Container lock intelligent identification dismounting method and device based on 3D vision

Abstract

The application relates to the technical field of industrial robot vision positioning and automatic control, in particular to a container lock intelligent identification dismounting method and device based on 3D vision, which comprises the following steps: collecting a structured light image of a target lock under multiple polarization angles; calculating pixel polarization characteristic parameters according to the structured light images under the multiple polarization angles, separating diffuse reflection light pixels and generating a three-dimensional point cloud of the target lock; performing global coarse positioning, local fine positioning and real-time dynamic tracking on the three-dimensional point cloud to obtain a real-time six-dimensional pose of the target lock; controlling a manipulator to perform lock dismounting operation according to the real-time six-dimensional pose; and adopting a standard calibration ball to perform online hand-eye calibration residual error compensation. The application mainly solves the technical problems of poor environmental adaptability, insufficient positioning accuracy, incapability of compensating dynamic disturbance and easy accumulation of calibration errors of an existing container lock dismounting system.

Description

Technical Field

[0001] This application relates to the field of industrial robot visual positioning and automation control technology, specifically to a method and device for intelligent identification and disassembly of container locks based on 3D vision. Background Technology

[0002] Container shipping is a core component of the modern logistics system, supporting over 90% of global international trade cargo transportation. The installation and removal of container locks is a critical process in port terminal operations, characterized by high frequency, high labor intensity, and significant safety risks. With the accelerated development of automated terminals and smart ports, 3D vision-based automated loading and unloading equipment is gradually replacing manual labor. However, under the complex conditions of actual ports, traditional vision systems still face many insurmountable technical bottlenecks.

[0003] The lighting conditions at the port vary drastically. At midday, direct sunlight can reach 100,000 lux, while nighttime operations are characterized by low light levels, along with complex lighting conditions such as backlighting, alternating shadows, and flickering lights. Adding to the challenge, the container locks are made of metal with galvanized or painted surfaces, making them highly susceptible to specular reflection. This results in large overexposed areas and numerous invalid noise points in images captured by ordinary structured light cameras, sometimes even preventing proper imaging. This directly leads to a lock positioning failure rate exceeding 40%, becoming the primary obstacle to the large-scale application of automated assembly and disassembly systems.

[0004] Meanwhile, most existing systems rely on static global 3D modeling for positioning, achieving an overall positioning accuracy of only about ±5mm. However, the diameter of container lock cylinder holes is typically only 12mm to 16mm, requiring a positioning error of less than 1mm to ensure the robotic arm can smoothly insert into the lock cylinder and complete the assembly / disassembly process. This current level of accuracy easily leads to problems such as lock insertion failure, robotic arm collisions, and lock deformation and damage, severely impacting operational efficiency and equipment lifespan.

[0005] Under dynamic operating conditions, the container itself may experience a positional shift of ±10cm and an angular tilt of ±5° due to operations such as forklift handling and platform clamping. Furthermore, during assembly and disassembly, the contact force between the robotic arm and the locks can cause slight shaking within ±5mm of the container. Traditional static positioning methods can only obtain the lock's position at a certain moment and cannot compensate for these dynamic changes in real time. This can easily lead to the continuous accumulation of positioning deviations, resulting in safety hazards such as improper insertion, jamming, or even equipment collisions.

[0006] Furthermore, traditional hand-eye calibration often employs an offline nine-point calibration method. During long-term system operation, the calibration results gradually drift due to factors such as changes in ambient temperature, equipment vibration, and mechanical wear. Actual operational data shows that after one month of operation, the hand-eye calibration error of a traditional system can rise to over ±3mm, requiring frequent shutdowns for recalibration by skilled technicians. This not only increases equipment maintenance costs but also disrupts dock operations, impacting overall production efficiency.

