A dynamic mask localization method for AGVs with dual temporal maps for dynamic warehousing

Abstract

This invention belongs to the field of AGV positioning technology and discloses an AGV dual-temporal map dynamic mask positioning method for dynamic warehousing, including the following steps: S1, multi-sensor data acquisition; S2, dual-temporal map construction; S3, dynamic mask generation; S4, multi-sensor fusion positioning; S5, sub-map dynamic management; S6, dynamic detection specialization of single-line laser. The beneficial effects of this invention are as follows: 1. Improved positioning robustness in dynamic scenarios; 2. Enabled rapid adaptation and updating to highly dynamic areas; 3. Enhanced dynamic detection capabilities based on single-line laser radar; 4. Improved adaptive update mechanism for positioning accuracy and computational efficiency; 5. Provided high-precision positioning and map updates.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of AGV positioning technology, in particular to an AGV double-time-domain map dynamic mask positioning method for dynamic warehousing. BACKGROUND

[0002] The current warehouse and discrete manufacturing workshop generally adopts the operation mode of manual forklift, cage truck and fixed station cooperation, the operation area is wide, the stations and platforms are scattered, the cross-zone inbound and outbound and in-process product transfer are frequent. Manual handling has limitations such as difficulty in standardizing path and speed, beat fluctuation caused by shift switching, fatigue and safety risk increase; at the same time, with the upgrading of WMS / MES, automated transportation and vertical warehouse and other technologies, the production line puts forward higher requirements for stability, availability and throughput capacity: maintaining a higher loading rate, lower waiting and empty running, better docking accuracy and process traceability without expanding the labor force. This trend makes the in-site logistics capacity of "continuous, high precision and all-weather" become the bottleneck link.

[0003] Under the above background, the latent AGV can realize flexible expansion and path optimization through software scheduling due to its low vehicle body, strong maneuverability and the ability to directly lift the bottom of the pallet / shelf, which can significantly reduce the labor intensity and safety risk, and improve the handling efficiency and operation consistency, thereby directly compensating for the problems described in the first paragraph. Due to the lack of high-reliability GNSS indoors, the navigation and docking of AGV rely on SLAM based on vehicle-mounted multi-sensor to continuously provide high-precision pose and maintainable map: typical sensors include wheel odometer and IMU to provide strong priori in short time, single-line two-dimensional laser radar and camera to provide environmental constraints; the SLAM result of the fusion of the two is the basis for path planning, obstacle avoidance, scheduling and centimeter-level docking, and its robustness directly determines the beat and availability of the vehicle fleet.

[0004] The existing mainstream scheme is mainly composed of grid / related matching (such as ICP, CSM, NDT) and graph optimization framework, or adopts VIO / VO mainly based on vision, and performs well under static or quasi-static assumption. However, in actual warehouse / factory, large goods or instruments are moved, stacking form changes frequently, forklifts pass through and curtains / gates open and close, etc., which will cause large-scale and high-frequency changes in the scene: near-field dynamic points will drag laser-map matching to make the optimization fall into the wrong minimum value; short-time objects are written into the historical map to cause pollution and drift; repeated shelf skeletons will cause multiple solutions and mislooping of attitude-lateral displacement; single-line laser only provides single-plane information and is easily blocked by the bottom of the goods, and the wheel speed slips and IMU bias will amplify the priori error; the small drift of time synchronization and external parameters is amplified under the above dynamic conditions. The above factors will cause the positioning to be greatly degraded or even fail (such as unable to reposition, whole map misplacement), and force frequent line stop reconstruction or manual map cleaning, which directly affects the production line beat and operation and maintenance cost. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a dynamic mask positioning method for AGVs with dual time-domain maps for dynamic warehousing. This invention is achieved through the following technical solutions.

[0006] A dynamic mask localization method for AGVs with dual temporal maps for dynamic warehousing includes the following steps:

[0007] S1, Multi-sensor data acquisition

[0008] Real-time warehouse environment information is acquired through sensors, including wheel speed odometers, IMUs, single-line lidar, and cameras. The IMUs include gyroscopes and accelerometers, and the single-line lidar is a 2D-LiDAR.

[0009] S2, Dual-Time Domain Map Construction

[0010] Construct a dual-time-domain map with an anchor layer (AM) and a real-time layer (LM), where the anchor layer is used to store long-term stable structural information and the real-time layer is used to quickly respond to dynamic changes;

[0011] Calculate the rate of change index (CRI) for each map cell in a dual-time-domain map to determine the dynamics of the map cell;

[0012] S3, Dynamic Mask Generation

[0013] Dynamic masks and their weights are generated based on CRI; stable and dynamic regions are determined based on the dynamic masks;

[0014] S4, Multi-sensor Fusion Positioning

[0015] S41 uses the initial pose estimation and motion constraints provided by the IMU and wheel speed odometer to perform precise localization by combining LiDAR-map matching with dynamic mask weights, and masks the dynamic area during LiDAR-map matching.

