle relative to the vehicle body, respectively. This parameter can be read from the configuration item _lidarConfig.m_installPose[0]. Based on the installation angle sine value With cosine value Construct a rotation matrix from the radar coordinate system to the vehicle coordinate system, and then transform the points in the radar coordinate system... Transform to the vehicle coordinate system.

[0067] The transformation formula between the radar coordinate system and the vehicle coordinate system is as follows:

[0068] ,

[0069] .

[0070] Next, using the global pose of the vehicle body Points in the vehicle coordinate system Transform to global coordinate system:

[0071] ,

[0072] .

[0073] Let the pose of the local map coordinates in the global coordinate system be... ,in, This represents the position of the origin of the local coordinate system on the global coordinate system. This represents the yaw angle of the local coordinate system relative to the global coordinate system. (This is used to map the local coordinate system relative to the global coordinate system.) After transforming to the local coordinate system of the map, the translation difference is first calculated to eliminate the global offset of the map origin:

[0074] .

[0075] Then, the coordinates of this point in the local coordinate system of the map are obtained through rotation transformation. :

[0076] ,

[0077] .

[0078] Furthermore, to eliminate time deviations caused by angular differences in lidar scanning, timestamp compensation is required for each point in the point cloud. Let the reciprocal of the laser emission angular velocity be... For the i-th point, its time compensation amount It can be represented as The corrected timestamp is , This is the base timestamp for the point cloud in this frame.

[0079] Through the aforementioned coordinate transformation chain and time compensation, the original laser point cloud data is sequentially transformed from the sensor coordinate system to the vehicle coordinate system, then to the global coordinate system, and finally unified to the local map coordinate system. This ensures that the point cloud data is aligned with the map coordinate reference under spatiotemporal consistency, providing an accurate reference for subsequent matching of grids and point clouds.

[0080] After data preprocessing is complete, the sensor data is matched with the currently maintained global map.

[0081] Furthermore, this matching step employs a two-stage matching process:

[0082] In the first stage, a correlation scanning matching algorithm is used to perform discrete searches within a preset position and angle search range. The coarse matching pose is determined by calculating the correlation score between the two-dimensional point cloud data and the map grid.

[0083] In the second stage, using the coarse matching pose as the initial value, the iterative nearest point algorithm is used for precise registration to obtain the precise matching pose and matching result.

[0084] In the first stage, coarse matching is performed. Specifically, a correlation-based scanning matching algorithm is used to perform a discrete search within a preset position search range and angle search range. The correlation score between the 2D point cloud data and the map grid is calculated based on the number of point clouds within the hit grid cells, determining a coarse matching pose as the initial pose. The matching rate can be evaluated by the ratio of the total number of point clouds in the hit grid cells to the total number of point clouds currently scanned. For example, the search radius can be set to 0.2 meters to 0.6 meters, and the search angle range can be set to 3 degrees to 30 degrees, which can be adjusted according to environmental characteristics (such as reflector distribution).

[0085] In the second stage, fine matching is performed. Using the coarse matching pose obtained in the first stage as the initial value, a 2D scan matcher based on the Ceres optimizer (CeresScanMatcher2D) is used for fine optimization, and the Point-to-Line ICP (PLICP) algorithm is employed for pose optimization. This process can also incorporate constraints from the inertial measurement unit or vision sensor, and apply different weights to specific features (such as reflectors) or historical grid information to improve matching accuracy.

[0086] The matching process described above, consisting of two stages, outputs a matching result. This matching result includes, but is not limited to, at least one of the following quality indicators:

[0087] Matching rate: The ratio of the number of successfully matched point clouds to the total number of point clouds;

[0088] Residual distance: the sum of squared distances between the actual location of a point cloud and its corresponding location on the map;

[0089] Relevance matching time / precise matching time: can be further subdivided into relevance matching time and exact matching time;

[0090] Matching iteration count: The number of iterations in the fine-matching optimization process;

[0091] Pose sequence: initial pose, intermediate pose and final pose during the matching process;

[0092] Match with independent graph: An indicator of whether a match was successfully made with an independent subgraph. .

