A lock buckle concurrent recognition method and system based on multi-level point cloud matching

Abstract

The application discloses a kind of based on multi-level point cloud matching's lock buckle concurrent identification method and system, including global request foolproof interception unit, configuration distribution and scheduling unit, point cloud dimension reduction and feature preprocessing unit, multi-level core matching unit and dynamic safety anti-collision evaluation unit, multi-level core matching unit is connected with point cloud dimension reduction and feature preprocessing unit.The beneficial effects of the present application are that high applicability, high real-time lock buckle concurrent identification is realized on industrial edge computing platform.Under normal operating conditions of the system, through the foolproof interception mechanism in front and the process scheduling of isolation, the impact of illegal instructions on algorithm engine is effectively blocked, and the system stability in multi-task environment is guaranteed;At the same time, through the design of multi-level point cloud matching, the waste of computing power caused by full iteration is avoided, the time consumption of single identification is controlled at a very low level, the recognition rate of the system under unknown lock buckle pose condition is improved, and the identification time is reduced.

Description

Technical Field

[0001] This invention relates to the fields of industrial machine vision, 3D point cloud processing and automated collaborative control technology, and more specifically, to a method and system for concurrent recognition of latches based on multi-level point cloud matching. Background Technology

[0002] With the development of industrial automation, precise grasping using 3D vision-guided robotic arms has become a crucial aspect of intelligent manufacturing. In the automated sorting and assembly of complex metal structural components such as latches, the system needs to quickly and accurately identify the model of the target latch from noisy and occluded local point clouds. In existing technologies, multi-camera acquisition is typically achieved through periodic command sending or master-slave triggering at the application layer. However, at the recognition algorithm level, the following problems commonly exist:

[0003] First, it suffers from the problems of getting stuck in local optima and high time consumption. Traditional point cloud global registration algorithms are prone to getting stuck in local optima when dealing with objects like locks, which are highly symmetrical and whose local features are easily occluded or missing, resulting in incorrect flip poses. If the number of iterative searches of the entire point cloud is blindly increased in order to improve accuracy, it will lead to extremely long recognition times per session, making it difficult to meet the time requirements of industrial automation systems.

[0004] Secondly, there is the risk of physical collisions. The locks that need to be identified at the work site are placed irregularly on the identification platform, which leads to the loss of prior pose and the unknown grasping point of the robotic arm. Due to the inconsistency in the type and size of the locks, the fixed safety height cannot adaptively avoid obstacles, which can easily cause the end effector of the robotic arm to physically collide with the metal locks during descent, thereby damaging the automated hardware system.

[0005] Third, the system suffers from high coupling and poor robustness. The computational load for registering large-scale point clouds is enormous. If the hardware acquisition and algorithm scheduling processes are not effectively decoupled, it can lead to latency during concurrent processing. Furthermore, existing vision service interfaces often lack a pre-emptive global error-proofing mechanism. When the host computer sends abnormal requests (such as unknown model numbers or invalid region parameters), erroneous instructions are directly transmitted to the underlying algorithm engine, easily causing point cloud read out-of-bounds errors, memory overflows, or process freezes, severely impacting the overall automated workflow.

[0006] It is worth noting that although multi-resolution point cloud registration or hierarchical ICP strategies have been explored in the field of general 3D reconstruction, existing solutions are usually designed for conventional objects with rich features and anisotropic structures, and lack active truncation constraints on the matching process. For metal structural components such as latches, which have high symmetry and severely missing local features due to stacking occlusion, directly applying the above strategies not only fails to solve the local optimum problem, but also leads to computational redundancy due to indiscriminate iteration. Therefore, how to organically combine the coarse matching truncation mechanism with the double-step ICP fine registration in the latch recognition scenario to balance recognition accuracy and real-time performance has become a technical challenge that urgently needs to be solved in this field.

