Inspection system and method for container yard

Abstract

The invention relates to an inspection system and method for a container yard, and belongs to the technical field of map construction, and the system comprises an unmanned aerial vehicle terminal subsystem which comprises a plurality of unmanned aerial vehicles and is used for obtaining and obtaining an odometer track based on a pose transformation matrix and inertial measurement data between adjacent frames of a target unmanned aerial vehicle, loopback detection is carried out on the odometer trajectory, and global graph optimization is carried out to obtain a global optimization trajectory; the operation management and control system is in communication connection with the unmanned aerial vehicle, and is used for acquiring plan data of the target container yard, constructing a harbor district operation situation map of the target container yard based on a global map corresponding to the odometer track and sensor data, and sending the situation map to the unmanned aerial vehicle; and comparing the operation situation map with a plan map corresponding to the plan data, carrying out anomaly detection and graded early warning to obtain a risk level, and carrying out path re-planning based on the risk level to obtain an optimal inspection path of the unmanned aerial vehicle. The inspection efficiency of the storage yard in the harbor district is improved, and the automatic management capability of the harbor district is improved.

Description

Technical Field

[0001] This invention relates to the field of map technology, and in particular to an inspection system and method for container yards. Background Technology

[0002] As ports become increasingly automated, higher demands are being placed on real-time inspection and intelligent management of operational areas. Large port container yards, as key nodes in logistics hubs, require real-time status awareness in a dynamic environment for operational safety and scheduling optimization.

[0003] Existing drone systems still have the following shortcomings when facing port yard scenarios: First, the stacking layout of containers in the yard often changes, which can easily cause errors in the drone's positioning and mapping, affecting navigation stability; second, flight path planning often lacks sufficient consideration of the distribution of obstacles on site, which can easily create blind spots or repeated inspections, reducing operational efficiency; third, existing systems have limited capabilities in identifying operational status, making it difficult to detect problems such as incorrect container numbering and equipment malfunctions in a timely manner, and lacking effective early warning mechanisms and feedback adjustment capabilities.

[0004] Therefore, there is an urgent need for a drone-based intelligent inspection system specifically designed for the environmental characteristics of large port yards, which can improve the accuracy and stability of mapping and positioning, have flexible path planning, efficient anomaly identification and multi-drone coordination capabilities, and improve overall inspection efficiency and the level of automation in port management. Summary of the Invention

[0005] In view of this, it is necessary to provide a container yard inspection system and method to improve the efficiency of port yard inspection and achieve the goal of automated port management.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a container yard inspection system, comprising: The UAV terminal subsystem includes several UAVs, which are used to acquire and obtain the pose transformation matrix between adjacent frames of the target UAV based on the sensor data of the target UAV, obtain the odometry trajectory based on the pose transformation matrix and inertial measurement data, perform loop closure detection on the odometry trajectory and perform global graph optimization to obtain the globally optimized trajectory. The sensor data of the target UAV includes: lidar point cloud, image data and IMU sensor data. The operation control system communicates with drones to acquire planned data of the target container yard. Based on the global map and sensor data corresponding to the globally optimized trajectory, it constructs a port operation status map of the target container yard. The operation status map is compared with the planned map corresponding to the planned data, and anomaly detection and graded early warning are performed to obtain the risk level. Based on the risk level, the path is replanned to obtain the optimal inspection path of the drone. The planned data includes: planned stacking location, expected number, and task status label.

[0007] In one possible implementation, the task status label includes: When the task at the planned stacking location is a loading task, the task status labels include: not started, stacked and waiting to be loaded, loading, loaded and departed. When the planned task at the storage location is a container pickup task, the task status labels include: waiting to arrive, waiting to be picked up, in progress, and already departed.

[0008] In one possible implementation, the unmanned aerial vehicle (UAV) terminal subsystem is further configured to: In the odometer trajectory, select historical keyframe point clouds that satisfy the preset spatial distance and preset time interval with the current odometer pose, and construct a set of loop closure candidate point clouds based on the historical keyframe point clouds; Calculate the difference between the current point cloud in the same slice and the historical point cloud in the set of loop closure candidate point clouds; The historical point cloud with a difference value below a preset threshold is used to form a point cloud pair with the current point cloud, and the relative pose transformation between the point cloud pairs is calculated. The relative pose transformation is added as a lapsing constraint to the preset factor graph optimization model to optimize the odometer trajectory and obtain the globally optimized trajectory.

[0009] Secondly, the present invention also provides a method for inspecting container yards, comprising: The global optimized trajectory and sensor data are acquired. The sensor data includes: the lidar point cloud of the target UAV, the image data of the target UAV, and the IMU sensor data of the target UAV. The global optimized trajectory is obtained by the target UAV based on the sensor data to obtain the pose transformation matrix between adjacent frames of the target UAV, the target UAV based on the pose transformation matrix and the inertial measurement data to obtain the odometry trajectory, and the target UAV performs loop closure detection and global graph optimization on the odometry trajectory. Obtain the planning data for the target container yard, which includes: planned stack location, expected number, and task status label; A port operation status map of the target container yard is constructed based on map and sensor data corresponding to the globally optimized trajectory; By performing anomaly detection and graded early warning on the operational status map and the corresponding planning map based on the planning data, the risk level can be obtained; Based on the risk level, path replanning is performed to obtain the optimal inspection path for the UAV.

[0010] In one possible implementation, obtaining the pose transformation matrix between adjacent frames of the target UAV based on sensor data includes: The relative height of each slice to the ground plane is determined based on the height of the container. Slice the container layer in the target yard at the current moment into multiple layers, taking each layer of containers as a layer; Perform layer-by-layer matching on the point cloud of each slice at the current time to obtain the pose transformation matrix between adjacent frames of the point cloud of each slice.

[0011] In one possible implementation, the odometry trajectory is obtained based on the pose transformation matrix and IMU sensor data, including: A kinematic model is built based on historical data from the IMU sensor. The IMU sensor data from the previous moment is input into the kinematic model to predict the UAV pose at the current moment. The target UAV's relative pose at the current moment is determined based on the pose transformation matrix and the UAV's pose at the current moment; the target UAV's odometry trajectory is determined based on the target UAV's relative pose at the current moment.