[0007] Therefore, how to achieve stable imaging, improve positioning accuracy, compensate for dynamic disturbances in real time, and maintain long-term reliable operation of the system in complex port environments has become a core technical problem that urgently needs to be solved in the field of automated container lock assembly and disassembly. Summary of the Invention

[0008] This application mainly addresses the technical problems of existing automated container lock assembly and disassembly equipment, such as susceptibility to imaging interference in harsh environments, insufficient positioning accuracy, difficulty in adapting to dynamic offsets, and accumulation of calibration errors over long-term operation.

[0009] In a first aspect, embodiments of this application provide a method for intelligent identification and disassembly of container locks based on 3D vision, the method comprising: Acquire structured light images of the target lock at multiple polarization angles; Based on the structured light images with multiple polarization angles, the polarization feature parameters of each pixel are calculated, the diffuse reflection light pixels are separated, and a three-dimensional point cloud of the target lock is generated. Multi-level dynamic positioning is performed on the three-dimensional point cloud to obtain the real-time six-dimensional pose of the target lock; Based on the real-time six-dimensional pose, the robotic arm is controlled to perform the disassembly and assembly operations of the lock.

[0010] Secondly, the embodiments of this application also provide a container lock intelligent identification and disassembly device based on 3D vision, including a polarization-enhanced three-dimensional image acquisition module, a processor module, a multi-manipulator execution module, and an online calibration and residual compensation module; The polarization-enhanced 3D image acquisition module is used to acquire structured light images of the target lock at multiple polarization angles; The processor module is communicatively connected to the polarization-enhanced 3D image acquisition module to execute the intelligent identification and disassembly method, obtain the real-time six-dimensional pose of the target lock, and generate robotic arm control commands. The multi-manipulator execution module is communicatively connected to the processor module and is used to perform the disassembly and assembly operations of the lock according to control commands; The online calibration and residual compensation module is used to correct hand-eye calibration errors in real time, ensuring the long-term operating accuracy of the system.

[0011] This application provides a 3D vision-based intelligent identification and disassembly method for container locks. Through an innovative imaging approach combining multi-angle polarization imaging and structured light 3D reconstruction, it effectively suppresses metal reflections and strong light interference from the imaging source by utilizing the difference in polarization characteristics between specular and diffuse light, enabling stable operation across the entire illumination range. A three-tiered progressive positioning strategy—global coarse positioning, local fine positioning, and real-time dynamic tracking—ensures millisecond-level response speed while improving static positioning accuracy and dynamic tracking accuracy. This reduces the disassembly and assembly time for a single container lock from 15-20 minutes manually to 3-5 minutes, achieving a lock identification and positioning success rate of 99.2%, fully meeting the high-speed, high-precision, and high-reliability requirements of modern automated terminals for container lock disassembly and assembly operations. Attached Figure Description

[0012] Figure 1 A flowchart of a 3D vision-based intelligent identification and disassembly method for container locks provided in this application embodiment; Figure 2 A flowchart of the three-level dynamic positioning method provided in the embodiments of this application; Figure 3 A flowchart of the online hand-eye calibration residual compensation method provided in the embodiments of this application; Figure 4 This is a structural block diagram of the container lock intelligent identification and disassembly device provided in the embodiments of this application.

[0013] In the figure, 401 is the polarization-enhanced 3D image acquisition module; 402 is the processor module; 403 is the multi-manipulator execution module; and 404 is the online calibration and residual compensation module. Detailed Implementation

[0014] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments. Similar elements in different embodiments are referred to by related similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the present application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to the present application are not shown or described in the specification. This is to avoid obscuring the core parts of the present application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0015] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0016] The serial numbers assigned to components in this document, such as "first" and "second," are used only to distinguish the described objects and have no sequential or technical meaning. The terms "connection" and "linkage" used in this application, unless otherwise specified, include both direct and indirect connections (linkages).