[0007] Therefore, there is a need for a lock concurrent recognition method and system that can balance extreme matching speed, high recognition rate, and high system error prevention robustness. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to provide a lock-and-tie concurrent recognition method and system based on multi-level point cloud matching. By introducing a unified error collection mechanism and an independent configuration distribution strategy at the interface layer, it achieves secure interception of service requests and decoupling of computing power isolation. At the underlying recognition algorithm layer, by constructing a multi-level matching architecture that combines "global coarse matching" and "double-step constraint fine matching", it eliminates a massive amount of invalid matrix operations without sacrificing accuracy.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A lock-and-latch concurrent recognition system based on multi-level point cloud matching, comprising:

[0011] The global request error prevention interception unit is used to receive external visual recognition requests and build a unified error collector to perform full verification and interception of multi-dimensional parameters in the request, and collect and merge all exceptions before throwing them out.

[0012] A configuration distribution and scheduling unit is connected to the global request fail-safe interception unit. It is used to dynamically load configuration parameters after the full verification interception is passed, and to schedule the underlying computing power in an asynchronous non-blocking sub-process manner to achieve multi-task computing power isolation.

[0013] The point cloud dimensionality reduction and feature preprocessing unit is used to calculate the normal vector and perform multi-resolution voxel downsampling on the point cloud to be identified, while preserving sharp features such as edges and holes.

[0014] A multi-level core matching unit is connected to the point cloud dimensionality reduction and feature preprocessing unit. It is used to perform global coarse matching based on the sparse features of the preprocessed point cloud to obtain the initial pose, and to perform fine registration and scoring through a restricted iterative double-step ICP fine registration stage.

[0015] The dynamic safety collision avoidance assessment unit, connected to the multi-level core matching unit, is used to dynamically adjust and output safety compensation coordinates based on the height difference between the highest physical point of the environmental point cloud and the target pose using an interpolation algorithm.

[0016] Furthermore, the global request error prevention interception unit is configured to: when encountering an invalid parameter, collect all exceptions into an error list, and merge and throw all errors in a single response through the same communication interface; wherein, the exceptions include out-of-bounds region IDs, missing candidate models, and invalid point cloud paths.

[0017] Furthermore, the configuration distribution and scheduling unit is configured to dynamically read an independent configuration file (YAML configuration file) and apply the corresponding coordinate transformation coefficients and safety height thresholds. This configuration file contains specific calibration parameters for this operating condition, such as the dynamic safety height threshold, the fixed offset angle of the camera coordinate system, and the point cloud downsampling resolution.

[0018] Furthermore, the point cloud dimensionality reduction and feature preprocessing unit is configured to: calculate the normal vector using a K-nearest neighbor search method with a preset neighborhood size; retain the latching edge features based on the K-nearest neighbor search normal vector calculation method; and losslessly transform the data structure with the normal vector and bridge it to the target registration engine.

[0019] Furthermore, the multi-level core matching unit is configured to: strictly truncate the maximum search time of global random sampling in the coarse matching stage; and set a convergence iteration upper limit in the subsequent double-step ICP fine registration stage to truncate invalid calculations.

[0020] Furthermore, the dual-step ICP fine registration stage is configured as follows: it includes an initial mesh stretching stage with a large search radius and an ultrafine locking stage with a small search radius; and the ultrafine locking stage is set with a maximum number of iterations. When the number of iterations reaches the maximum number of iterations and the convergence condition is still not met, the current candidate pose is directly eliminated.

[0021] Furthermore, the dynamic safety collision avoidance evaluation unit is configured to: extract a preset minimum difference threshold T. min With the maximum difference threshold T max When the target height difference is within the range, a linear proportional mapping method is used, according to the formula: H=H min +(H max -H min )×(ΔZ-T) min ) / (T max -T min ), real-time generation between the minimum safe height H min With maximum safe height H max The dynamic Z-axis command H is issued between them.

[0022] This invention also provides a method for concurrent lock identification based on multi-level point cloud matching, comprising the following steps:

[0023] S1. Multi-dimensional error prevention verification: Receive external visual recognition requests, build a unified error collector, and perform non-blocking batch error prevention verification on environmental identifiers, area identifiers, candidate model library and local file path parameters through a unified error collection mechanism; if multiple illegal parameters exist, package all error information and reject the request at once.