[0012] In one possible implementation, the step of performing loop closure detection and global graph optimization on the odometer trajectory to obtain a globally optimized trajectory includes: In the odometer trajectory, select historical keyframe point clouds that satisfy the preset spatial distance and preset time interval with the current odometer pose, and construct a set of loop closure candidate point clouds based on the historical keyframe point clouds; Calculate the difference between the current point cloud in the same slice and the historical point cloud in the set of loop closure candidate point clouds; The historical point cloud with a difference value below a preset threshold is used to form a point cloud pair with the current point cloud, and the relative pose transformation between the point cloud pairs is calculated. The relative pose transformation is added as a lapsing constraint to the preset factor graph optimization model to optimize the odometer trajectory and obtain the globally optimized trajectory.

[0013] In one possible implementation, the expression for the globally optimized trajectory is:

[0014] In the formula, This indicates the globally optimized trajectory of the drone. Denotes the odometer edge set, This indicates the odometer error term. Indicates inter-frame odometer measurement. Indicate its covariance, This represents the closure error term. Indicate its covariance, This represents the relative pose transformation between point cloud pairs. Indicates the first Pose variables of each pose node Indicates the first Pose variables of each pose node Indicates the first Pose variables of each pose node This indicates the odometer trajectory.

[0015] In one possible implementation, the expression for the preset GTSAM factor graph optimization model is:

[0016] In the formula, This represents the optimal state estimate obtained after optimization of the factor graph model. Odometer factor The corresponding residual function, Odometer factor The observed covariance matrix, Represents the cyclic factor The observed covariance matrix, This represents the target UAV pose node. For odometer factor, As a cyclic factor, This represents the frame of the point cloud at the current moment. This represents the frame in the candidate point cloud set that matches the historical point cloud at the current moment.

[0017] In one possible implementation, the step of performing anomaly detection and graded early warning on the plan map corresponding to the operational status map and the plan data to obtain the risk level includes: The spatial structure similarity analysis of the operational status map and the corresponding planning map based on the planning data is used to obtain similarity values; The consistency of the operational status map and the corresponding planning map based on the numbering confidence is determined to obtain the numbering confidence value. The stability of the operational status diagram and the corresponding planning diagram based on the planning data are assessed to obtain the stability value. The joint assessment function is determined based on the similarity value, the number confidence value, and the stability value. The risk level is then determined based on the value of the joint assessment function.

[0018] The beneficial effects of this invention are as follows: This invention provides a container yard inspection system, including a UAV terminal subsystem and an operation control system. The UAV terminal subsystem includes several UAVs, used to acquire and obtain the pose transformation matrix between adjacent frames of the target UAV based on sensor data, obtain the odometer trajectory based on the pose transformation matrix and inertial measurement data, perform loop closure detection on the odometer trajectory and perform global graph optimization to obtain a globally optimized trajectory. The sensor data of the target UAV includes: lidar point cloud, image data, and IMU sensor data. The addition of a loop closure detection mechanism to the UAV terminal subsystem improves the global graph optimization in long-term port inspection tasks. Figure 1 Consistency and navigation stability provide a high-confidence trajectory basis for anomaly localization and task scheduling. The operation control system, communicating with the UAV, acquires planned data from the target container yard. Based on the global map corresponding to the globally optimized trajectory and sensor data, it constructs a port operation status map of the target container yard. The operation status map is compared with the planned map corresponding to the planned data, and anomaly detection and graded early warning are performed to obtain the risk level. Based on the risk level, path replanning is performed to obtain the optimal inspection path for the UAV. The planned data includes: planned stacking positions, expected numbers, and task status tags, thereby enabling automated early warning. This invention improves navigation accuracy by adding a loop closure detection mechanism to the trajectory generation of the UAV terminal subsystem, enabling the UAV to be scheduled more quickly according to the task, improving the efficiency of port yard inspection, and further enhancing the port's automated management capabilities by classifying risks based on the generated path. Attached Figure Description

[0019] Figure 1 A system architecture diagram of an embodiment of a container yard inspection system provided by the present invention; Figure 2 A front view of a drone in one embodiment of a container yard inspection system provided by the present invention; Figure 3 A side view of a system according to an embodiment of the container yard inspection system provided by the present invention; Figure 4 A top view of an embodiment of a container yard inspection system provided by the present invention; Figure 5 This invention provides an embodiment of a container yard inspection system, including a block diagram of UAV pose and map generation. Figure 6 A schematic diagram of a slice indicator on a container, representing an embodiment of a container yard inspection system provided by the present invention; Figure 7 An anomaly detection flowchart of an embodiment of a container yard inspection system provided by the present invention; Figure 8 A flowchart illustrating an embodiment of a container yard inspection method provided by the present invention; The components are: 1-propeller, 2-fuselage, 3-frame, 4-edge computing module, 5-IMU, 6-RTK receiver, 7-LiDAR, 8-camera and thermal imager, 9-landing gear. Detailed Implementation

[0020] 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 a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0022] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0023] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] This invention provides a container yard inspection system and method, which are described below.

[0025] Figure 1 A system architecture diagram of an embodiment of the container yard inspection system provided by the present invention is shown below. Figure 1 As shown, the container yard inspection system 100 includes: The UAV terminal subsystem 101 includes several UAVs, which are used to acquire and obtain the pose transformation matrix between adjacent frames of the target UAV based on the sensor data of the target UAV, obtain the odometry trajectory based on the pose transformation matrix and inertial measurement data, perform loop closure detection on the odometry trajectory and perform global graph optimization to obtain the globally optimized trajectory. The sensor data of the target UAV includes: lidar point cloud, image data and IMU sensor data. The drone is equipped with an IMU sensor, LiDAR, camera and thermal imager, and edge processing module.

[0026] The operation control system 102 communicates with the UAV to obtain the planned data of the target container yard. Based on the global map corresponding to the global optimized trajectory and sensor data, it constructs the port operation status map of the target container yard. It compares the operation status map with the planned map corresponding to the planned data and performs anomaly detection and graded early warning to obtain the risk level. Based on the risk level, it performs path replanning to obtain the optimal inspection path of the UAV. The planned data includes: planned stacking location, expected number and task status label. The operations management system serves as the command center for the entire system, responsible for tasks such as configuration, path planning, multi-machine scheduling, and operations status identification. Based on high-precision maps and positioning data provided by Slice-SLAM, the system dynamically constructs a port operations situation map, integrating sensor data and scheduling information to form a multi-dimensional status map including the structure of the operations area, container numbers, and equipment status. Simultaneously, the platform can identify structural anomalies and path risks in real time based on the database map, and through a path feedback adjustment mechanism, form a closed-loop task management system from "map creation - identification - feedback - scheduling".