[0016] S42, in dynamic environments, if laser information is insufficient or matching fails, switch to visual-inertial odometry as a supplement to ensure stable positioning output;

[0017] S43, localization is achieved through ESKF (Error State Kalman Filter);

[0018] S5, Subgraph Dynamic Management

[0019] A regional playback strategy enables rapid updates to hot zones and stable maintenance of cold zones;

[0020] S6, Dynamic Detection Specialization for Single-Line Laser

[0021] Based on temporal consistency analysis, semantic projection fusion, and occlusion chain tracking, dynamic objects are identified and CRI is updated.

[0022] As a further aspect of the present invention, the method for calculating the rate of change index (CRI) in step S2 is as follows:

[0023] S21, Observation count update, specifically:

[0024]

[0025]

[0026]

[0027] Among them Map unit; and These are the updated and unupdated values ​​of the map unit hit strength status, respectively. and These are the values ​​of the map unit's free space state after the update and the values ​​before the update, respectively. and These are the values ​​after the map unit state change intensity update and the values ​​before the update, respectively. , and The attenuation coefficient is, and ; This is an indicator function; it is denoted as 1 if the condition is true, and 0 otherwise. Used to determine whether a map cell is occupied; Used to determine whether a map cell is free space; Used to determine whether the state of a map cell has changed;

[0028] S22, Calculate the frequency of map cell occupancy. and uncertainty Specifically:

[0029]

[0030]

[0031] in, It is a constant greater than zero;

[0032] S23, the fusion calculation of the rate of change index (CRI), specifically:

[0033]

[0034] in, This is a 0-1 normalization function; and A fusion weight ≥ 0; To calculate the length of the statistical window; This is a semantic dynamic prior from the camera; , is the geometric residual index;

[0035] S24, Determining the dynamic level of map units, for map units Its corresponding The larger the value, the higher its dynamic degree.

[0036] As a further aspect of the present invention, the calculation method for the dynamic mask in step S3 is as follows:

[0037]

[0038] in, For map units Dynamic mask; The threshold is determined dynamically.

[0039] like Then the map unit is a dynamic area; if If so, then the map unit is a stable region.

[0040] As a further aspect of the present invention, the calculation method for the dynamic mask weights in step S3 is as follows:

[0041]

[0042] in, for The weights; The sigmoid function is used to map input to... scope; The parameters used to control the slope of the Sigmoid curve.

[0043] As a further aspect of the present invention, step S41 specifically comprises:

[0044] In a 2D plane, with pose increments To optimize the variables, a LiDAR-map matching objective function with dynamic mask weights is constructed, specifically as follows:

[0045]

[0046] in, Represents a matrix transpose; ; and These represent the translation increment on the x-axis, the translation increment on the y-axis, and the rotation increment, respectively. The currently observed number The pose characteristics of each point; This is a pose transformation function used to transform... Transform from local coordinate system to global coordinate system; for The squared distance to the corresponding point in AM; The currently observed number Dynamic mask weights for each point; For robust kernel functions;

[0047] The output of the objective function for LiDAR-map matching with dynamic mask weights is as follows:

[0048]

[0049] in For the optimized pose estimation, and These represent the position on the x-axis, the position on the y-axis, and the rotation angle, respectively.

[0050] and output covariance .

[0051] As a further aspect of the present invention, step S43 specifically includes:

[0052] S431, Initialization: Calculates gyroscope and accelerometer biases using the static window mean and sets the initial covariance; specifically:

[0053] Calculate the average angular velocity measured by the gyroscope while the object is stationary. and the average acceleration measured by the accelerometer Specifically:

[0054]

[0055]

[0056] in, This is the window length in a static state. and The first Gyroscope and accelerometer readings at each sampling point;

[0057] Gyroscope bias ;

[0058] Accelerometer bias ;

[0059] in, This is the rotation matrix from the world coordinate system to the body coordinate system; Let be the gravity vector in the world coordinate system, and ; It is the acceleration due to gravity;

[0060] initial covariance = ;

[0061] in, and These are the initial covariances of position, velocity, yaw angle, gyroscope bias, accelerometer bias, and gravitational acceleration, respectively.

[0062] S432, prediction, performs debiasing processing on IMU data based on gyroscope and accelerometer biases; predicts position, velocity, and attitude based on IMU data integration; and performs covariance prediction; specifically:

[0063] The formula for IMU data bias correction is as follows:

[0064] ; ;in and These are the debiased values ​​for angular velocity and acceleration, respectively. and These are the measured values ​​of angular velocity and acceleration, respectively.