[0024] S2. Concurrent Scheduling: After the verification is passed, the corresponding environment configuration parameters are dynamically loaded, and the underlying algorithm subprocess is started;

[0025] S3. Point cloud preprocessing: Perform normal vector estimation and multi-resolution downsampling on the on-site point cloud;

[0026] S4. Global coarse matching: Perform coarse matching based on the global sampling matching algorithm to extract the initial pose matrix of the target latch;

[0027] S5. Bi-step ICP fine registration: Based on the initial pose matrix, a bi-step ICP fine registration stage is performed to perform fine registration with a limit on the maximum number of iterations and the search radius, and the highest-scoring target and its spatial pose are output; the bi-step ICP fine registration stage includes the initial mesh stretching stage and the ultra-fine locking stage.

[0028] S6. Dynamic safety compensation: Compare the highest physical point on site with the initial pose point, calculate the dynamic safety compensation height, and correct the final gripping coordinate system sent to the robotic arm.

[0029] Furthermore, the initial mesh stretching stage uses a large search radius for global convergence;

[0030] The extremely fine locking phase uses a small search radius and a limited maximum number of iterations to perform extremely fine locking.

[0031] By integrating iteration error and overlap rate, the target lock model with the highest score and its corresponding six-degree-of-freedom spatial pose are output.

[0032] Further, step S6 specifically includes:

[0033] The highest physical point is obtained by traversing the point cloud of the site, and the height difference between the locking matching feature point and the highest physical point is calculated; based on the preset interpolation strategy, according to the formula: H=H min +(H max -H min )×(ΔZ-T) min ) / (T max -T minThe target is set as the safe gripping height H for dynamic compensation and adaptive spatial pose. Based on the height threshold, it is determined whether the target needs to undergo mechanical arm posture adjustment or meets the adsorption conditions, and the final gripping control command is output.

[0034] By adopting the above technical solution, the beneficial effects of the present invention are as follows:

[0035] 1. A high-pass global request error prevention interception unit constructs a unified error collector to perform non-blocking batch verification of multi-dimensional parameters. When illegal parameters are encountered, all exceptions are merged and thrown out in a single instance, achieving millisecond-level full interception, eliminating the risk of underlying algorithm engine crashes and robotic arm malfunctions, and improving the reliability of the system of this invention.

[0036] 2. By dynamically loading independent configuration files through the configuration distribution and scheduling unit, the underlying computing power is scheduled in an asynchronous non-blocking child process manner, realizing the isolation of multi-task computing power and effectively solving the problems of resource competition and blocking delay during collaborative operations.

[0037] 3. Through a hierarchical architecture of point cloud dimensionality reduction preprocessing, global coarse matching truncation constraints, and dual-step ICP fine registration, the computational load of global registration is reduced by more than 60%, the time for a single recognition is reduced from 850ms to 280ms, and the recognition success rate is increased from 82% to 97.5%, which is conducive to the high real-time recognition performance of the system of the present invention.

[0038] 4. A dynamic safety height compensation mechanism that is linked to multi-level matching results is introduced. Based on the height difference between the highest physical point on site and the target recognition pose, an adaptive safety height is generated in real time using a linear interpolation strategy. Compared with a fixed safety height or coarse compensation based solely on the highest point, it can more accurately avoid physical collision interference in complex stacking scenarios and improve safety. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the concurrent identification method for locks based on multi-level point cloud matching in Embodiment 1 of the present invention.

[0040] Figure 2 This is a structural block diagram of the lock concurrent recognition system based on multi-level point cloud matching in Embodiment 2 of the present invention. Detailed Implementation

[0041] Reference Figures 1 to 2 The embodiments of the present invention will be further described below.

[0042] Example 1: System Fault Prevention and Concurrency Scheduling Based on Edge Computing Platform

[0043] This embodiment describes in detail the specific deployment and workflow of the multi-level point cloud matching-based lock concurrent recognition system of the present invention on an edge computing platform.