[0027] Compared with existing technologies, this embodiment provides a container yard inspection system, including a UAV terminal subsystem and an operation control system. The UAV terminal subsystem includes several UAVs, used to acquire and obtain the pose transformation matrix between adjacent frames of the target UAV based on the sensor data of the target UAV, obtain the odometer trajectory based on the pose transformation matrix and inertial measurement data, perform loop closure detection on the odometer trajectory and perform global graph optimization to obtain a globally optimized trajectory. The sensor data of the target UAV includes: lidar point cloud, image data and IMU sensor data. A loop closure detection mechanism is added to the UAV terminal subsystem, improving the global graph optimization in long-term port inspection tasks. Figure 1Consistency and navigation stability provide a high-confidence trajectory basis for anomaly localization and task scheduling. The operation control system, communicating with the UAV, acquires planned data from the target container yard. Based on the global map corresponding to the globally optimized trajectory and sensor data, it constructs a port operation status map of the target container yard. The operation status map is compared with the planned map corresponding to the planned data, and anomaly detection and graded early warning are performed to obtain the risk level. Based on the risk level, path replanning is performed to obtain the optimal inspection path for the UAV. The planned data includes: planned stacking positions, expected numbers, and task status tags, thereby enabling automated early warning. This invention improves navigation accuracy by adding a loop closure detection mechanism to the trajectory generation of the UAV terminal subsystem, enabling the UAV to be scheduled more quickly according to the task, improving the efficiency of port yard inspections, and further enhancing the port's automated management capabilities by classifying risks based on the generated path.

[0028] It is understood that the application scenarios of this invention are not limited to container yards, but also applicable to other scenarios that require drone inspections.

[0029] In a specific embodiment of the present invention, multiple quadcopter drone platforms are used, equipped with high-performance LiDAR, inertial measurement unit (IMU), real-time dynamic differential (RTK), high-definition camera and thermal imager and other multi-source sensors to achieve accurate attitude estimation, real-time 3D mapping and image inspection.

[0030] The drone is equipped with an edge computing module, which can perform image recognition, SLAM mapping and status assessment locally, reducing communication load.

[0031] During flight, the Slice-SLAM mapping and localization algorithm proposed in this invention can provide robust mapping capabilities in environments with varying container stacking heights, supporting subsequent mission planning and anomaly identification.

[0032] In a specific embodiment of this invention, to meet the autonomous positioning and inspection needs of unmanned aerial vehicles (UAVs) in complex environments such as large-scale, highly dynamic, and structurally repetitive port yards, this invention proposes a Slice-SLAM method based on multi-layer tiled maps, combining multi-source sensor information such as LiDAR, IMU, and RTK. This method is suitable for large port yard environments with regular container heights and significant planar features. The method provides high-precision mapping and positioning capabilities, and the specific process is as follows: Figure 5 As shown.

[0033] In some embodiments of the present invention, the unmanned aerial vehicle (UAV) terminal subsystem is further configured to: The relative height of each slice to the ground plane is determined based on the height of the container. Slice the container layer in the target yard at the current moment into multiple layers, taking each layer of containers as a layer; Perform layer-by-layer matching on the point cloud of each slice at the current time to obtain the pose transformation matrix between adjacent frames of the point cloud of each slice.

[0034] In a specific embodiment of the present invention, the UAV terminal subsystem is further used for tile map construction and matching, specifically: Multi-slice extraction method based on prior information The core objective of point cloud registration is to maximize the number of corresponding points between matching point clouds and to solve for the optimal rigid body transformation (translation and rotation) parameters to minimize the difference between the overall observed point and the true position. Assume the coordinates of a point in the current frame's point cloud... First, we need to find the nearest neighbor set of points in the point cloud between adjacent frames: Traditional algorithms search the entire frame of point cloud, considering three degrees of freedom. This invention imposes constraints on the height dimension, limiting the search range to a specific height interval between adjacent frames, effectively eliminating interference from irrelevant point clouds and improving registration efficiency and robustness.

[0035] In a port yard environment, the map is sliced ​​into multiple layers, with each layer of containers representing a slice. At the same time, the thickness of each slice is minimized as much as possible, reducing the slice area to near the top surface of the containers, thereby approximating the point cloud of the top surface of a single layer of containers.

[0036] The process of extracting the top surface of the container is as follows Figure 6 As shown. Containers are stacked horizontally on the yard floor, with the top surface of each individual container parallel to the ground. In large port areas, the open ground area is large, and the ground features are very obvious. By fitting the gravity direction using IMU data and using it as the target normal vector, the plane farthest from the radar center point is extracted as the ground. The ground equation is fitted using the general plane equation as follows: (2-1) Since the standard container size is fixed, the container height is utilized. This prior information allows for the rapid determination of the relative heights of different slice planes to the ground plane during the slicing operation.

[0037] Multi-layer slice matching method based on inter-layer reuse After extracting slices containing point clouds of the upper surfaces of each container layer, invalid slices with insufficient planar information are filtered out based on the number of valid points contained in the slice point cloud. Layer-by-layer matching is then performed on the slice point clouds of each layer. The goal of the matching is to solve for the optimal rigid transformation between the same-layer slice point clouds of adjacent frames. ,in It is a rotation matrix. It is a translation vector. In this process, the previous slice matching result... Used as the current layer slice match Initial value: (2-2) This initial value reuse strategy effectively shortens the convergence time of the matching algorithm.

[0038] To improve robustness, this invention designs an inter-layer consistency check. When the difference between consecutive inter-layer matching results exceeds a preset threshold, it is considered that there may be a mismatch. The system constructs a confidence scoring mechanism based on point cloud registration residuals and the number of valid points. The matching result of each layer is assigned a score based on factors such as its error and occlusion degree, and the matching results are filtered according to the score. For example, when the confidence is low or the number of points is insufficient, the system will reduce the influence weight of the layer's result on the final pose estimation, or directly exclude the interfering data of that layer.

[0039] In some embodiments of the present invention, the unmanned aerial vehicle (UAV) terminal subsystem is further configured to: A kinematic model is built based on historical data from the IMU sensor. The IMU sensor data from the previous moment is input into the kinematic model to predict the UAV pose at the current moment. The target UAV's relative pose at the current moment is determined based on the pose transformation matrix and the UAV's pose at the current moment; the target UAV's odometry trajectory is determined based on the target UAV's relative pose at the current moment.