[0093] By calculating and outputting the aforementioned matching quality indicators, the system can quantitatively assess the reliability, accuracy, and computational load of the current matching before making map update decisions. These indicators provide crucial basis for subsequent judgments on whether map structure needs optimization and whether matching parameters are appropriate, thereby ensuring the stability and efficiency of the map update process.

[0094] Based on the obtained matching results, the target cells in the map are updated and evaluated. This update evaluation integrates the immediate cleanup criterion based on spatial consistency with the long-term value criterion based on historical state information to achieve intelligent maintenance of map cells in a dynamic environment.

[0095] First, implement immediate cleanup criteria based on spatial consistency. Specifically, such as... Figure 2 As shown, the guideline includes the following steps:

[0096] S210. Based on the matching results, calculate the matching score between the target unit and the matched sensor data.

[0097] In one specific implementation, calculating the matching score includes: after converting the sensor data to the world coordinate system, measuring the difference between the scan point and the corresponding map grid using Mahalanobis distance, and calculating the score accordingly.

[0098] Match score The calculation formula is as follows:

[0099] .

[0100] in, This represents the mean of the raster. The Mahalanobis distance, Represents the covariance matrix. This represents the preset maximum variance. A higher value indicates a higher degree of match. For example, given the raster mean... =(2.0,3.0), scan point P=(2.1,2.9), maximum variance =0.1, and the calculation yields... ≈0.447, ≈0.368.

[0101] S220. When the matching score is lower than the first threshold, calculate the continuity metric based on the matching status of each adjacent unit of the target unit.

[0102] If the matching score is higher than or equal to the first threshold, the target unit is determined to be retained.

[0103] Furthermore, such as Figure 3 As shown, the calculation of the continuity metric specifically includes the following steps:

[0104] S231. Based on the matching results, determine whether each adjacent unit of the target unit matches the current sensor data, and count the number of matching adjacent units.

[0105] S232. Divide the number by the total number of adjacent units to obtain the basic continuity score;

[0106] S233. Multiply the basic continuity score by a boundary protection coefficient based on the target unit location to obtain the final continuity metric.

[0107] Specifically, for target units located at the edge or corner of the map, the boundary protection factor (β) is greater than 1; for target units in other locations, the boundary protection factor (β) is equal to 1.

[0108] The boundary protection coefficient is designed to prevent accidental deletion of edge region cells due to limited viewing angles or incomplete data by increasing the continuity metric of these cells, thereby effectively protecting the integrity of the map's boundary structure. For example, β=1.5 can be set for cells located at edges or corners.

[0109] S230. Based on the continuity metric, the confidence level of the target unit is attenuated.

[0110] In one specific embodiment, the confidence decay process is as follows: Let the continuity metric be... The basic attenuation coefficient is The actual attenuation coefficient Where N is the total number of adjacent cells. Subsequently, the new confidence level of the target cell... Updated to .

[0111] Assuming the central grid cell does not match, 5 out of its 8 neighbors will match. ,and ,but If the original confidence level Then the new confidence level This result is higher than the direct decay result without considering neighbors. The central grid is preserved. This mechanism reflects the mitigating effect of spatial continuity constraints on confidence decay, helping to preserve the static units that constitute the continuous structure.

[0112] If the aforementioned central grid cell is located precisely at the edge or corner of the map, the boundary protection mechanism described in step S233 will be activated. Its continuity metric will be based on the boundary protection coefficient (assuming...). The score was adjusted to obtain the enhanced score. Then, substitute the values ​​into the attenuation formula from step S230 for calculation. This allows edge cells to obtain a smaller actual attenuation coefficient when the number of neighbor matches is the same, thus providing more effective protection during iterative updates and preventing premature erosion of the map outline due to incomplete observations.

[0113] S240. If the confidence level after decay is lower than the second threshold, then a decision tendency to delete the target unit is generated.