[0017] Please refer to Figure 1 To address the problems of unstable imaging, insufficient positioning accuracy, and difficulty in adapting to dynamic changes in the complex environment of ports, existing container lock assembly and disassembly equipment provides a 3D vision-based intelligent identification and disassembly method for container locks. This method is applicable to the automated disassembly and disassembly of various types of container locks, such as twist locks, rotary locks, and fixed locks, in automated terminals. The method includes the following steps: Step S101: Acquire structured light images of the target lock at multiple polarization angles. In this step, after the container is transported to the disassembly / assembly platform by a forklift and positioned and fixed by a hydraulic clamping device, the processor sends a pulse control signal to the stepper motor driver. This drives the high-precision stepper motor to rotate the linear polarizer in front of the camera lens sequentially to four standard polarization angles: 0°, 45°, 90°, and 135°. At each polarization angle, the 3D structured light camera projects a set of sinusoidal fringe structured light patterns onto the lock area through a DLP projection unit, and simultaneously acquires the corresponding deformed fringe images. After the images at the four different polarization angles are acquired sequentially, they are transmitted to the processor's memory via gigabit Ethernet for subsequent polarization calculation processing.

[0018] In this embodiment, the 3D structured light camera has a resolution of 1920×1200, a frame rate of 30fps, and a measurement distance range of 1 to 3m, which can cover the installation positions of all locks on the side of the container; the linear polarizer has an extinction ratio of not less than 1000:1 and an effective wavelength range of 400 to 700nm, which can effectively distinguish between specular reflection light from the metal surface and diffuse reflection light from the lock body; the stepper motor has a step angle of 0.9°, which, together with the microstepping driver, can achieve a rotational accuracy of 0.01°, ensuring that the polarization angle error is less than 1°.

[0019] In some embodiments, the system is equipped with an adaptive auxiliary light source consisting of 12 high-brightness LEDs. The system can detect the ambient light intensity in real time through a photosensitive sensor and automatically adjust the output power and illumination angle of the auxiliary light source. This improves image contrast in low-light environments and reduces output power in strong light environments to avoid further aggravating specular reflection on metal surfaces, ensuring that the imaging process remains stable and reliable under various lighting conditions.

[0020] Step S102: Based on the structured light images at multiple polarization angles, calculate the polarization feature parameters of each pixel, separate the diffuse reflection pixels, and generate a 3D point cloud of the target lock. In this step, the processor performs pixel-by-pixel parallel calculations on the grayscale values ​​of the images acquired at four angles: 0°, 45°, 90°, and 135°, obtaining three components of the Stokes vector: S0 is the total light intensity, equal to the sum of the grayscale values ​​at polarization angles of 0° and 90°; S1 is the difference between the horizontal and vertical polarization components, equal to the difference between the grayscale values ​​at polarization angles of 0° and 90°; and S2 is the difference between the polarization components at polarization angles of 45° and 135°. Based on the Stokes vector, the degree of polarization corresponding to each pixel is further calculated. The degree of polarization reflects the degree of polarization of light; specular reflection light typically has a higher degree of polarization, while diffuse reflection light has a lower degree of polarization.

[0021] Based on an optimal polarization threshold of 0.7 determined through 1000 sets of experiments, pixels with a polarization degree greater than 0.7 are identified as noise points from metallic mirror reflections and directly discarded, retaining only effective diffuse reflection pixels with a polarization degree less than or equal to 0.7. Phase unpacking and 3D reconstruction are then performed using the deformed fringe images corresponding to these effective pixels, ultimately generating a clean, complete, and extremely low-noise 3D point cloud of the target lock. This processing method significantly reduces point cloud noise caused by metallic reflections and strong light at the source, reducing noise points in reflective areas by 92.7%, fundamentally solving the imaging failure problem caused by strong light and metallic reflections in ports.