[0065] Location prediction value Specifically:

[0066]

[0067] Speed ​​prediction value Specifically:

[0068]

[0069] Attitude prediction value Specifically:

[0070]

[0071] in, , and These are the position, velocity, and attitude readings obtained from the IMU in the world coordinate system, respectively. For time step; This is an exponential mapping function used to convert a rotation vector into a rotation matrix;

[0072] Attitude prediction value Specifically:

[0073]

[0074] in, This is the state transition matrix;

[0075] , A function for generating block diagonal matrices;

[0076] S433, Update, specifically includes wheel speed odometer observation update, LiDAR-map matching observation update and covariance update;

[0077] The wheel speed odometer observation update is as follows: the wheel speed odometer directly obtains the speed observation value under the machine system, constructs an observation residual model to compare the difference between the speed prediction value and the speed observation value; the sensitivity of speed and attitude to the observation is quantified by calculating the Jacobian matrix, and finally the speed and attitude are dynamically corrected by using Kalman gain.

[0078] The LiDAR-map matching observation update is as follows: global pose correction is obtained by registering the LiDAR point cloud with the pre-built map, forming the observation residuals of position and attitude; the influence of position and attitude on point cloud matching error is quantified by the Jacobian matrix, and spatial geometric constraints are incorporated into the filtering framework to correct the robot's global pose.

[0079] The covariance update is specifically as follows: the state covariance matrix is ​​dynamically adjusted based on the Kalman gain, and the local motion constraints of the wheel speed sensor and the global geometric constraints of LiDAR are fused; the uncertainty of position, velocity and attitude is reduced by weighted fusion of the predicted covariance and the observation noise.

[0080] S434, Reset and tangent space reprojection, projects the updated covariance onto the new tangent space.

[0081] As a further aspect of the present invention, step S5 specifically includes:

[0082] S51, Hot Zone Determination, for any region ,like And the duration exceeds If the temperature is high, it is considered a hot zone; otherwise, it is considered a cold zone.

[0083] S52, Playback and Write, creates or resets local subgraphs for hotspots in LM. The area corresponding to the AM hot zone remains frozen;

[0084] S53, Isolation Strategy: When optimizing and matching AGV positioning, constraints are first established on the AM (Advanced Positioning Component); when the AM is sparse or mismatched, [the strategy is implemented by...]. It provides local auxiliary matching and performs weight reduction processing through dynamic mask weights;

[0085] S54, Promotion / Demotion Mechanism, when And the duration exceeds the duration At that time, from Select a stable boundary or skeleton feature to promote to AM; if the temperature rises again, reverse and degrade back to LM.

[0086] S55, adaptive gating, relaxes the upper limit of gating in hot zones to reduce false negatives; tightens and relaxes the upper limit of gating in cold zones to prevent false loop closure.

[0087] in and This represents the temperature threshold.

[0088] As a further aspect of the present invention, step S6 specifically includes:

[0089] S61, Timing Consistency Analysis

[0090] For the Compare the measured values ​​of the single-line lidar measurement line. Compared with the optimized pose estimation The difference yields the residual. Large residual continuous segments are classified as dynamic objects.

[0091] right Update, specifically ,in It is a non-negative increasing function;

[0092] Assume continuous residual segments The set of map units covered by its projection is ,calculate The CRI increment is updated as follows:

[0093]

[0094] in, For residual evidence weights, for The length; if If the CRI increases significantly, it is determined to be a dynamic object;

[0095] S62, Semantic Projection Fusion

[0096] Projecting the camera detection frame onto the ground yields a polygon. For each map unit Calculate its relationship with The overlap rate is as follows:

[0097]

[0098] in The overlap rate calculation function is used to determine regions with high overlap rates as semantically dynamic regions.

[0099] S63, Obstruction Chain Tracking

[0100] For a single-line lidar survey line, if the free space it traverses is occupied, then backtrack along the path of that survey line to find the continuous map cell chain that transitions from free space to occupied space. Then, for that continuous map cell chain... The weights;

[0101] S64, update CRI, and update the CRI. , and Substitute into step S23, for Update.

[0102] The beneficial effects of this invention are as follows:

[0103] 1. To improve the robustness of localization in dynamic scenes, a combination of dual temporal maps (long-term stable / short-term) and dynamic masks is introduced, enabling rapid isolation and updating of dynamic regions. While maintaining a low update frequency in long-term stable regions (anchor layer), the real-time layer can quickly respond to dynamic changes and update highly dynamic regions in a timely manner. Through dynamic masks, the system can suppress the negative impact of dynamic regions on pose estimation, performing accurate localization only in stable regions, thereby effectively avoiding the contamination of map building and localization by dynamic objects.