[0040] In a specific embodiment of the present invention, in odometry pose estimation, the present invention utilizes multi-layer slice registration results. This serves as the initial state of the system, used to initialize the generalized iterative nearest point of the overall point cloud. Simultaneously, a kinematic model is constructed on the IMU data in parallel, with the IMU state vector as follows: (2-3) (2-4) In the formula The IMU sampling time interval Rotation matrix derived from quaternions These are the raw acceleration and angular velocity measurements from the IMU. It is the acceleration due to gravity. This is the quaternion multiplication operator. For the exponential mapping from Lie algebras to Lie groups, The noise level is zero. Based on this, the system uses LiDAR registration results as observation input and performs joint optimization and updates with IMU predictions. The final output is the UAV's current spatial position and attitude information.

[0041] In a container yard environment, the area includes empty ground and lane lines. When drones fly over these sparsely characterized empty areas, it leads to insufficient effective data in the sliced ​​point cloud, resulting in reduced mapping accuracy. In this situation, odometry calculations are temporarily maintained solely by relying on motion state predictions from IMU data. (2-5) The IMU measurements (acceleration and angular velocity) are always taken at time k+1. It is process noise. It can prevent large drifts in the mapping process due to a lack of effective constraints.

[0042] In some embodiments of the present invention, the unmanned aerial vehicle (UAV) terminal subsystem is further configured to: In the odometer trajectory, select historical keyframe point clouds that satisfy the preset spatial distance and preset time interval with the current odometer pose, and construct a set of loop closure candidate point clouds based on the historical keyframe point clouds; The point cloud at the current moment is matched with the historical point cloud in the loop closure candidate point cloud set to obtain the slice matching result, and the difference value between the current point cloud slice and the historical point cloud slice is calculated. Historical point clouds with differences below a preset threshold are identified as having loop closures, and the relative pose transformation between the current point cloud and the corresponding historical point cloud is calculated based on the slice matching results. The relative pose transformation is added as a lapsing constraint to the preset factor graph optimization model to optimize the odometer trajectory and obtain the globally optimized trajectory.

[0043] Regional division based on scene distribution In a container yard setting, container stacking is highly random, and LiDAR has a narrow field of view. Even when returning to a historical location during flight, the scanned point cloud rarely maintains a large area overlap with historical scan data. Performing overall matching between keyframe point clouds containing high-dynamic points easily leads to loop closure failures. By further utilizing slicing operations, the keyframe point cloud can be divided into a yard area and a high-dynamic area. The yard area has a regular structure and is suitable for loop closure detection, while the high-dynamic area, containing crane and vehicle movement, experiences significant matching interference and should only be used as a reference or excluded from areas where cranes are located.

[0044] Loop Optimization Based on Two-Stage Registration During the mapping process, the nearest keyframes are continuously searched, and slices are extracted and matched between two keyframes. The differences between the matching results of slices in the same layer of all stockpile areas are calculated. Matching results with differences below a threshold are grouped together. , Number of slices within a group: (2-6) Select the group with the highest number of slices. The group with the highest voting score was selected, and the transformation matrix with the highest confidence score was used as the initial value for loop closure detection. Fine-grained point cloud matching was then performed on two keyframes from which high-dynamic-range points were removed, ultimately yielding the overall fine-grained matching result. Add it as a closure edge to the GTSAM factor graph to construct a GTSAM factor graph optimization model: (2-7) In the formula, This represents the optimal state estimate obtained after factor graph optimization. Odometer factor The corresponding residual function, Odometer factor The observed covariance matrix, Represents the cyclic factor The observed covariance matrix, This represents the target UAV pose node. For odometer factor, As a cyclic factor, This represents the frame of the point cloud at the current moment. This represents the frame in the candidate point cloud set that matches the historical point cloud at the current moment.

[0045] The overall trajectory of the system ( Keyframe Global optimization is performed on the pose of the object. (2-8) In the formula, This indicates the globally optimized trajectory of the drone. Denotes the odometer edge set, This indicates the odometer error term. Indicates inter-frame odometer measurement. Indicate its covariance, It is the hysteresis error term. It is its covariance. This represents the relative pose transformation between point cloud pairs. Indicates the first Pose variables of each pose node Indicates the first Pose variables of each pose node Indicates the first Pose variables of each pose node This indicates the odometer trajectory.

[0046] The optimization process minimizes the weighted sum of squares of all error terms, thus obtaining a smooth and globally consistent trajectory estimate.

[0047] The loop factor graph optimization mechanism significantly improves the overall performance of long-term port inspection tasks. Figure 1 Consistency and navigation stability provide a high-confidence trajectory basis for anomaly localization and task scheduling.

[0048] In some embodiments of the present invention, the operation control system is further configured to perform spatial structure similarity analysis on the operation status map and the plan map corresponding to the plan data to obtain a similarity value; The consistency of the operational status map and the corresponding planning map based on the numbering confidence is determined to obtain the numbering confidence value. The stability of the operational status diagram and the corresponding planning diagram based on the planning data are assessed to obtain the stability value. The joint assessment function is determined based on the similarity value, the number confidence value, and the stability value. The risk level is then determined based on the value of the joint assessment function.

[0049] In some embodiments of the present invention, the task status label includes: When the task at the planned stacking location is a loading task, the task status labels include: not started, stacked and waiting to be loaded, loading in progress, loaded and departed. When the planned task at the storage location is a container pickup task, the task status labels include: waiting to arrive, waiting to be picked up, in the process of picking up the container, and already departed.

[0050] In some embodiments of the present invention, to ensure continuous and stable SLAM mapping and localization of the UAV in the complex, dynamic, and interference-prone working environment of the port area, an abnormal scene detection and fault tolerance mechanism is introduced. Through multi-source data consistency judgment, point cloud sparsity analysis, and time synchronization deviation estimation, potential failures during system operation are dynamically identified, and the operating state is automatically switched according to the anomaly level to ensure stable flight of the UAV and uninterrupted core functions.

[0051] In IMU drift detection, the UAV constructs a position residual term between the IMU pre-integration trajectory and the LiDAR point cloud odometry trajectory. When the residual is in continuous... If all frames exceed the set threshold, it will be determined that the current IMU has significant drift, and it will enter laser-dominated mode, reducing the weight of IMU observation data.