[0022] In some embodiments, the system adaptively fine-tunes the polarization threshold based on real-time illumination intensity. In low-light environments such as cloudy days, dusk, and nighttime, the threshold is appropriately lowered to 0.6 to ensure a sufficient number of effective point clouds. In strong midday light, the threshold is appropriately increased to 0.75 to further enhance noise suppression. Simultaneously, the generated 3D point cloud is denoised using a combination of statistical filtering and radius filtering, and downsampling is performed using a voxel mesh method. This reduces the point cloud data volume by more than 70% without losing key geometric features, significantly improving the computational speed for subsequent localization and registration.

[0023] Step S103: Perform multi-level dynamic positioning on the three-dimensional point cloud to obtain the real-time six-dimensional pose of the target lock. During this step, the system calls... Figure 2The corresponding three-level dynamic positioning method performs progressive positioning processing on the preprocessed 3D point cloud, and finally outputs the real-time six-dimensional pose of the lock.

[0024] Step S104: Based on the real-time six-dimensional pose, control the robotic arm to perform the disassembly and assembly operations of the lock. In this step, the processor uses a hand-eye conversion matrix to transform the final output real-time six-dimensional pose of the lock from the camera coordinate system to the robotic arm's base coordinate system. Combining the lock model and the preset disassembly and assembly sequence, a fifth-order polynomial interpolation algorithm is used to generate a smooth, collision-free motion trajectory and execution commands. Multiple six-axis industrial robots simultaneously receive commands and collaboratively complete the disassembly or assembly of the lock. Throughout the entire disassembly and assembly process, the real-time dynamic positioning module continuously outputs the corrected pose results, and the robotic arm controller fine-tunes the motion trajectory in real time to ensure smooth and accurate operation, avoiding insertion failures or equipment collisions due to accumulated deviations.

[0025] After the operation is completed, the system automatically collects the final state image of the lock, verifies the proper installation and removal of the lock using image recognition algorithms, and synchronously uploads the collected images, point cloud data, execution records, verification results and other information to the cloud server for storage, realizing full traceability of the entire operation process.

[0026] Step S105: Perform online hand-eye calibration residual compensation using a standard calibration ball. During this step, the system calls... Figure 3 The corresponding online hand-eye calibration residual compensation method automatically executes the calibration compensation process after the system is started or has been running for a certain period of time, continuously correcting hand-eye conversion errors and ensuring the long-term operating accuracy of the system.

[0027] Please refer to Figure 2 The three-level dynamic positioning method provided in this application embodiment is specifically designed for rapid and high-precision acquisition and dynamic tracking of lock pose. The method includes the following steps: Step S201: Input the polarization-optimized 2D image and 3D point cloud. In this step, the low-noise 3D point cloud generated in step S102 and the corresponding polarization-optimized 2D image are used as input to provide basic data for subsequent positioning at each level.

[0028] Step S202: Perform global coarse localization to determine the region of interest (ROI) for each lock. In this step, a lightweight YOLOv8n object detection algorithm is used for rapid inference on the 2D image, outputting the bounding boxes, center coordinates, and category information for all locks. The global coarse localization has an accuracy of approximately ±2cm and takes less than 50ms, quickly narrowing down the processing area for subsequent steps, defining the ROI for local fine localization, and significantly improving the overall system response speed.

[0029] Step S203: Perform local fine localization to obtain the initial six-dimensional pose of the lock. In this step, based on the bounding box coordinates obtained from coarse localization, a local point cloud corresponding to a single lock is precisely cropped from the complete 3D point cloud. After denoising and downsampling preprocessing of the local point cloud, an improved iterative nearest-point algorithm is used to accurately register it with the pre-stored standard lock point cloud in the database. A polarization degree weighting mechanism is introduced during the registration process: diffuse reflection effective points with low polarization degree are given higher matching weights, while residual high polarization degree noise points are given lower weights, thereby further suppressing the impact of noise on registration accuracy. The accuracy of local fine localization can reach ±0.5mm, with a time consumption of less than 100ms, outputting the initial six-dimensional pose of the lock.