[0104] 2. To achieve rapid adaptation and updating of highly dynamic areas, the system introduces the Rate of Change Index (CRI) to estimate the degree of dynamic change for each raster and feature unit. CRI allows the system to identify areas that have undergone significant changes and quickly update these areas using dynamic masks, avoiding the computational overhead of global updates. Furthermore, CRI enables regionalized playback based on hot / cold zone determination methods, isolating dynamic areas from long-term maps and updating them locally and rapidly only.

[0105] 3. Enhance the dynamic detection capabilities based on single-line LiDAR. By introducing dynamic detection specialization into the LiDAR-map matching process, combined with residual segment clustering and semantic projection fusion, dynamic objects can be effectively distinguished from static environments. Specifically, by analyzing the temporal consistency of LiDAR measurement residuals, when dynamic objects are detected, these areas are downweighted or masked using a dynamic mask. Furthermore, semantic information (such as object recognition) can be used to discriminate dynamic objects, further improving the accuracy of dynamic detection.

[0106] 4. An adaptive update mechanism improves positioning accuracy and computational efficiency. By introducing subgraphs and regional playback, the system can apply different update frequencies to different regions. In highly dynamic regions, the system can achieve higher-frequency local reconstruction and updates, while in stable regions, it uses a lower update rate, thus making more efficient use of computational resources. Furthermore, the system uses dynamic mask guidance to further reduce unnecessary computational overhead and improve computational efficiency.

[0107] 5. Provides high-precision positioning and map updates. By combining dual-time-domain maps (AM / LM), rate of change index (CRI), and dynamic masking technologies, it can simultaneously achieve high-precision positioning and map updates. In highly dynamic areas, the system ensures positioning accuracy while avoiding map contamination through rapid local updates and dynamic masking. For stable areas, a lower update frequency ensures long-term map stability, thus achieving a balance between high precision and efficient updates. Attached Figure Description

[0108] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the specific embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0109] Figure 1 Flowchart of a dynamic mask localization method for AGVs with dual time-domain maps for dynamic warehousing. Detailed Implementation

[0110] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0111] like Figure 1 As shown, the AGV dual-temporal-domain map dynamic mask localization method for dynamic warehousing includes the following steps:

[0112] S1, Multi-sensor data acquisition

[0113] Real-time warehouse environment information is acquired through sensors, including wheel speed odometers, IMUs, single-line lidar, and cameras. The IMUs include gyroscopes and accelerometers, and the single-line lidar is a 2D-LiDAR.

[0114] S2, Dual-Time Domain Map Construction

[0115] Construct a dual-time-domain map with an anchor layer (AM) and a real-time layer (LM), where the anchor layer is used to store long-term stable structural information and the real-time layer is used to quickly respond to dynamic changes;

[0116] Calculate the rate of change index (CRI) for each map cell in a dual-time-domain map to determine the dynamics of the map cell.

[0117] The rate of change index (CRI) is calculated as follows:

[0118] S21, Observation count update, specifically:

[0119]

[0120] If a map cell is occupied (e.g., by a LiDAR hit), then Otherwise, it is 0, determined by the attenuation rate. Gradually weaken old information while incorporating new observations; if map units remain occupied, It will remain at a high value; if it is no longer occupied, Gradually decrease.

[0121]

[0122] If the map unit is free space (e.g., laser light penetrates without reflection), then Otherwise, it is 0, determined by the attenuation rate. Gradually weaken old information while incorporating new observations; if map units remain free space, It will remain at a high value; if map units are occupied, It gradually decays, thereby dynamically maintaining free space information and adapting to environmental changes (such as the disappearance of temporary obstacles).

[0123]

[0124] If the state of a map cell changes (e.g., from free space to occupied, or from occupied to free space), then Otherwise, it is 0, determined by the attenuation rate. Gradually weaken old information while incorporating new observations; if the state of map units changes drastically, It will remain at a high value; if the state of the map unit does not change, Gradually decaying, thus marking high dynamic areas (such as areas where moving obstacles frequently appear) for subsequent dynamic masking or local updates.

[0125] Among them Map unit; and These are the updated and unupdated values ​​of the map unit hit strength status, respectively. and These are the values ​​of the map unit's free space state after the update and the values ​​before the update, respectively. and These are the values ​​after the map unit state change intensity update and the values ​​before the update, respectively. , and The attenuation coefficient is, and ; This is an indicator function; it is denoted as 1 if the condition is true, and 0 otherwise. Used to determine whether a map cell is occupied; Used to determine whether a map cell is free space; Used to determine whether the state of a map cell has changed.