[0114] The generated deletion decision tendency will be incorporated into the final map update decision. To efficiently process large-scale point cloud data to support the above evaluation, the system can use a spatially partitioned data structure (such as KD-Tree) for fast nearest neighbor search, and allocate time-consuming processing tasks to independent threads for asynchronous execution to avoid blocking the SLAM main thread. Simultaneously, motion compensation can be applied to the point cloud data to correct biases caused by sensor motion, improving the real-time performance and accuracy of the evaluation.

[0115] In addition, the system performs topology consistency checks to improve the accuracy of decisions regarding the deletion of isolated grids.

[0116] Specifically, for candidate grids that are in the medium probability range (e.g., 0.5) after confidence decay and exhibit isolated behavior (i.e., most of their neighboring cells did not match in this match), the system will further initiate a topology consistency check. This check aims to verify the spatial connectivity of the grid with the surrounding known static environment to distinguish whether it belongs to a fixed structural object (such as a column) or a temporarily appearing dynamic object (such as a moving AGV).

[0117] During the inspection, the system combines historical observation data to analyze whether the grid maintains a reasonable spatial connection with surrounding stable structures. Its evaluation weight is dynamically adjusted: if the grid is continuously hit by sensor data, its weight as a static structure increases, making it more likely to be retained; conversely, if it has not been observed for a long time, its weight decreases, making it more likely to be removed. This mechanism ensures that even if a side of a fixed object is not seen in a single scan due to occlusion or other reasons, it will not be mistakenly identified as a dynamic object and deleted; at the same time, it can quickly remove transient point clouds generated by truly moving objects.

[0118] To improve the overall efficiency of the update and evaluation process, the system also adopts the following optimization and management strategies:

[0119] 1. Asynchronous parallel processing: The computational tasks such as point cloud matching, raster evaluation and topology inspection are assigned to independent worker threads for asynchronous execution, avoiding blocking the SLAM main thread and ensuring the real-time performance of the system.

[0120] 2. Layered cleaning frequency control: Based on the type, location and state change frequency of grid cells, the trigger cycle for evaluation and cleaning is set differently to optimize the allocation of computing resources.

[0121] 3. Parallel management of sub-maps: The global map is divided into multiple sub-maps for maintenance, supporting parallel processing, improving system throughput, and facilitating map expansion to multiple floors and regions.

[0122] 4. Storage and memory optimization: Set an upper limit on the number of point cloud storage points for each grid cell (e.g., using a FIFO strategy to retain a maximum of 5 points), automatically perform data deduplication and cleaning, effectively prevent unlimited memory growth, and improve map data quality.

[0123] By employing the aforementioned spatial consistency-based real-time cleanup criteria (integrating matching degree assessment, neighborhood continuity measurement, confidence decay, and topology consistency check), and combined with hierarchical, asynchronous, and parallel engineering optimizations, the system can make accurate and efficient update assessments of map units in complex and dynamic environments, providing a reliable basis for subsequent deletion and other update actions.

[0124] After performing an immediate evaluation of map units based on spatial consistency, the system enters the hierarchical fusion decision-making stage. Specifically, based on target units that have been quantitatively evaluated using the immediate cleanup criterion and marked as having a tendency to be deleted, the system generates a set of target units (i.e., a candidate set) to be deeply evaluated. For each target unit in this set, the system triggers and executes a deep evaluation based on a long-term value criterion using historical state information. The core of this criterion is to use a reinforcement learning model to predict the long-term expected value of each candidate unit performing different update actions (such as deletion or retention), thereby providing a basis for the final decision that balances immediate matching benefits with long-term value. Figure 1 Consistent global value basis. This mechanism of "triggering targeted deep evaluation based on real-time quantification results" significantly reduces the average computational cost of the system and improves the overall processing efficiency by limiting the computationally intensive reinforcement learning evaluation to a limited number of candidate units that have been pre-screened, while ensuring the intelligence of decision-making.