[0052] To address the issue of sparse point clouds caused by LiDAR occlusion or flight into open areas, the UAV terminal system continuously counts the number of valid points in each frame's slice point cloud. If the total number of valid slice layers is less than a preset threshold or the number of points in a critical altitude layer suddenly drops, the current frame is considered to be in a structurally degraded region. The UAV terminal system then implements a fault-tolerant strategy: temporarily switching to a short-term localization mode based on IMU and RTK, or performing automatic hovering, return-to-home, or soft landing operations after the continuous occlusion time exceeds a threshold. Furthermore, the UAV terminal system also features a multi-sensor time synchronization monitoring mechanism. This is achieved by maintaining a sliding window of timestamps for data streams from multiple sensors, including IMU, LiDAR, and cameras. If the maximum time difference is observed Exceeding the synchronous tolerance error If a time synchronization anomaly is detected, the mapping thread will be paused and the synchronization module will be restarted.

[0053] Based on the above anomaly types and severity levels, the UAV terminal system defines three operating states: normal mode, degraded mode (some functions are disabled, only the core navigation module is retained), and fault-tolerant mode (maintaining the current position and waiting for manual intervention).

[0054] To enable continuous environmental mapping and structural awareness of large port yards under dynamic operation, this embodiment designs a multi-frame map fusion mechanism based on slice point clouds, and adds structural change detection and map update functions to improve the accuracy, real-time performance and maintainability of the system during long-term deployment.

[0055] After localization is completed for each frame, the UAV terminal system transforms the current frame's point cloud into the world coordinate system. It then slices the same altitude layer from different time frames. By performing coordinate transformation and joint overlay, a multi-layered structured global port yard map can be formed. The UAV terminal system introduces voxel filtering (Voxel Grid) during the map construction process to control the point cloud density, and adds timestamp labels to all point clouds to facilitate subsequent perception and analysis of structural changes.

[0056] To achieve intelligent, dynamic, and risk-controllable management of port yard operations, this invention designs a yard operation control system oriented towards the business layer, covering key functions such as operation status modeling, abnormal status identification, dynamic path intervention, and risk early warning linkage, forming a closed-loop port intelligent operation and maintenance framework of "perception-understanding-decision-intervention".

[0057] Port Area Operation Status Map Design This system achieves graph-structure modeling of multi-source information such as the structural status of the port yard, equipment operation, and operational processes by deeply integrating UAV perception data and port operation scheduling data, thus constructing an operational situation map for operation management. This situation map expresses the spatial and task relationships between key elements such as containers, cranes, and stacking lanes in the port yard in a graph structure, and dynamically constructs a port operation map that links "operation status – spatial location – equipment interaction" by integrating perception information and planning data.

[0058] During each UAV flight mission, the UAV terminal system obtains the container location through point cloud slicing extraction and image recognition modules. Numbered labels Attitude vector Structural features Construct the state of the quintuple node: (3-1) in This represents the recognition confidence level. The edges between nodes are defined by both spatial adjacency and task flow relationships; for example, the edges along a crane path represent equipment. In time Move to node The task operation performed.

[0059] To represent the multi-layered stacking characteristics of the port area structure, the system organizes the operational status map into a multi-layered sub-map structure, with each layer corresponding to a height level of container stacking. This approach not only preserves spatial layering information but also facilitates intra-layer analysis of local changes and the propagation of intra-map constraints.

[0060] While constructing the structure diagram, the operation control system aligns the perception diagram with the pre-set plan diagram in the port area scheduling database. Specifically, for each perception node... The operation control system searches for the planned stacking location in the scheduling records. Expected number and task status labels Create a quintuple of planned states: (3-2) This is used to perform consistency checks on the nodes in the graph.

[0061] Routine detection and graded early warning methods During drone inspections of the port's storage yard, the operation control system needs to accurately detect anomalies in large-scale, highly dynamic work areas. To improve identification efficiency and reduce computational load, this system is designed with an anomaly candidate area identification triggering mechanism. This ensures that high-overhead consistency analysis is selectively performed only within suspected anomaly areas, thereby achieving resource optimization and real-time performance assurance. The overall risk warning process is as follows: Figure 7 As shown.

[0062] This mechanism is based on the continuous updating process of the operational status map. It utilizes the structural differences between the current frame image obtained during inspection and the historical stable reference image to quickly determine whether there are abnormal trends in local areas. The system uses node-level low-dimensional statistical features as criteria to construct a lightweight anomaly prior judgment model. Key indicators include: Point cloud density change rate It is used to identify whether an object is added, missing, or dumped. Infrared thermal imaging field changes This reflects a sudden change in the target's thermal state or abnormal heating of the equipment; Spatial center of gravity drift It is used to monitor structural movement or stack position deviation; Number identification confidence decay The message indicates that the serial number is obscured, worn, or that the camera is malfunctioning.

[0063] The above multi-source features are integrated into an anomaly candidate scoring function: (3-3) in This is a feature weighting factor, adaptively set based on node type and work area. When scoring... When the preset threshold is exceeded, the system will [remove] the node. Mark the node as an abnormal candidate and record the heap area to which it belongs as a subgraph to be investigated in detail.

[0064] Due to the complex issues present at the port area, such as dense stacking, severe obstruction, and incomplete numbering, the system employs a two-stage matching strategy for abnormal candidate nodes, taking into account the complementary advantages of numbering information and spatial structural features. The first stage is explicit numbering matching: for containers with clear numbering identification results and a confidence level higher than a threshold, a one-to-one binding is established between the number and the registered target in the database. The second stage is spatial location matching: when numbering information is missing or unreliable, the system calculates the Euclidean distance based on the centroid of the point cloud and the planned stacking location in the database, selecting the nearest neighbor as a candidate match. If the candidate is unique or has significant locational priority, a weak binding relationship is established and the matching confidence level is recorded.

[0065] After completing the initial matching of the perceived target and planned task data, the system initiates a consistency analysis module to perform in-depth verification of the matching results using a multi-dimensional fusion judgment approach. This module no longer relies on direct comparison of a single dimension, but instead constructs a high-resolution anomaly identification model that can distinguish between minor deviations and systematic anomalies through joint modeling of multiple factors such as spatial similarity, attribute confidence, and structural stability.