[0126] S22, Calculate the frequency of map cell occupancy. and uncertainty Specifically:

[0127]

[0128]

[0129] in, It is a constant greater than zero;

[0130] S23, the fusion calculation of the rate of change index (CRI), specifically:

[0131]

[0132] in, This is a 0-1 normalization function; and A fusion weight ≥ 0; To calculate the length of the statistical window; This is a semantic dynamic prior from the camera; , which is the geometric residual index.

[0133] S24, Determining the dynamic level of map units, for map units Its corresponding The larger the value, the higher its dynamic degree.

[0134] S3, Dynamic Mask Generation

[0135] Dynamic masks and their weights are generated based on CRI; stable and dynamic regions are determined based on the dynamic masks.

[0136] The calculation method for dynamic masks is as follows:

[0137]

[0138] in, For map units Dynamic mask; The threshold is determined dynamically.

[0139] like Then the map unit is a dynamic area; if If so, then the map unit is a stable region.

[0140] The calculation method for dynamic mask weights is as follows:

[0141]

[0142] in, for The weights; The sigmoid function is used to map input to... scope; The parameters used to control the slope of the Sigmoid curve.

[0143] The deviation between CRI and the threshold is converted into continuous weights by using the Sigmoid function, thus avoiding abrupt changes caused by hard masking.

[0144] When CRI(c)≫τ, the σ output is close to 1, w(c)≈0, indicating a strong reduction in weight;

[0145] When CRI(c)≪τ, the σ output is close to 0, w(c)≈1, and the high weight is maintained.

[0146] The larger the value, the more sensitive it is to changes in weight (steep sigmoid curve).

[0147] The smaller the value, the smoother the weight change (a smoother sigmoid curve).

[0148] Extremely high dynamic regions can be directly shielded through the masking function of dynamic masks; regions close to the threshold can be smoothly weighted through dynamic mask weighting to avoid performance fluctuations caused by hard truncation; the combination of the two can achieve a trade-off between stability and adaptability in dynamic environments, improving the robustness of localization, mapping or detection.

[0149] S4, Multi-sensor Fusion Positioning

[0150] S41 uses the initial pose estimation and motion constraints provided by the IMU and wheel speed odometer to perform precise localization by combining LiDAR-map matching with dynamic mask weights, and masks dynamic areas during LiDAR-map matching.

[0151] Step S41 is as follows:

[0152] In the 2D plane, with pose increment To optimize the variables, a LiDAR-map matching objective function with dynamic mask weights is constructed, specifically as follows:

[0153]

[0154] in, Represents a matrix transpose; ; and These represent the translation increment on the x-axis, the translation increment on the y-axis, and the rotation increment, respectively. The currently observed number The pose characteristics of each point; This is a pose transformation function used to transform... Transform from local coordinate system to global coordinate system; for The squared distance to the corresponding point in AM; The currently observed number Dynamic mask weights for each point; For robust kernel functions;

[0155]

[0156] Since AM is a long-term stable map layer, it stores reliable static structural information and maintains long-term consistency.

[0157] like If it is a high dynamic range region (high CRI), then Approaching zero reduces its influence;

[0158] like If it is a stable region (low CRI), then If the value is close to 1, its constraints are preserved.

[0159] By using pose increment optimization, combined with dynamic mask weights and robust kernel functions, accurate estimation of robot pose in dynamic environments is achieved.

[0160] The output of the objective function for LiDAR-map matching with dynamic mask weights is as follows:

[0161]

[0162] in For the optimized pose estimation, and These represent the position on the x-axis, the position on the y-axis, and the rotation angle, respectively.

[0163] and output covariance .

[0164] S42, in dynamic environments, if laser information is insufficient or matching fails, it switches to visual-inertial odometry as a supplement to ensure stable positioning output.

[0165] S43 uses error state Kalman filter (ESKF) fusion for localization.

[0166] Step S43 specifically includes:

[0167] S431, Initialization: Calculates gyroscope and accelerometer biases using the static window mean and sets the initial covariance; specifically:

[0168] Calculate the average angular velocity measured by the gyroscope while the object is stationary. and the average acceleration measured by the accelerometer Specifically:

[0169]

[0170]

[0171] in, This is the window length in a static state. and The first Gyroscope and accelerometer readings at each sampling point;

[0172] Gyroscope bias ;

[0173] Accelerometer bias ;

[0174] in, This is the rotation matrix from the world coordinate system to the body coordinate system; Let be the gravity vector in the world coordinate system, and ; It is the acceleration due to gravity;

[0175] initial covariance = ;

[0176] in, and These are the initial covariances of position, velocity, yaw angle, gyroscope bias, accelerometer bias, and gravitational acceleration, respectively.