[0030] Step S204: Perform real-time dynamic tracking and output the real-time six-dimensional pose of the lock. During this step, as the robotic arm moves towards the lock, the system continuously acquires local images of the lock at a high frequency of 30Hz, extracts Shi-Tomasi corner points as key feature points, and uses the Lucas-Kanade optical flow method to continuously track the displacement changes of these feature points. It then calculates in real-time the minute offsets and attitude corrections caused by the contact force of the robotic arm, dynamically correcting the initial six-dimensional pose. The accuracy of real-time dynamic tracking can reach ±0.3mm, effectively compensating for the shaking errors generated during container assembly and disassembly, and finally outputting the real-time six-dimensional pose of the lock.

[0031] The total time for Level 3 positioning is less than 200ms, and the positioning success rate reached 99.2% in 1000 tests under different environmental conditions. In some embodiments, the dynamic tracking module has an automatic re-initialization mechanism for feature loss. When feature points are occluded or tracking fails during operation, the system can automatically switch back to the local fine positioning mode to reacquire the precise pose of the lock, ensuring uninterrupted continuous operation. At the same time, the system automatically switches the tracking update frequency according to the real-time distance between the robot and the lock. At long distances, the frequency is reduced to 10Hz to save computing power, and at short distances, the frequency is increased to 30Hz to ensure tracking accuracy, achieving an optimal balance between computing resources and system performance.

[0032] Please refer to Figure 3 The online hand-eye calibration residual compensation method provided in this application embodiment is specifically designed to solve the problem of error accumulation over time in traditional offline calibration, ensuring the long-term operational accuracy of the system. The method includes the following steps: Step S301: System initialization, control the robot arm to move to the calibration position. During this step, after the system starts and completes hardware initialization and software self-test of each module, the probe at the end of the robot arm is moved to a safe position above the standard ceramic calibration ball fixed in the upper left corner of the disassembly / assembly platform, ready to perform the calibration operation.

[0033] Step S302: Collect multi-position contact data of the calibration ball. In this step, control the robotic arm to slowly descend at a speed of 0.5 mm / s until the probe contacts the top of the calibration ball, and record the joint angle of the robotic arm at this point. Repeat the above operation, sequentially contacting five different positions on the calibration ball: bottom, front, rear, and left side, collecting a total of five sets of joint angle data to ensure accurate fitting of the calibration ball's center coordinates.

[0034] Step S303: Calculate the center coordinates of the calibration ball in both coordinate systems. In this step, based on the five sets of joint angle data collected, the center coordinates P_robot of the calibration ball in the manipulator's base coordinate system are calculated using the robot's forward kinematics. Simultaneously, images of the calibration ball are acquired using a 3D camera, and the center coordinates P_camera of the calibration ball in the camera coordinate system are calculated using a sphere fitting algorithm.

[0035] Step S304: Calculate the hand-eye conversion matrix residual. In this step, the current actual hand-eye conversion matrix T_current is calculated based on two sets of coordinates and compared with the hand-eye matrix T0 obtained from the initial offline calibration of the system. The residual matrix ΔT = T_current × T0⁻¹ is then obtained. This residual matrix reflects the calibration drift generated after long-term system operation.

[0036] Step S305: Apply residual compensation and update the error compensation model. In this step, the calculated residual matrix ΔT is updated into the system's error compensation model. In all subsequent positioning calculations, the point cloud in the camera coordinate system is first multiplied by the residual matrix ΔT for compensation, and then transformed to the robot's coordinate system, thereby effectively suppressing calibration drift caused by factors such as temperature, vibration, and wear.

[0037] This online residual compensation method reduces the manual calibration error from the traditional ±2mm to ±0.3mm. After a month of continuous operation, the positioning accuracy remains within ±0.5mm, significantly reducing the frequency of manual calibration and lowering equipment maintenance costs. In some embodiments, the system monitors the change in the residual matrix in real time. When the residual exceeds the preset allowable range, it automatically triggers an audible and visual fault alarm, prompting operators to check the camera's fixation status, the robot arm's installation accuracy, and whether the calibration ball has shifted, thus preventing system accuracy degradation and ensuring operational safety.