[0125] This long-term value criterion based on historical state information is implemented through a reinforcement learning model, specifically, such as... Figure 4 As shown, the reinforcement learning model is configured to perform the following steps:

[0126] S310. The update decision process for the target unit is modeled as a Markov decision process, wherein the state includes at least the current matching degree, current confidence degree, historical update count, and historical variance of the observations of the target unit.

[0127] S320. Based on the Bellman optimal equation, calculate the state value function value corresponding to different update actions performed on the target unit, as the basis for evaluating its long-term expected value.

[0128] This embodiment models the update decision process for map cells as a Markov decision process and calculates the long-term expected value of different update actions based on the Bellman optimality equation. Specifically, the "whether to delete a grid" decision is modeled as a reinforcement learning problem. By calculating the long-term payoffs of actions such as "keep" and "delete," a basis is provided for the final decision. The implementation is as follows:

[0129] 1. Markov Decision Process Modeling

[0130] state This represents the historical and current metrics for each map unit (grid), including at least: current matching degree (i.e., spatial consistency). ), current confidence level Historical update frequency and historical variance of observations .

[0131] action : The update actions that can be performed, including insert, keep, and delete.

[0132] award In state Next action The instant reward received.

[0133] State transition probability In state Next action Afterwards, transition to the new state. The probability of.

[0134] Discount factor : Used to adjust the relative importance of future returns (0≤γ<1).

[0135] 2. Value Function and Bellman Optimality Equation

[0136] Define the value state function Indicates from state Starting from the optimal strategy, we determine the long-term expected total return achievable by following the optimal strategy. This is accomplished through the Bellman optimality equation. Perform iterative updates to evaluate the long-term value of each action:

[0137] .

[0138] in, This represents the expected future payoff considering all possible subsequent states.

[0139] 3. Reward Function Design

[0140] reward function Designed as:

[0141] .

[0142] in, , , These are the weights for confidence level, matching degree, and historical variance, respectively. If the action is to retain (… The immediate reward is 0; if the action is deletion ( The immediate reward is determined by the weighted sum of the three states. This design allows for higher rewards for deleting rasters with low confidence, low matching, or high variance, while lower or even negative rewards are given for deleting high-confidence and stable rasters.

[0143] 4. State transition probability design

[0144] .

[0145] State transition probability Set according to the action: If the action is delete ( The state transition is deterministic; after deletion, the raster becomes an "empty" state, and its state value... Set to baseline value 1.0; if the action is reserved ( The transition probability depends on the matching rate of the surrounding neighbors of the grid. If the neighbor matching rate is high, it indicates that the area is relatively static, and the grid is likely to remain stable after retention; conversely, if the neighbor matching rate is low, the future benefits after retention will be lower.

[0146] This reinforcement learning method, through Markov decision processes and Bellman optimal equations, comprehensively weighs immediate rewards against long-term consequences, enabling intelligent identification and deletion of grids corresponding to dynamic obstacles during map maintenance. This method is particularly suitable for complex and dynamically changing environments, improving the long-term consistency and accuracy of maps and achieving intelligent map maintenance.

[0147] To facilitate understanding, the decision-making process is explained below with specific numerical data:

[0148] Suppose the state of a target grid is as follows: , ;

[0149] Weight settings: , , ;

[0150] Discount factor Decision threshold .

[0151] Instant Rewards for Deletion Actions Future benefits Then the long-term value of the deletion action. .

[0152] Retain Action Instant Reward =0, assuming the future state value corresponding to the neighbor matching rate is... =0.85, then the future return = This preserves the long-term value of the action. .

[0153] The calculation results show that the deletion action yields a higher long-term expected value (1.66 > 0.765). This long-term value ( and This will be integrated with the decision tendency output by the instant cleanup criterion based on spatial consistency, and together they will serve as the "result of the update assessment" to ultimately determine the update action.