[0177] S432, prediction, performs debiasing processing on IMU data based on gyroscope and accelerometer biases; predicts position, velocity, and attitude based on IMU data integration; and performs covariance prediction; specifically:

[0178] The formula for IMU data bias correction is as follows:

[0179] ; ;in and These are the debiased values ​​for angular velocity and acceleration, respectively. and These are the measured values ​​of angular velocity and acceleration, respectively.

[0180] Location prediction value Specifically:

[0181]

[0182] Speed ​​prediction value Specifically:

[0183]

[0184] Attitude prediction value Specifically:

[0185]

[0186] in, , and These are the position, velocity, and attitude readings obtained from the IMU in the world coordinate system, respectively. For time step; This is an exponential mapping function used to convert a rotation vector into a rotation matrix;

[0187] Attitude prediction value Specifically:

[0188]

[0189] in, This is the state transition matrix;

[0190] , A function for generating block diagonal matrices;

[0191] S433, Update, specifically includes wheel speed odometer observation update, LiDAR-map matching observation update and covariance update;

[0192] The wheel speed odometer observation update is as follows: the wheel speed odometer directly obtains the speed observation value under the machine system, constructs an observation residual model to compare the difference between the speed prediction value and the speed observation value; the sensitivity of speed and attitude to the observation is quantified by calculating the Jacobian matrix, and finally the speed and attitude are dynamically corrected by using Kalman gain.

[0193] The LiDAR-map matching observation update is as follows: global pose correction is obtained by registering the LiDAR point cloud with the pre-built map, forming the observation residuals of position and attitude; the influence of position and attitude on point cloud matching error is quantified by the Jacobian matrix, and spatial geometric constraints are incorporated into the filtering framework to correct the robot's global pose.

[0194] The covariance update is specifically as follows: the state covariance matrix is ​​dynamically adjusted based on the Kalman gain, and the local motion constraints of the wheel speed sensor and the global geometric constraints of LiDAR are fused; the uncertainty of position, velocity and attitude is reduced by weighted fusion of the predicted covariance and the observation noise.

[0195] S434, Reset and tangent space reprojection, projects the updated covariance onto the new tangent space.

[0196] S5, Subgraph Dynamic Management

[0197] A regionalized playback strategy enables rapid updates to hot zones and stable maintenance of cold zones.

[0198] Step S5 specifically includes:

[0199] S51, Hot Zone Determination, for any region ,like And the duration exceeds If the temperature is high, it is considered a hot zone; otherwise, it is considered a cold zone.

[0200] S52, Playback and Write, creates or resets local subgraphs for hotspots in LM. The area corresponding to the AM hot zone remains frozen;

[0201] S53, Isolation Strategy: When optimizing and matching AGV positioning, constraints are first established on the AM (Advanced Positioning Component); when the AM is sparse or mismatched, [the strategy is implemented by...]. It provides local auxiliary matching and performs weight reduction processing through dynamic mask weights;

[0202] S54, Promotion / Demotion Mechanism, when And the duration exceeds the duration At that time, from Select a stable boundary or skeleton feature to promote to AM; if the temperature rises again, reverse and degrade back to LM.

[0203] S55, adaptive gating, relaxes the upper limit of gating in hot zones to reduce false negatives; tightens and relaxes the upper limit of gating in cold zones to prevent false loop closure.

[0204] in and This represents the temperature threshold.

[0205] S6, Dynamic Detection Specialization for Single-Line Laser

[0206] Based on temporal consistency analysis, semantic projection fusion, and occlusion chain tracking, dynamic objects are identified and CRI is updated.

[0207] Step S6 specifically includes:

[0208] S61, Timing Consistency Analysis

[0209] For the Compare the measured values ​​of the single-line lidar measurement line. Compared with the optimized pose estimation The difference yields the residual. Large residual continuous segments are classified as dynamic objects.

[0210] right Update, specifically ,in It is a non-negative increasing function;

[0211] Assume continuous residual segments The set of map units covered by its projection is ,calculate The CRI increment is updated as follows:

[0212]

[0213] in, For residual evidence weights, for The length; if If the CRI increases significantly, it is determined to be a dynamic object;

[0214] S62, Semantic Projection Fusion

[0215] Projecting the camera detection frame onto the ground yields a polygon. For each map unit Calculate its relationship with The overlap rate is as follows:

[0216]

[0217] in The overlap rate calculation function is used to determine regions with high overlap rates as semantically dynamic regions.

[0218] By directly marking the areas where dynamic objects may appear using semantic information (such as people and vehicles), the reliance on geometric residuals is reduced, and the detection capability for small or fast-moving objects is improved.