[0038] Please refer to Figure 4To reliably implement the aforementioned intelligent identification and disassembly functions for container locks, this application provides a 3D vision-based intelligent identification and disassembly device for container locks. The device includes a polarization-enhanced 3D image acquisition module 401, a processor module 402, a multi-manipulator execution module 403, and an online calibration and residual compensation module 404. All modules are connected via high-speed industrial Ethernet communication, with a data transmission delay of less than 1ms, meeting the system's real-time requirements.

[0039] The polarization-enhanced 3D image acquisition module 401 is the perception core of the system, mainly composed of a 3D structured light camera, a rotatable linear polarizer, a high-precision stepper motor, a motor driver, and an adaptive auxiliary light source. Four sets of this module are configured, installed at the four corners of the assembly / disassembly platform, at a height of 2.5m and an angle of 45°, fully covering all locking areas on all four sides of the container. This module effectively suppresses metallic reflections and strong light interference through multi-angle polarization imaging, providing high-quality images and point cloud data for subsequent processing.

[0040] The processor module 402 is the core of the system's computing and control. It adopts an industrial-grade high-performance industrial control computer, equipped with an Intel Core i7-12700K CPU, 32GB of memory, and an NVIDIA RTX 3060 GPU, which can meet the high-intensity computing requirements of real-time image processing, 3D reconstruction, multi-level positioning, and trajectory planning. The module has a built-in standard lock model library, disassembly and assembly strategy library, and error compensation model, which can automatically call the corresponding processing algorithms and operation processes according to different lock models.

[0041] The multi-manipulator execution module 403 consists of four six-axis industrial robots, each mounted on one of the four sides of the assembly / disassembly platform. Each robot has a payload of 10kg, a repeatability of ±0.1mm, and features rapid response and high-precision motion capabilities. The robot end effector is equipped with dedicated lock assembly / disassembly tools to accommodate different lock models.

[0042] The online calibration and residual compensation module 404 uses a standard ceramic calibration ball fixed on the assembly and disassembly platform as a reference. Combined with the measurement data from the robot and camera, it calculates and compensates for hand-eye calibration errors in real time, ensuring the stability of the system's positioning accuracy during long-term operation.

[0043] The system described in this application refers to the collaborative operation of the following multiple modules, which include at least: a polarization-enhanced 3D image acquisition module 401, which acquires images of the calibration sphere and calculates the coordinates of the sphere's center in the camera coordinate system; a processor module 402, which acts as the control core, initiating the calibration process, calculating the hand-eye transformation matrix residual, and updating the error compensation model; a multi-manipulator execution module 403, which executes the action of the probe contacting the calibration sphere and acquires joint angle data; and an online calibration and residual compensation module 404, which stores calibration parameters and performs residual compensation calculations.

[0044] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0045] The above examples illustrate this application only to aid understanding and are not intended to limit its scope. Those skilled in the art to which this application pertains can make various simple deductions, modifications, or substitutions based on the ideas presented.

Claims

1. A method for intelligent identification and disassembly of container locks based on 3D vision, characterized in that, The intelligent identification and disassembly method includes: Acquire structured light images of the target lock at multiple polarization angles; Based on the structured light images with multiple polarization angles, the polarization feature parameters of each pixel are calculated, the diffuse reflection light pixels are separated, and a three-dimensional point cloud of the target lock is generated. Multi-level dynamic positioning is performed on the three-dimensional point cloud to obtain the real-time six-dimensional pose of the target lock; Based on the real-time six-dimensional pose, the robotic arm is controlled to perform the disassembly and assembly operations of the lock.

2. The intelligent identification and disassembly method for container locks based on 3D vision as described in claim 1, characterized in that, The acquisition of structured light images of the target lock at multiple polarization angles includes: The linear polarizer is controlled to rotate to different angles in sequence, and a sinusoidal fringe pattern is projected at each angle and the corresponding deformed fringe image is acquired.