[0154] Specifically, the system uses the output of long-term value criteria as the core decision-making basis, along with preset decision thresholds. A comparison is made to determine the initial inclination towards the immediate cleanup principle. The rules are as follows:

[0155] when > + Regardless of the immediate criterion's bias, the system determines to perform the deletion operation based on the long-term value assessment results. This rule establishes the decisive role of the long-term value criterion in decision-making; that is, when the expected long-term benefits of deletion are significantly higher than retention, the system will adopt the assessment conclusion to ensure the long-term consistency of the map and the reliability of navigation, effectively overcoming decision-making fluctuations caused by the uncertainty of single or instantaneous observations.

[0156] when + In such cases, the tendency to delete based on the immediate cleanup criterion is rejected, and the unit is ultimately retained. In this situation, although the current observational match is poor, the long-term value assessment indicates that retaining the unit will bring higher or acceptable future expected benefits, thus preventing the accidental deletion of valuable static structures due to single or short-term observational biases.

[0157] As in the example above, the target raster (1.66) is significantly greater than (0.765) and threshold The sum of (0.7) is (0.765 + 0.7 = 1.465). Therefore, the system directly determines deletion as the final update action based on the first case. This fusion mechanism clarifies the master-slave relationship between the two criteria: the immediate criterion acts as a "sensitive trigger," efficiently screening potential changes; the long-term criterion acts as an "intelligent arbiter," making the final decision based on historical and future benefits. It institutionally avoids erroneous deletions caused by instantaneous observation noise, ensuring the long-term accuracy, stability, and rationality of map updates in dynamic environments.

[0158] It should be noted that the system also supports different operating modes to adapt to the needs of mapping at different stages, specifically including the initial mapping mode and the optimization and maintenance mode.

[0159] In the initial mapping mode, the system employs a relatively lenient matching threshold (i.e., higher first and second threshold tolerances) and a higher point cloud insertion frequency. This strategy aims to ensure rapid map framework construction and maintain positioning stability in unfamiliar scenes, but may introduce dynamic object point clouds or lead to excessively high local point cloud density.

[0160] In the optimized maintenance mode, the system adopts stricter matching thresholds and the aforementioned complete real-time cleaning criteria based on spatial consistency and long-term value criteria based on historical state information to conduct refined evaluation and maintenance of map units, aiming to clean up low-value data and optimize map quality and consistency.

[0161] In one specific embodiment, such as Figure 5 As shown, the method also includes a map version replacement decision step:

[0162] S410. Maintain and update multiple map versions, including the currently used map and at least one backup map;

[0163] S420. Employ a multi-dimensional scoring function to comprehensively score multiple map versions. The multi-dimensional scoring function is at least based on... Figure 1 Construction of consistency score, positioning accuracy score and coverage integrity score;

[0164] S430. Based on the scoring results, determine whether to replace the currently used map with the backup map that has the highest overall score; wherein, a lag mechanism is introduced in the determination, and the replacement operation is only performed when the overall score of the backup map continuously exceeds the overall score of the currently used map and the difference is greater than the set replacement threshold.

[0165] To address the local optima problem that can arise from greedy decisions at each step of the Bellman equation (e.g., mistakenly deleting a raster and triggering a series of subsequent positioning and mapping errors), this embodiment further introduces a map version replacement decision mechanism. This mechanism does not directly apply update operations based on dynamic value assessment (such as adding or deleting rasters) to the currently used online map. Instead, it records these operations in a separate backup map. By comprehensively evaluating the long-term performance of multiple map versions, it makes a global, delayed decision on whether to replace the current map, thereby ensuring the long-term stability and optimal decision-making of the system.

[0166] During operation, the system continuously maintains a currently used map for positioning and navigation. At the same time, maintain one or more backup maps. All update actions generated by the aforementioned fusion decision-making mechanism are applied asynchronously to the backup map, while the currently used map remains unchanged. Furthermore, the system can be configured to retain a historical version depth. This creates a multi-version map collection that includes historical versions. This is used for subsequent comprehensive evaluation and decision-making.

[0167] To quantitatively evaluate map quality, the system defines a comprehensive scoring function. This function integrates sub-scores from multiple dimensions, and its general form is: .