[0219] S63, Obstruction Chain Tracking

[0220] For a single-line lidar survey line, if the free space it traverses is occupied, then backtrack along the path of that survey line to find the continuous map cell chain that transitions from free space to occupied space. Then, for that continuous map cell chain... The weights;

[0221] S64, update CRI, and update the CRI. , and Substitute into step S23, for Update.

[0222] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A dynamic mask localization method for AGVs with dual time-domain maps for dynamic warehousing, characterized in that, Includes the following steps: S1, Multi-sensor data acquisition Real-time warehouse environment information is acquired through sensors, including wheel speed odometers, IMUs, single-line lidar, and cameras. The IMUs include gyroscopes and accelerometers, and the single-line lidar is a 2D-LiDAR. S2, Dual-Time Domain Map Construction Construct a dual-time-domain map with an anchor layer (AM) and a real-time layer (LM), where the anchor layer is used to store long-term stable structural information and the real-time layer is used to quickly respond to dynamic changes; Calculate the rate of change index (CRI) for each map cell in a dual-time-domain map to determine the dynamics of the map cell; The rate of change index (CRI) is calculated as follows: S21, Observation count update, specifically: ； ； ； Among them Map unit; and These are the updated and unupdated values ​​of the map unit hit strength status, respectively. and These are the updated and unupdated values ​​of the map unit's free space state, respectively. and These are the updated values ​​of the map unit state change intensity and the values ​​before the update, respectively. , and The attenuation coefficient is, and ; This is an indicator function; it is denoted as 1 if the condition is true, and 0 otherwise. Used to determine whether a map cell is occupied; Used to determine whether a map cell is free space; Used to determine whether the state of a map cell has changed; S22, Calculate the frequency of map cell occupancy. and uncertainty Specifically: ； ； in, It is a constant greater than zero; S23, the fusion calculation of the rate of change index (CRI), specifically: ； in, This is a 0-1 normalization function; and A fusion weight ≥ 0; To calculate the length of the statistical window; This is a semantic dynamic prior from the camera; , is the geometric residual index; S24, Determining the dynamic level of map units, for map units Its corresponding The larger the value, the higher its dynamic degree; S3, Dynamic Mask Generation Dynamic masks and their weights are generated based on CRI; stable and dynamic regions are determined based on the dynamic masks; The calculation method for dynamic masks is as follows: ； in, For map units Dynamic mask; The threshold is determined dynamically. like Then the map unit is a dynamic area; if If so, the map unit is a stable region; The calculation method for dynamic mask weights is as follows: ； in, for The weights; The sigmoid function is used to map input to... scope; Parameters used to control the slope of the Sigmoid curve; S4, Multi-sensor Fusion Positioning S41 uses the initial pose estimation and motion constraints provided by the IMU and wheel speed odometer to perform precise localization by combining LiDAR-map matching with dynamic mask weights, and masks the dynamic area during LiDAR-map matching. S42, in dynamic environments, if laser information is insufficient or matching fails, switch to visual-inertial odometry as a supplement to ensure stable positioning output; S43, localization is achieved through ESKF (Error State Kalman Filter); S5, Subgraph Dynamic Management A regional playback strategy enables rapid updates to hot zones and stable maintenance of cold zones; S6, Dynamic Detection Specialization for Single-Line Laser Based on temporal consistency analysis, semantic projection fusion, and occlusion chain tracking, dynamic objects are identified and CRI is updated.

2. The AGV dual-temporal-domain map dynamic mask positioning method for dynamic warehousing according to claim 1, characterized in that, Step S41 specifically involves: In a 2D plane, with pose increments To optimize the variables, a LiDAR-map matching objective function with dynamic mask weights is constructed, specifically as follows: ； in, Represents a matrix Transpose of; ; and These represent the translation increment on the x-axis, the translation increment on the y-axis, and the rotation increment, respectively. The currently observed number The pose characteristics of each point; This is a pose transformation function used to transform... Transform from local coordinate system to global coordinate system; for The squared distance to the corresponding point in AM; The currently observed number Dynamic mask weights for each point; For robust kernel functions; The output of the objective function for LiDAR-map matching with dynamic mask weights is as follows: ； in For the optimized pose estimation, and These represent the position on the x-axis, the position on the y-axis, and the rotation angle, respectively. and output covariance .