[0066] In determining spatial consistency, in addition to examining the Euclidean distance between the sensing center point and the planned stack location center point, directional projection and regional overlap rate are introduced as supplementary features. The projection deviation along the main axis of the stack location (e.g., the port area X-axis) is: (3-4) in This represents the unit vector along the pile location axis. The system compares this projection error with the physical dimensions of the pile location, and combines this with the volume percentage of the sliced ​​point cloud within the pile location voxel mesh to construct a spatial matching score. .like If the position is significantly offset, the system will mark it as misplaced.

[0067] Regarding numbering consistency, the system checks whether corresponding point numbers are consistent. If all candidate numbers do not meet the requirements, it is considered an anomaly in numbering conflicts. This mechanism is highly robust in dealing with issues such as numbering ambiguity, occlusion, and delayed database updates.

[0068] The system performs multi-scale evaluation of structural attitude consistency. First, at the local level (slice level), the angle between the principal direction vector of the point cloud and the standard attitude is evaluated. If the offset exceeds a set angle threshold, it is considered an attitude deviation. Furthermore, the system estimates the tilting trend index based on the normal distance between the centroid and the supporting plane fitted by the 3D point cloud boundary. When the following conditions are met: (3-5) This means that the target is considered to be at risk of overturning or sinking.

[0069] By integrating three indices—spatial structure similarity, numbering confidence consistency, and attitude stability—into a joint evaluation function: (3-6) in The system uses adaptive weighting coefficients, adjusting them based on target category, work area, and historical error experience. This is done according to the comprehensive anomaly score. The system is equipped with Level 1 (Attention), Level 2 (Warning), and Level 3 (Danger) warning signals.

[0070] Route Adjustment and Emergency Response Strategies To achieve rapid response after anomaly identification and dynamic avoidance of operational interference, the inspection system of this container yard introduces a real-time closed-loop control mechanism based on the early warning status map into the UAV path planning module. This mechanism decouples the risk identification results from the flight path and integrates them into the path cost map modeling process, constructing a trajectory adaptive adjustment framework that integrates "perception-evaluation-feedback-reconstruction". This effectively addresses various factors in the port operation environment, such as sudden risks, dynamic crane obstruction, and changes in scheduling plans.

[0071] The operation control system, based on the port area map and yard status, first constructs a three-dimensional path cost map. The cost of each voxel unit is determined by the base terrain impedance, spatial accessibility, and risk level. Risk areas identified by the perception module are mapped as high-cost areas, specifically determined by the anomaly risk level function. Convert to cost increment term: (3-7) in The penalty weight is used. The system uses the augmented cost graph as input for path search, through improved... The algorithm combines node state constraints to reconstruct the path and introduces spline smoothing interpolation when necessary to improve trajectory continuity and ensure the executability of the path in complex environments.

[0072] The system formulates three response strategies based on the different types and urgency levels of path interference: Minor interference (such as slight stack position offset or mismatched numbering): The system adopts a path fine-tuning method to locally avoid interference around the original path without interrupting the task.

[0073] Moderate interference (such as regional congestion, overlapping crane paths): The system automatically triggers the path replanning module, generates a new flight path in real time based on the cost map, and pushes the update to the UAV terminal.

[0074] High risk (such as tipping risk, low-altitude obstacles): Immediately suspend the current flight segment, perform hovering or return-to-base maneuvers until the risk is eliminated or administrators intervene.

[0075] To ensure the integrity of flight missions and the system's fault tolerance, the container yard inspection system has also added mission interruption recovery capabilities. When a UAV fails to complete its planned mission segment due to obstacle avoidance, flight pause, or other operations, the operation control system, by maintaining the mission execution stack and status flags, reallocates the incomplete segment to subsequent flight batches or other collaborating UAVs to ensure full coverage of inspection missions without omissions. Path closed-loop control is not only applicable to abnormal avoidance but also supports dynamic intervention from the business side, such as scheduling mission changes, area blockades, and temporary priority tasks. After injecting new operational constraints or strategy instructions into the path cost map through the control platform, the system will trigger trajectory replanning with a unified mechanism, achieving flexible collaboration between autonomous UAV inspection missions and port operation scheduling.

[0076] This path feedback mechanism effectively connects the response chain between the sensing end and the execution end, realizing the full-link control capability of local avoidance of dynamic risks, path correction and task recovery, and significantly improving the system's operational robustness and scheduling flexibility in complex and dynamic port environment.

[0077] Management and control platform and human-machine interface To achieve visualized supervision, strategic intervention, and cross-departmental collaboration in the port area inspection system, this system is equipped with a business-level intelligent port area management and control platform that integrates information display, task configuration, anomaly handling, and multi-drone scheduling. Through modular design and a multi-channel communication mechanism, the platform connects the front-end drone sensing system with the back-end port area business system, constructing a complete closed-loop management and control system encompassing data perception, intelligent judgment, strategic decision-making, and task delivery, thus becoming the interactive hub of the port area's intelligent inspection system.

[0078] The platform's core consists of four main functional modules: situational awareness visualization, risk event alerts and backtracking, task configuration and path monitoring, and manual intervention and scheduling collaboration. Firstly, the system uses 3D map rendering and graph structure reconstruction technology to dynamically present the operational situational awareness map in a real-time visual interface. Node information (number, stack location, status), path trajectory, and equipment location are expressed through different layers, and the system supports operations such as adjusting layer transparency, object filtering, and historical status queries, facilitating managers to quickly locate targets and analyze the site.

[0079] In terms of risk warning, the platform and the anomaly identification module achieve information linkage. All warning events are recorded in a timeline format and overlaid with heat map markers on the map, supporting warning classification filtering, event confirmation, and tracking playback functions. For key areas and critical equipment, users can set monitoring strategies and rules (such as tolerance thresholds, inspection frequency, and automatic review cycles). The system will automatically trigger task scheduling or administrator reminders according to the rules, building a semi-automatic to fully automatic adjustable warning response system.

[0080] In terms of task configuration, the platform provides a graphical task editor, allowing users to directly select inspection areas, set take-off and landing points, and plan time windows and frequencies based on a map. The system will automatically generate inspection tasks and assign them to the drone queue. Simultaneously, the task execution status is dynamically refreshed on the interface, supporting task progress tracking, flight path playback, and rescheduling of failed tasks. If there are abnormal conditions or task failures on-site, users can trigger a one-click full-process command flow of flight path modification – task takeover – path re-issuance, and the platform immediately synchronizes with the flight system, achieving efficient human-machine collaborative scheduling. To meet the collaborative needs of multiple systems in the port area, the platform opens a northbound interface, which can connect to container scheduling systems, video surveillance systems, and operation scheduling platforms to achieve data interoperability, task linkage, and information sharing. The platform also supports user permission management and log auditing mechanisms to ensure task transparency, clear responsibility, and data traceability in multi-department, multi-user scenarios.