[0168] in, This represents the weight of the i-th rating dimension, satisfying... ; Let i represent the sub-rating function of the i-th rating dimension.

[0169] In this embodiment, the scoring function is at least based on the ground. Figure 1 The system is constructed based on three dimensions: consistency, positioning accuracy, and coverage integrity, as defined below:

[0170] (1) land Figure 1 Consistency score : Assess the degree of matching between the map and historical sensing data to ensure that map updates do not deviate from reliable observation history.

[0171] .

[0172] in, To assess the number of historical scan points within a time window; For the first Coordinates of historical scan points With map Mahalanobis distance; The standard deviation parameter controls the tolerance for deviations. It reflects the map's compatibility with historical perception data, ensuring that the map remains consistent with past perception data after updates, avoiding the contradiction of "new maps negating historical observations," and possessing spatiotemporal continuity characteristics.

[0173] (2) Positioning accuracy score : Evaluate the deviation between the estimated pose and the true pose (or high-precision reference pose) when using the map for localization.

[0174] .

[0175] in, To assess the time step; It must be a very small positive number to prevent division by zero; Let be the estimated pose at time t; Let t be the true pose at time t. It reflects the map's ability to support positioning accuracy, ensuring that the map can support high-precision positioning and avoiding positioning drift caused by map inaccuracies.

[0176] (3) Coverage integrity score : Evaluate the completeness and accuracy of the map's coverage of the real-world environment.

[0177] .

[0178] in, Indicates the actual ground coverage area (or a high-confidence reference area); This is an area measurement function; Indicates symmetrical difference; It represents the intersection of the map and the real area, i.e., the area accurately covered; Indicates the missing coverage area of ​​the map; Indicates the overall reference range. It can ensure that the map fully covers the real environment, with no redundant or missing areas.

[0179] Finally, replacement decisions are made based on the rating results of each version, and a lag mechanism is introduced to improve stability.

[0180] First, the system calculates the posterior probability of replacing the current map with a backup map based on a Bayesian inference framework, in order to achieve robust decision-making in a probabilistic sense:

[0181] .

[0182] Among them, decision variables This indicates that a replacement will be performed. Indicates no replacement. Likelihood function Modeling using Gaussian mixture model:

[0183] .

[0184] in, The mixed weighting of the scores for each version satisfies ; For the first Gaussian distribution parameters for each version's rating The probability density function is Gaussian. This model fully considers the uncertainty of the scoring.

[0185] To avoid frequent map switching due to minor fluctuations in ratings, the system uses a threshold decision function with a hysteresis mechanism to make deterministic replacement decisions:

[0186] .

[0187] in, It has the highest rating among all evaluated map versions; The static score difference threshold; It is a lagged function based on historical score differences, and its value is non-negative, used to further suppress frequent switching.

[0188] The system will perform a replacement operation and update the currently used map to the optimal version only when the above conditions are met (i.e., when there is a map version with a consistently and significantly higher overall score than the currently used map).

[0189] The dynamic process of this decision can be represented by a Lyapunov function. Analysis was conducted. Theoretical analysis shows that, under certain conditions (such as a learning rate sequence), Under these conditions, the expected convergence of the system's decision state is... ,in, This is the set of all historical information up to time t. This mathematically guarantees the long-term stability and convergence of the replacement decision mechanism.

[0190] By introducing a map version replacement decision mechanism, immediate and local update actions are placed within a global and long-term evaluation framework for adjudication, effectively avoiding the accumulation of errors and local optima problems that may be caused by single-step greedy decisions. Combining multi-dimensional comprehensive scoring and Bayesian decision-making with a lag mechanism significantly improves the robustness, stability, and long-term consistency of the dynamic map update system, ensuring that the robot can achieve reliable and continuous autonomous operation in complex dynamic environments.