3. The intelligent identification and disassembly method for container locks based on 3D vision as described in claim 2, characterized in that, The calculation of polarization feature parameters for each pixel to separate diffuse reflection pixels includes: Calculate the Stokes vector components of each pixel based on the image grayscale values ​​at different polarization angles; The degree of polarization of each pixel is calculated based on the Stokes vector components; Pixels with a polarization degree greater than a preset threshold are identified as specular reflection noise points and removed, while diffuse reflection light pixels with a polarization degree less than or equal to the preset threshold are retained.

4. The intelligent identification and disassembly method for container locks based on 3D vision as described in claim 1, characterized in that, The process of performing multi-level dynamic positioning on the three-dimensional point cloud to obtain the real-time six-dimensional pose of the target lock includes: Perform global coarse localization of the target lock to obtain its approximate location and category information; Based on the global coarse localization results, the target lock is local fine localized to obtain the initial six-dimensional pose of the lock. During the process of the robotic arm approaching the lock, the target lock is dynamically tracked in real time, and the initial six-dimensional pose is corrected to obtain the real-time six-dimensional pose.

5. The intelligent identification and disassembly method for container locks based on 3D vision as described in claim 4, characterized in that, The global coarse positioning of the target lock includes: The target detection algorithm is used to detect the bounding box of the target lock from the two-dimensional image after polarization imaging, and the center coordinates and category information of the lock are output.

6. The intelligent identification and disassembly method for container locks based on 3D vision as described in claim 4, characterized in that, The local fine positioning of the target lock includes: The local 3D point cloud of the lock is cropped from the bounding box obtained by global coarse localization. The local point cloud is then registered with the standard lock point cloud using an iterative nearest point algorithm. During the registration process, different matching weights are assigned according to the polarization degree of each point.

7. The intelligent identification and disassembly method for container locks based on 3D vision as described in claim 4, characterized in that, The real-time dynamic tracking of the target lock includes: Key feature points on the surface of the lock are extracted, and the displacement changes of the feature points are tracked using optical flow method to calculate the attitude correction amount of the lock in real time.

8. The intelligent identification and disassembly method for container locks based on 3D vision as described in claim 1, characterized in that, The intelligent identification and disassembly method also includes: A standard calibration ball fixed on the assembly / disassembly platform is used for online hand-eye calibration residual compensation to correct the conversion error between the camera coordinate system and the robot coordinate system in real time.

9. A container lock intelligent identification and disassembly device based on 3D vision, characterized in that, The intelligent identification and disassembly device includes a polarization-enhanced 3D image acquisition module, a processor module, and a multi-manipulator execution module. The polarization-enhanced 3D image acquisition module is used to acquire structured light images of the target lock at multiple polarization angles; The processor module is communicatively connected to the polarization-enhanced three-dimensional image acquisition module and is used to execute the method described in any one of claims 1-8 to obtain the real-time six-dimensional pose of the target lock and generate robotic arm control commands. The multi-manipulator execution module is communicatively connected to the processor module and is used to perform the disassembly and assembly operations of the lock according to control commands.

10. The intelligent identification and disassembly device for container locks as described in claim 9, characterized in that, The polarization-enhanced 3D image acquisition module includes a 3D structured light camera, a rotatable linear polarizer, and a stepper motor. The extinction ratio of the linear polarizer is sufficient to separate specular reflection light from diffuse reflection light. The stepper motor can drive the polarizer to rotate to at least four different polarization angles. The preset polarization degree threshold can effectively eliminate specular reflection noise points. The global coarse positioning can quickly determine the approximate position of the lock, the local fine positioning can meet the accuracy requirements of lock assembly and disassembly, and the real-time dynamic tracking can compensate for the slight shaking of the lock. The standard calibration ball is a high-precision spherical reference component, which is fixedly installed on the assembly and disassembly platform.

Images