[0044] I. Hardware Deployment Environment

[0045] In this embodiment, the system of the present invention is deployed on an industrial edge computing platform with multi-core processing capabilities. This computing platform is connected to an external host computer (such as an industrial PLC or bus controller) via industrial Ethernet or fieldbus to receive and respond to visual recognition commands in real time.

[0046] II. How the Global Request Foolproof Interception Unit Works

[0047] In this embodiment, when an external control device sends an identification request to the system, it first enters the global request error prevention interception unit. This unit at the interface layer monitors and verifies the request payload. The global request error prevention interception unit employs a non-blocking batch verification mechanism and simultaneously verifies the following parameters:

[0048] Environment identifier: Verify whether the environment ID in the request is in the system registration list;

[0049] Raise the area marker: Verify whether the area parameters exceed the preset physical boundaries;

[0050] Candidate lock set: Verify whether the candidate model name exists in the model library;

[0051] Point cloud data path: Verify that the local file path exists and is readable.

[0052] If any illegal parameters are detected in the request, the error-prevention interception unit, without initializing the underlying algorithm module or consuming disk I / O resources, collects all illegal parameter items (including out-of-bounds region IDs, missing candidate models, and invalid point cloud paths) into an error list. Specifically, the global request error-prevention interception unit constructs a list-type data structure in memory as the error list; during the verification process, for each illegal parameter encountered, a key-value pair containing an error code and error description is generated and appended to this list; when all parameters have been verified, if the list is not empty, it is serialized into an error response body through the same communication interface and thrown all at once, terminating the subsequent identification process. Subsequently, the system packages multiple error messages together and sends them back to the external device in a single response.

[0053] III. Specific Operating Methods of the Configuration Distribution and Scheduling Unit

[0054] In this embodiment, for valid recognition requests, the configuration distribution and scheduling unit parses the corresponding external parameter configuration file (YAML configuration file) based on the environment identifier carried in the request to extract specific calibration parameters for that working condition (such as dynamic safety height threshold, fixed offset angle of camera coordinate system, etc.). After the parameters are loaded, the configuration distribution and scheduling unit uses the operating system's underlying process management mechanism (such as fork or CreateProcess) to independently derive a restricted algorithm subprocess for the request, instead of directly calling the algorithm module in the main thread. Each subprocess is assigned independent CPU core affinity and memory limit. Through the above-mentioned process soft isolation and scheduling method, the independence of parameter configuration between different recognition requests is achieved; at the same time, the CPU computing power occupied by a single recognition task is effectively limited; and the risk of system crash or response stagnation due to local algorithm anomalies or high concurrency requests is avoided.

[0055] In this embodiment, under concurrent scenarios, when the system receives multiple recognition requests within a short period, the configuration distribution and scheduling unit spawns a corresponding algorithm subprocess for each legitimate request. The operating system kernel schedules these subprocesses to different CPU cores for parallel execution based on the current system load and preset process priorities. Since each subprocess has its own independent memory address space and resource quota, the processing of any request (including point cloud data loading, matching calculations, etc.) will not block or interfere with the processing threads of other requests, thus achieving true concurrent recognition capability at the system level.

[0056] Through the above implementation method, in the concurrent stress test, when 10 requests containing illegal parameters are received at the same time, the system completes full verification within 10 milliseconds and returns a unified error code, and the underlying algorithm process is not started; in the scenario of legal concurrent requests, each request runs independently in different sub-processes, the CPU usage of a single recognition task is limited to within 25%, and the system as a whole does not experience blocking or freezing.

[0057] Example 2: Multi-level point cloud matching and dynamic security compensation

[0058] This embodiment describes in detail the specific working method of the multi-level core matching unit and dynamic safety collision avoidance evaluation unit of the present invention.