3. The AGV dual-temporal-domain map dynamic mask positioning method for dynamic warehousing according to claim 2, characterized in that, Step S43 specifically includes: S431, Initialization: Calculates gyroscope and accelerometer biases using the static window mean and sets the initial covariance; specifically: Calculate the average angular velocity measured by the gyroscope while the object is stationary. and the average acceleration measured by the accelerometer Specifically: ； ； in, This is the window length in a static state. and The first Gyroscope and accelerometer readings at each sampling point; Gyroscope bias ; Accelerometer bias ; in, This is the rotation matrix from the world coordinate system to the body coordinate system; Let be the gravity vector in the world coordinate system, and ; It is the acceleration due to gravity; initial covariance = ; in, and These are the initial covariances of position, velocity, yaw angle, gyroscope bias, accelerometer bias, and gravitational acceleration, respectively. S432, prediction, performs debiasing processing on IMU data based on gyroscope and accelerometer biases; predicts position, velocity, and attitude based on IMU data integration; and performs covariance prediction; specifically: The formula for IMU data bias correction is as follows: ; ;in and These are the debiased values ​​for angular velocity and acceleration, respectively. and These are the measured values ​​of angular velocity and acceleration, respectively. Location prediction value Specifically: ； Speed ​​prediction value Specifically: ； Attitude prediction value Specifically: ； in, , and These are the position, velocity, and attitude readings obtained from the IMU in the world coordinate system, respectively. For time step; This is an exponential mapping function used to convert a rotation vector into a rotation matrix; Attitude prediction value Specifically: ； in, This is the state transition matrix; , A function for generating block diagonal matrices; S433, Update, specifically includes wheel speed odometer observation update, LiDAR-map matching observation update and covariance update; The wheel speed odometer observation update is as follows: the wheel speed odometer directly obtains the speed observation value under the machine system, constructs an observation residual model to compare the difference between the speed prediction value and the speed observation value; the sensitivity of speed and attitude to the observation is quantified by calculating the Jacobian matrix, and finally the speed and attitude are dynamically corrected by using Kalman gain. The LiDAR-map matching observation update is as follows: global pose correction is obtained by registering the LiDAR point cloud with the pre-built map, forming the observation residuals of position and attitude; the influence of position and attitude on point cloud matching error is quantified by the Jacobian matrix, and spatial geometric constraints are incorporated into the filtering framework to correct the robot's global pose. The covariance update is specifically as follows: the state covariance matrix is ​​dynamically adjusted based on the Kalman gain, and the local motion constraints of the wheel speed meter and the global geometric constraints of LiDAR are fused; the uncertainty of position, velocity and attitude is reduced by weighted fusion of the predicted covariance and the observation noise. S434, Reset and tangent space reprojection, projects the updated covariance onto the new tangent space.

4. The AGV dual-temporal-domain map dynamic mask positioning method for dynamic warehousing according to claim 1, characterized in that, Step S5 specifically includes: S51, Hot Zone Determination, for any region ,like And the duration exceeds If the temperature is high, it is considered a hot zone; otherwise, it is considered a cold zone. S52, Playback and Write, creates or resets local subgraphs for hotspots in LM. The area corresponding to the AM hot zone remains frozen; S53, Isolation Strategy: When optimizing and matching AGV positioning, constraints are first established on the AM (Advanced Positioning Component); when the AM is sparse or mismatched, [the strategy is implemented by...]. It provides local auxiliary matching and performs weight reduction processing through dynamic mask weights; S54, Promotion / Demotion Mechanism, when And the duration exceeds the duration At that time, from Select a stable boundary or skeleton feature to promote to AM; if the temperature rises again, reverse and degrade back to LM. S55, adaptive gating, relaxes the upper limit of gating in hot zones to reduce false negatives; tightens and relaxes the upper limit of gating in cold zones to prevent false loop closure. in and This represents the temperature threshold.

5. The AGV dual-temporal-domain map dynamic mask positioning method for dynamic warehousing according to claim 2, characterized in that, Step S6 specifically includes: S61, Timing Consistency Analysis For the Compare the measured values ​​of the single-line lidar measurement line. Compared with the optimized pose estimation The difference yields the residual. Large residual continuous segments are classified as dynamic objects. right Update, specifically ,in It is a non-negative increasing function; Assume continuous residual segments The set of map units covered by its projection is ,calculate The CRI increment is updated as follows: ； in, For residual evidence weights, for The length; if If the CRI increases significantly, it is determined to be a dynamic object; S62, Semantic Projection Fusion Projecting the camera detection frame onto the ground yields a polygon. For each map unit Calculate its relationship with The overlap rate is as follows: ； in The overlap rate calculation function is used to determine regions with high overlap rates as semantically dynamic regions. S63, Obstruction Chain Tracking For a single-line lidar survey line, if the free space it traverses is occupied, then backtrack along the path of that survey line to find the continuous map cell chain that transitions from free space to occupied space. Then, for that continuous map cell chain... The weights; S64, update CRI, and update the CRI. , and Substitute into step S23, for Update.

Images