[0191] In one specific embodiment, the method further includes an update trigger control step. This step acts as a pre-filter in the map update process, designed to prevent ineffective consumption of system resources or unstable map disturbances caused by sensor noise, minor vibrations of the carrier, or excessively high frequency of sensing data. The system will only allow subsequent matching and update evaluation processes to be triggered if all of the following constraints are not met simultaneously; if any one condition is met, map updates will be prohibited from being triggered:

[0192] a. Time interval limit: Current moment The time interval since the last successful map update The interval between them is less than the set minimum time threshold. ,Right now ,in, The time limit is dynamically adjusted according to the system's operational phase. For example, it is set to 1 second during the initial mapping phase and 5 seconds during the optimization and maintenance phase.

[0193] b. Update frequency limit: Within a preset statistical time window (e.g., 5 seconds), the number of successfully executed map updates has reached the maximum threshold. ,Right now , Similarly, the process is divided into phases: the initial mapping phase can have 10 iterations, and the optimization phase can have 5 iterations.

[0194] c. Pose stability limit: Since the last update, the pose change of the carrier is less than the preset displacement threshold. And the time elapsed is less than the time threshold. That is, simultaneously satisfying , where the threshold pair During the initial mapping phase, it can be set to... During the optimization phase, it is set to This dual-threshold condition ensures that a substantial change in the observed vision is only considered after the carrier has undergone significant displacement or after a sufficiently long period of time, thus allowing an update to be triggered.

[0195] To enhance system robustness, this embodiment also supports a rollback mechanism for update decisions. When the system detects a quality degradation after an update (such as an abnormal increase in point cloud density or a significant degradation in positioning accuracy), a rollback function (such as undoLastUpdate()) can be called to undo the update, restoring the map to its previous consistent state. All update tasks are executed asynchronously in a background thread. The main thread is only responsible for adding the update actions generated by trigger control and fusion decisions to the execution queue and recording their timing, thereby ensuring that the real-time positioning and navigation main loop is not blocked.

[0196] To ensure the long-term availability and consistency of map data, the system also supports periodic and event-triggered map saving. The triggering conditions for persistence operations can be expressed as: .

[0197] in, (e.g., 50) is the counting threshold. (e.g., 30 seconds) is the time threshold. The saved map file will overwrite the file with the same name used for system playback or relocation, ensuring that the map is in the latest and optimal state the next time it is loaded, thus forming a closed management loop.

[0198] This application achieves accurate map update decisions by intelligently integrating both immediate and long-term criteria. By combining global scoring and fault-tolerant control mechanisms, it significantly improves system operating efficiency and stability while ensuring long-term map consistency and high accuracy, providing reliable and adaptive navigation support for robots in dynamic environments.

[0199] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for updating a planar 2D point cloud map based on dynamic value assessment, characterized in that, Includes the following steps: Obtain sensor data for the current frame and match the sensor data with the currently maintained global map to obtain the matching result; Based on the matching results, the target units in the map are updated and evaluated, wherein the update evaluation integrates the immediate cleanup criterion based on spatial consistency and the long-term value criterion based on historical state information. Based on the results of the update evaluation, an update action for the target unit is determined, and the update action includes at least a deletion operation; Perform the update action to update the global map.

2. The planar 2D point cloud map update method based on dynamic value assessment according to claim 1, characterized in that, The instant cleanup criteria based on spatial consistency include: Based on the matching results, a matching score is calculated between the target unit and the matched sensor data. When the matching score is lower than the first threshold, a continuity metric is calculated based on the matching status of each adjacent unit of the target unit. Based on the continuity metric, the confidence level of the target unit is attenuated; If the confidence level after decay is lower than the second threshold, a decision tendency to delete the target unit is generated.

3. The planar 2D point cloud map update method based on dynamic value assessment according to claim 2, characterized in that, The calculation of the continuity metric includes: Based on the matching results, determine whether each adjacent unit of the target unit matches the current sensor data, and count the number of matching adjacent units; Divide the number by the total number of adjacent units to obtain the basic continuity score; The basic continuity score is multiplied by a boundary protection coefficient based on the target cell location to obtain the final continuity metric. Specifically, for target units located at the edge or corner of the map, the boundary protection coefficient is greater than 1; for target units at other locations, the boundary protection coefficient is equal to 1.