[0059] I. Point Cloud Dimensionality Reduction and Feature Preprocessing

[0060] In this embodiment, after the underlying multi-level core matching unit acquires the preprocessed local point cloud, the point cloud dimensionality reduction and feature preprocessing unit first constructs a neighborhood based on KD-Tree search. It then uses a K-nearest neighbor-based normal vector calculation method (e.g., k=10) to calculate precise normal vectors, preserving the sharp features of the latch edges and holes. Subsequently, multi-resolution voxel filtering technology (e.g., voxel sizes set to 5mm, 2mm, and 1mm respectively) is used to construct feature point cloud libraries of different scales to balance feature representation accuracy with system computational burden.

[0061] II. Global coarse matching

[0062] In this embodiment, during the global coarse matching stage, the system executes a global sampling matching algorithm at extremely sparse feature levels (e.g., voxel size of 5 mm). To control the computational scale, the system sets an internal timer to 100 milliseconds or a sampling threshold of 5000 times, imposing strict truncation constraints on the global search process of coarse matching. Within this constrained period, the algorithm quickly calculates the approximate translation matrix and basic rotational pose of the target latch relative to the model library, thereby generating a reliable initial spatial pose.

[0063] The truncation mechanism of this invention checks the system timer or counter at the beginning of each iteration or sampling loop. When a preset threshold is reached, the algorithm does not simply terminate, but evaluates the matching scores of all currently calculated candidate transformation matrices and selects the one with the highest score as the coarse matching result. Furthermore, the truncation mechanism also sets a preset minimum confidence score threshold; if the highest selected score is still lower than the minimum confidence score threshold when the preset threshold is reached, the matching is deemed unreliable in the current scenario, and a matching failure is returned. If no valid candidates have been calculated at the time of truncation, a matching failure is returned.

[0064] III. Two-step ICP fine registration

[0065] In this embodiment, during the double-step ICP fine registration stage, the system starts with the initial spatial pose output from the initial mesh stretching stage (coarse matching) and initiates double-step ICP fine registration on a higher resolution feature point cloud (2mm voxel size). This fine registration process is divided into two consecutive sub-processes:

[0066] Initial mesh stretching stage: The system sets a large neighborhood search radius for initial constraints, prompting the source point cloud to converge quickly to the target surface in order to eliminate initial spatial deviation;

[0067] The fine-grained locking phase: The system reduces the search radius to a preset microscale (e.g., to 0.5 mm) and sets strict convergence error thresholds and maximum iteration limits (e.g., a convergence error threshold of 0.05 mm and a maximum iteration limit of 30). During fine-grained registration, if the system determines that the matching error tends to oscillate, or if the convergence condition is not met even after reaching the maximum iteration limit, the termination mechanism is immediately triggered to prematurely terminate invalid matching calculations.

[0068] In a specific implementation, the fine-grained registration stage of the two-step ICP can be implemented as follows: First, ICP is run with the first set of parameters (e.g., maximum corresponding point search distance of 5mm, transformation matrix change threshold of 0.5mm, convergence error threshold of 0.5mm, and maximum number of iterations of 50) to obtain the transformation matrix T1; then, using T1 as the initial pose, ICP is automatically switched to the second set of more stringent parameters (e.g., maximum corresponding point search distance = 0.5mm, transformation matrix change threshold = 0.05mm, convergence error threshold = 0.05mm, and maximum number of iterations = 30) and run again. If the second stage iterations exceed 30 times and still fail to converge, the candidate pose registration is directly determined to have failed.

[0069] By fusing the final iteration error with the point cloud overlap rate, each candidate lock model is scored. The specific scoring rule is as follows: Calculate the score for each candidate pose: Score = W1 × (1 - E) fitness )+W2×R overlap E fitness R represents the normalized fitness error. overlap W1 and W2 are the point cloud overlap ratios, and W1 and W2 are preset weighting coefficients (e.g., both 0.5). Output the target lock model with the highest score and its corresponding six-DOF spatial pose.

[0070] By employing the aforementioned coarse-to-fine and hierarchical computational architecture, computational redundancy in the spatial matrix iteration process is reduced, significantly decreasing the time required for a single lock recognition while ensuring recognition accuracy.