[0081] As the management and decision-making end of the UAV intelligent inspection system, the platform effectively integrates front-end perception, mid-platform computing and back-end scheduling resources, empowering intelligent supervision and flexible response throughout the entire port operation process, and providing technical foundation and operational support for realizing a safe, efficient and collaborative modern port management model.

[0082] It should be noted that the methods in each subsystem of the container yard inspection system have already been explained, therefore the steps in the following container yard inspection methods will not be explained redundantly. Secondly, as... Figure 8 As shown, the present invention also provides a method for inspecting container yards, including: S801. Acquire the global optimized trajectory and sensor data. The sensor data includes: the lidar point cloud of the target UAV, the image data of the target UAV, and the IMU sensor data of the target UAV. The global map is the map corresponding to the global optimized trajectory. The global optimized trajectory is obtained by the target UAV based on the pose transformation matrix between adjacent frames of the target UAV obtained from the sensor data. The target UAV obtains the odometry trajectory based on the pose transformation matrix and inertial measurement data. The target UAV performs loop closure detection on the odometry trajectory and optimizes the global map. S802. Obtain the planning data for the target container yard. The planning data includes: planned location, expected number, and task status label. S803. Construct a port operation status map of the target container yard based on the map and sensor data corresponding to the globally optimized trajectory; S804. Perform anomaly detection and graded early warning on the operation status map and the corresponding plan map of the plan data to obtain the risk level; S805: Based on the risk level, path replanning is performed to obtain the optimal inspection path for the UAV.

[0083] In some embodiments of the present invention, obtaining the pose transformation matrix between adjacent frames of the target UAV based on sensor data includes: The relative height of each slice to the ground plane is determined based on the height of the container. Slice the container layer in the target yard at the current moment into multiple layers, taking each layer of containers as a layer; Perform layer-by-layer matching on the point cloud of each slice at the current time to obtain the pose transformation matrix between adjacent frames of the point cloud of each slice.

[0084] In some embodiments of the present invention, the odometry trajectory is obtained based on the pose transformation matrix and IMU sensor data, including: A kinematic model is built based on historical data from the IMU sensor. The IMU sensor data from the previous moment is input into the kinematic model to predict the UAV pose at the current moment. The target UAV's relative pose at the current moment is determined based on the pose transformation matrix and the UAV's pose at the current moment; the target UAV's odometry trajectory is determined based on the target UAV's relative pose at the current moment.

[0085] In some embodiments of the present invention, the step of performing loop closure detection and global graph optimization on the odometer trajectory to obtain a globally optimized trajectory includes: In the odometer trajectory, select historical keyframe point clouds that satisfy the preset spatial distance and preset time interval with the current odometer pose, and construct a set of loop closure candidate point clouds based on the historical keyframe point clouds; The point cloud at the current moment is matched with the historical point cloud in the loop closure candidate point cloud set to obtain the slice matching result, and the difference value between the current point cloud slice and the historical point cloud slice is calculated. Historical point clouds with differences below a preset threshold are identified as having loop closures, and the relative pose transformation between the current point cloud and the corresponding historical point cloud is calculated based on the slice matching results. The relative pose transformation is added as a lapsing constraint to the preset factor graph optimization model to optimize the odometer trajectory and obtain the globally optimized trajectory.

[0086] In some embodiments of the present invention, the expression for the globally optimized trajectory is:

[0087] In the formula, This indicates the odometer trajectory of the drone. Denotes the odometer edge set, This indicates the odometer error term. Indicates inter-frame odometer measurement. Indicate its covariance, It is the hysteresis error term. It is its covariance.

[0088] In some embodiments of the present invention, the expression of the preset GTSAM factor graph optimization model is:

[0089] In the formula, This represents the optimal state estimate obtained after factor graph optimization modeling. Odometer factor The corresponding residual function, Odometer factor The observed covariance matrix, Represents the cyclic factor The observed covariance matrix, This represents the target UAV pose node. For odometer factor, As a cyclic factor, This represents the frame of the point cloud at the current moment. This represents the frame in the candidate point cloud set that matches the historical point cloud at the current moment.

[0090] In some embodiments of the present invention, the step of performing anomaly detection and graded early warning on the plan map corresponding to the operation status map and the plan data to obtain the risk level includes: The spatial structure similarity analysis of the operational status map and the corresponding planning map based on the planning data is used to obtain similarity values; The consistency of the operational status map and the corresponding planning map based on the numbering confidence is determined to obtain the numbering confidence value. The stability of the operational status diagram and the corresponding planning diagram based on the planning data are assessed to obtain the stability value. The joint assessment function is determined based on the similarity value, the number confidence value, and the stability value. The risk level is then determined based on the value of the joint assessment function.

[0091] In summary, the embodiments provided by this invention achieve centimeter-level real-time positioning accuracy by combining multi-source data from LiDAR and IMU and using a layered slicing mapping method in port yard environments with severe obstruction and dense structures, significantly reducing vertical drift error.

[0092] In terms of mapping performance, this invention effectively suppresses error accumulation and maintains global trajectory closure consistency through loop closure detection and factor graph optimization methods, ensuring mapping accuracy over a wide range and long period of time, and supporting long-term unattended operation.

[0093] In terms of path planning and multi-drone collaboration, this invention constructs a three-dimensional path cost map based on dynamic changes in the yard and risk classification information. The system can automatically avoid obstacles such as cranes and equipment, and dynamically adjust the flight path to ensure continuous mission execution. In multi-drone collaborative scenarios, the system supports path recovery and secondary scheduling mechanisms after mission failure. It can automatically transfer tasks in the event of flight control anomalies, insufficient power, or link loss, significantly improving mission completion rate and system robustness.

[0094] In terms of risk response, this invention significantly reduces system resource consumption and improves the focus and path linkage efficiency of key area identification through a lightweight anomaly candidate area screening mechanism. Integrating multi-source sensing methods such as point cloud, image, and thermal imaging, it can automatically identify problems such as misplacement of stacked items, equipment malfunctions, numbering conflicts, and overturning risks, issue early warnings according to risk levels, and take corresponding measures to improve the overall safety and operational efficiency of the system.