4. The planar 2D point cloud map update method based on dynamic value assessment according to claim 1, characterized in that, The long-term value criterion based on historical state information is implemented through a reinforcement learning model, which is configured to perform the following steps: The update decision process for the target unit is modeled as a Markov decision process, wherein the state includes at least the current matching degree, current confidence degree, historical update count, and historical variance of the observations of the target unit. Based on the Bellman optimal equation, the state value function value corresponding to different update actions performed on the target unit is calculated, which serves as the basis for evaluating its long-term expected value.

5. The planar 2D point cloud map update method based on dynamic value assessment according to claim 4, characterized in that, The reward function in the reinforcement learning model is set as follows: When the update action is deletion, the output instant reward value is negatively correlated with the current confidence level of the target unit, positively correlated with the current mismatch level, and positively correlated with the historical observation variance.

6. The planar 2D point cloud map update method based on dynamic value assessment according to claim 1, characterized in that, The update assessment step is implemented through a decision fusion mechanism, which integrates the decision tendency generated by the immediate cleanup criteria with the expected value assessed by the long-term value criteria to determine the final update action.

7. The planar 2D point cloud map update method based on dynamic value assessment according to claim 1, characterized in that, The method also includes a map version replacement decision step: Maintain and update multiple map versions, including the currently used map and at least one backup map; A multi-dimensional scoring function is used to comprehensively score the multiple map versions. The multi-dimensional scoring function is constructed based on at least map consistency score, positioning accuracy score, and coverage integrity score. Based on the scoring results, it is determined whether to replace the currently used map with the backup map that has the highest overall score; wherein, the determination introduces a lag mechanism, and the replacement operation is only performed when the overall score of the backup map continuously exceeds the overall score of the currently used map and the difference is greater than a set replacement threshold.

8. The planar 2D point cloud map update method based on dynamic value assessment according to claim 1, characterized in that, The global map consists of multiple sub-maps.

9. A planar 2D point cloud map update method based on dynamic value assessment according to claim 8, characterized in that, The target unit is a sub-map, or the target unit is a grid unit that constitutes a sub-map.

10. A planar 2D point cloud map update method based on dynamic value assessment according to claim 1, characterized in that, The sensor data is two-dimensional point cloud data obtained by lidar scanning.

11. The planar 2D point cloud map update method based on dynamic value assessment according to claim 10, characterized in that, The process includes a data preprocessing step before matching the sensor data with the currently maintained global map: The sensor data is converted from the sensor coordinate system to the global coordinate system used by the global map, based on the sensor installation parameters and the carrier pose.

12. The planar 2D point cloud map update method based on dynamic value assessment according to claim 11, characterized in that, The matching step employs a two-stage matching process: In the first stage, a correlation scanning matching algorithm is used to perform a discrete search within a preset position and angle search range. The coarse matching pose is determined by calculating the correlation score between the two-dimensional point cloud data and the map grid. In the second stage, using the coarse matching pose as the initial value, the iterative nearest point algorithm is used for precise registration to obtain the precise matching pose and the matching result.

13. The planar 2D point cloud map update method based on dynamic value assessment according to claim 12, characterized in that, The matching results include at least one of the following: matching rate, residual distance, matching calculation time, and number of matching iterations.

14. The planar 2D point cloud map update method based on dynamic value assessment according to claim 1, characterized in that, The steps for performing the update action are executed asynchronously by a background thread independent of the main thread that performs the positioning and navigation functions.

15. The planar 2D point cloud map update method based on dynamic value assessment according to claim 1, characterized in that, The method further includes an update trigger control step, which prevents map updates from being triggered when any of the following restrictions are met: a. Time interval limit: The time interval between the current moment and the last successful update is less than the minimum time threshold; b. Update frequency limit: Within the preset statistical time window, the number of updates has reached the maximum threshold. c. Pose stability limit: Since the last update, the pose change of the carrier is less than the displacement threshold, and the time elapsed is less than the time threshold.

Images