[0071] IV. Dynamic Security Compensation

[0072] In this embodiment, after receiving the above optimal matching result, the dynamic safety collision avoidance evaluation unit performs the following steps:

[0073] a. Traverse the original point cloud of the site to obtain the Z-coordinate value of the highest physical point. max ;

[0074] b. Calculate the Z-coordinate of the center point of the locking matching feature point. target ;

[0075] c. Calculate the height difference ΔZ = Z max -Ztarget ;

[0076] d. Extract the preset minimum difference threshold T min =20mm and the maximum difference threshold T max =100mm;

[0077] e. When T min ≤ΔZ≤T max At that time, a linear scaling method is used to generate data in real time between the minimum safe height H. min =50mm and maximum safe height H max The dynamic Z-axis command H is issued within a range of 150mm, and the calculation formula is: H = H min +(H max -H min )×(ΔZ-T) min ) / (T max -T min )

[0078] If ΔZ < T min If the target is determined to be severely occluded, an attitude adjustment command is output; if ΔZ > T max Adopting the maximum safe height H max =150mm for compensation.

[0079] Using the above implementation method, tests were conducted in complex scenarios containing stacked latches: the total time for a single recognition was reduced from an average of 850ms in the traditional method to 280ms (a reduction of approximately 67%); the recognition success rate was increased from 82% to 97.5%; the dynamic safety compensation mechanism effectively avoided collisions between the end effector and the protruding latches, and no physical interference occurred during 5000 grasping tests.

[0080] In summary, this invention achieves highly applicable and real-time concurrent lock recognition on an industrial edge computing platform. Under normal system operation conditions, the pre-implemented error-proofing interception mechanism and isolated process scheduling effectively block illegal commands from impacting the algorithm engine, ensuring system stability in a multi-tasking environment. Simultaneously, the multi-level point cloud matching design avoids the waste of computing power caused by full iteration, keeping the single recognition time extremely low, improving the system's recognition rate and reducing recognition time in cases of unknown lock poses.

[0081] In this invention, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, a direct connection, or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0082] The control method of this invention is automatic control through a controller. The control circuit of the controller can be implemented by simple programming by those skilled in the art. The power supply is also common knowledge in the art. Furthermore, since this invention is mainly used to protect mechanical devices, the control method and circuit connection will not be explained in detail here.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any ordinary changes and substitutions made by those skilled in the art within the scope of the technical solution of the present invention should be included within the protection scope of the present invention.

Claims

1. A concurrent lock recognition system based on multi-level point cloud matching, characterized in that, include: The global request error prevention interception unit is used to receive external visual recognition requests and build a unified error collector to perform full verification and interception of multi-dimensional parameters in the request. A configuration distribution and scheduling unit is connected to the global request error prevention and interception unit, which is used to dynamically load configuration parameters after the full verification interception is passed, and schedule the underlying computing power in a subprocess manner. The point cloud dimensionality reduction and feature preprocessing unit is used to perform normal vector calculation and multi-resolution voxel downsampling on the point cloud to be identified. A multi-level core matching unit, connected to the point cloud dimensionality reduction and feature preprocessing unit, is used to perform global coarse matching based on the sparse features of the preprocessed point cloud to obtain the initial pose, and to perform fine registration and scoring through a restricted iterative bi-step ICP fine registration stage; the bi-step ICP fine registration stage includes an initial mesh stretching stage and an extremely fine locking stage. The dynamic safety collision avoidance assessment unit, connected to the multi-level core matching unit, is used to dynamically adjust and output safety compensation coordinates based on the height difference between the highest physical point of the environmental point cloud and the target pose using an interpolation algorithm.

2. The lock-and-latch concurrent recognition system based on multi-level point cloud matching according to claim 1, characterized in that, The global request error prevention interception unit is configured to: when encountering invalid parameters, collect all exceptions into an error list, and merge and throw all errors in a single response through the same communication interface; wherein, the exceptions include out-of-bounds region IDs, missing candidate models, and invalid point cloud paths.