[0095] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0096] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A container yard inspection system, characterized in that, include: The UAV terminal subsystem includes several UAVs, which are used to acquire and obtain the pose transformation matrix between adjacent frames of the target UAV based on the sensor data of the target UAV, obtain the odometry trajectory based on the pose transformation matrix and inertial measurement data, perform loop closure detection on the odometry trajectory and perform global graph optimization to obtain the globally optimized trajectory. The sensor data of the target UAV includes: lidar point cloud, image data and IMU sensor data. The operation control system communicates with drones to acquire planned data of the target container yard. Based on the global map and sensor data corresponding to the globally optimized trajectory, it constructs a port operation status map of the target container yard. The operation status map is compared with the planned map corresponding to the planned data, and anomaly detection and graded early warning are performed to obtain the risk level. Based on the risk level, the path is replanned to obtain the optimal inspection path of the drone. The planned data includes: planned stacking location, expected number, and task status label.

2. The container yard inspection system according to claim 1, characterized in that, The task status labels include: When the task at the planned stacking location is a loading task, the task status labels include: not started, stacked and waiting to be loaded, loading in progress, loaded and departed. When the planned task at the stacking location is a container pickup task, the task status labels include: waiting to arrive, waiting to be picked up, in the process of picking up, and already departed. The plan diagram includes: planned stack location, expected number, and task status label.

3. The container yard inspection system according to claim 1, characterized in that, The UAV terminal subsystem is also used for: Select historical keyframe point clouds from the odometer trajectory that satisfy the preset spatial distance and preset time interval with the current odometer pose, and construct a set of loop closure candidate point clouds based on the historical keyframe point clouds; Calculate the difference between the current point cloud in the same slice and the historical point cloud in the set of loop closure candidate point clouds; The historical point cloud with a difference value below a preset threshold is used to form a point cloud pair with the current point cloud, and the relative pose transformation between the point cloud pairs is calculated. The relative pose transformation is added as a lapsing constraint to the preset factor graph optimization model to optimize the odometer trajectory and obtain the globally optimized trajectory.

4. A method for inspecting a container yard, characterized in that, include: The global optimized trajectory and sensor data are acquired. The sensor data includes: the lidar point cloud of the target UAV, the image data of the target UAV, and the IMU sensor data of the target UAV. The global optimized trajectory is obtained by the target UAV based on the sensor data to obtain the pose transformation matrix between adjacent frames of the target UAV, the target UAV based on the pose transformation matrix and the inertial measurement data to obtain the odometry trajectory, and the target UAV performs loop closure detection and global graph optimization on the odometry trajectory. Obtain the planning data for the target container yard, which includes: planned stack location, expected number, and task status label; A port operation status map of the target container yard is constructed based on map and sensor data corresponding to the globally optimized trajectory; By performing anomaly detection and graded early warning on the operational status map and the corresponding planning map based on the planning data, the risk level can be obtained; Based on the risk level, path replanning is performed to obtain the optimal inspection path for the UAV.

5. The container yard inspection method according to claim 4, characterized in that, The process of obtaining the pose transformation matrix between adjacent frames of the target UAV based on sensor data includes: The relative height of each slice to the ground plane is determined based on the height of the container. Slice the container layer in the target yard at the current moment into multiple layers, taking each layer of containers as a layer; Perform layer-by-layer matching on the point cloud of each slice at the current time to obtain the pose transformation matrix between adjacent frames of the point cloud of each slice.

6. The container yard inspection method according to claim 4, characterized in that, The odometry trajectory is obtained based on the pose transformation matrix and IMU sensor data, including: A kinematic model is built based on historical data from the IMU sensor. The IMU sensor data from the previous moment is input into the kinematic model to predict the UAV pose at the current moment. Obtain and determine the relative pose of the target UAV at the current moment based on the pose transformation matrix and the UAV pose at the current moment; The odometry trajectory of the target drone is determined based on its current relative pose.

7. The container yard inspection method according to claim 4, characterized in that, The process of performing loop closure detection and global graph optimization on the odometer trajectory to obtain a globally optimized trajectory includes: Select historical keyframe point clouds from the odometer trajectory that satisfy the preset spatial distance and preset time interval with the current odometer pose, and construct a set of loop closure candidate point clouds based on the historical keyframe point clouds; Calculate the difference between the current point cloud in the same slice and the historical point cloud in the set of loop closure candidate point clouds; The historical point cloud with a difference value below a preset threshold is used to form a point cloud pair with the current point cloud, and the relative pose transformation between the point cloud pairs is calculated. The relative pose transformation is added as a lapsing constraint to the preset factor graph optimization model to optimize the odometer trajectory and obtain the globally optimized trajectory.

8. The container yard inspection method according to claim 7, characterized in that, The expression for the globally optimized trajectory is: In the formula, This indicates the globally optimized trajectory of the drone. Denotes the odometer edge set, This indicates the odometer error term. Indicates inter-frame odometer measurement. Indicate its covariance, It is the hysteresis error term. It is its covariance. This represents the relative pose transformation between point cloud pairs. Indicates the first pose variables of each pose node Indicates the first pose variables of each pose node Indicates the first pose variables of each pose node This indicates the odometer trajectory.

9. The container yard inspection method according to claim 7, characterized in that, The expression for the preset GTSAM factor graph optimization model is: In the formula, This represents the optimal state estimate obtained after optimization of the factor graph model. Odometer factor The corresponding residual function, Odometer factor The observed covariance matrix, Represents the cyclic factor The observed covariance matrix, This represents the target UAV pose node. For odometer factor, As the cyclic factor, This represents the frame of the point cloud at the current moment. This represents the frame in the candidate point cloud set that matches the historical point cloud at the current moment.

10. The container yard inspection method according to claim 4, characterized in that, The process of performing anomaly detection and graded early warning on the operational status map and the corresponding planning map based on the planning data to obtain the risk level includes: The spatial structure similarity analysis of the operational status map and the corresponding planning map based on the planning data is used to obtain similarity values; The consistency of the operational status map and the corresponding planning map based on the numbering confidence is determined to obtain the numbering confidence value. The stability of the operational status diagram and the corresponding planning diagram based on the planning data are assessed to obtain the stability value. The joint assessment function is determined based on the similarity value, the number confidence value, and the stability value. The risk level is then determined based on the value of the joint assessment function.

Images