3. The lock-and-latch concurrent recognition system based on multi-level point cloud matching according to claim 1, characterized in that, The configuration distribution and scheduling unit is configured to dynamically read independent configuration files and apply corresponding coordinate transformation coefficients and safety height thresholds.

4. The lock-and-latch concurrent recognition system based on multi-level point cloud matching according to claim 1, characterized in that, The point cloud dimensionality reduction and feature preprocessing unit is configured to: employ a normal vector calculation method based on K-nearest neighbor search to preserve latching edge features, and losslessly transform the data structure with normal vectors and bridge it to the target registration engine.

5. A lock-and-latch concurrent recognition system based on multi-level point cloud matching according to claim 1, characterized in that, The multi-level core matching unit is configured to: truncate the maximum search time of global random sampling in the coarse matching stage; and set a convergence iteration upper limit in the subsequent double-step ICP fine registration stage to truncate invalid calculations.

6. A lock-and-latch concurrent recognition system based on multi-level point cloud matching according to claim 1, characterized in that, The fine-grained locking stage is set with a maximum number of iterations. When the number of iterations reaches the maximum number of iterations and the convergence condition is still not met, the current candidate pose is directly eliminated.

7. A lock-and-latch concurrent recognition system based on multi-level point cloud matching according to claim 1, characterized in that, The dynamic safety collision avoidance assessment unit is configured to: extract a preset minimum difference threshold T. min With the maximum difference threshold T max When the target height difference is within the range, a linear scaling method is used to generate a value between the minimum safe height H in real time. min With maximum safe height H max The dynamic Z-axis command H is issued between them.

8. A concurrent lock identification method based on multi-level point cloud matching, characterized in that, Includes the following steps: S1. Multi-dimensional error prevention verification: Receive external visual recognition requests, build a unified error collector, and perform non-blocking batch error prevention verification on environmental identifiers, area identifiers, candidate model library and local file path parameters through a unified error collection mechanism; if multiple illegal parameters exist, package all error information and reject the request at once. S2. Concurrent Scheduling: After the verification is passed, the corresponding environment configuration parameters are dynamically loaded, and the underlying algorithm subprocess is started; S3. Point cloud preprocessing: Perform normal vector estimation and multi-resolution downsampling on the on-site point cloud; S4. Global coarse matching: Coarse matching is performed based on a global sampling matching algorithm with a time or iteration count truncation mechanism to extract the initial pose matrix of the target latch; S5. Bi-step ICP fine registration: Based on the initial pose matrix, a bi-step ICP fine registration stage is performed to perform fine registration with a limit on the maximum number of iterations and the search radius, and the highest-scoring target and its spatial pose are output; the bi-step ICP fine registration stage includes the initial mesh stretching stage and the ultra-fine locking stage. S6. Dynamic safety compensation: Compare the highest physical point on site with the initial pose point, calculate the dynamic safety compensation height, and correct the final gripping coordinate system sent to the robotic arm.

9. A lock-and-latch concurrent identification method based on multi-level point cloud matching according to claim 8, characterized in that, The initial mesh stretching stage uses a large search radius for global convergence; The extremely fine locking phase uses a small search radius and a limited maximum number of iterations to perform extremely fine locking. By integrating iteration error and overlap rate, the target lock model with the highest score and its corresponding six-degree-of-freedom spatial pose are output.

10. A lock-and-latch concurrent identification method based on multi-level point cloud matching according to claim 8, characterized in that, Step S6 specifically includes: The highest physical point is obtained by traversing the point cloud of the site, and the height difference between the locking matching feature point and the highest physical point is calculated; based on the preset interpolation strategy, according to the formula: H=H min +(H max -H min )×(ΔZ-T) min ) / (T max -T min The target is set as the safe gripping height H for dynamic compensation and adaptive spatial pose. Based on the height threshold, it is determined whether the target needs to undergo mechanical arm posture adjustment or meets the adsorption conditions, and the final gripping control command is output.

Images