Staged lightweight attitude estimation method in multi-agent cooperative system
Through a phased lightweight posture estimation method, using a lightweight target detection network and key point detection combined with a nonlinear optimization algorithm, the accuracy and real-time problems of posture estimation in multi-agent collaborative systems are solved, and efficient posture recognition and collaborative operations are achieved in resource-constrained environments.
Patent Information
- Application Number
- CN202510885639.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2025-10-10
AI Technical Summary
In multi-agent collaborative systems, especially in outdoor operation scenarios, traditional multi-agent collaborative technologies have difficulty achieving high-precision, low-latency posture estimation, resulting in insufficient reliability of collaborative operations between agents. In particular, on resource-constrained mobile platforms, pure visual perception is unable to meet the posture recognition requirements over the entire distance range.
A staged lightweight attitude estimation method is adopted. Through a two-stage collaborative mechanism of long-distance attitude range estimation and short-distance attitude calibration, a lightweight target detection network and key point detection are combined with a nonlinear optimization algorithm to perform attitude estimation and calibration at long and short distances respectively, reducing computational overhead and compensating for motion accumulated errors.
It significantly improves the accuracy and real-time performance of posture recognition, ensures that the executor and target can efficiently complete collaborative tasks in complex outdoor scenes, reduces dependence on depth cameras and computing resources, and is suitable for mobile platforms with limited computing power.
Smart Images

Figure CN120765728A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to multi-agent collaboration technology, and in particular to a phased lightweight posture estimation method in a multi-agent collaboration system. Background Art
[0002] In multi-agent collaborative systems, especially in outdoor operations, the posture relationship between agents plays a critical role in the smooth execution of tasks. For example, in unmanned vehicle-robot collaborative systems, the vehicle frequently needs to be loaded with a robot, and the robot's accuracy in recognizing the vehicle's posture is crucial for successful loading. Furthermore, the robot's recognition of the vehicle's posture also places high demands on the range of recognition. However, on resource-constrained mobile platforms, traditional multi-agent collaborative technologies struggle to achieve high-precision, low-latency posture estimation across the entire range in dynamic environments using only visual perception, thus failing to ensure the reliability of collaborative operations between agents.
[0003] It should be noted that the information disclosed in the above background technology section is only used to understand the background of this application, and therefore may include information that does not constitute prior art known to ordinary technicians in this field. Summary of the Invention
[0004] The main purpose of the present invention is to overcome the defects existing in the above-mentioned background technology and provide a staged lightweight posture estimation method in a multi-agent collaborative system.
[0005] To achieve the above object, the present invention adopts the following technical solutions:
[0006] A phased lightweight pose estimation method in a multi-agent collaborative system comprises the following steps:
[0007] S1. Long-distance pose range estimation: The actuator captures the target using its onboard visual sensor, identifies the target based on the captured image using a lightweight object detection network, and uses a classification network to classify the target's pose into discrete categories to obtain a rough pose range of the target.
[0008] S2. Close-range pose calibration: When the actuator detects that the distance between it and the target is less than a predetermined range, a rough pose estimate of the target is obtained by a marker detection algorithm based on preset markers or natural feature points of the target in the captured image. The rough pose estimate is used as the initial pose of the target, and the pose of the target is iteratively updated by key point detection combined with a nonlinear optimization algorithm to obtain a precise pose estimate of the target.
[0009] Among them, dynamic posture tracking is performed through a two-stage collaborative mechanism, a rough posture range is obtained in the long-distance stage, and the motion cumulative error is compensated in the close-distance stage, thereby realizing real-time lightweight posture tracking until the docking collaboration task between the execution body and the target body is completed.
[0010] Furthermore, in step S1, the classification network uses channel attention and spatial attention mechanisms to enhance feature extraction capabilities, and strengthens the discrimination of non-adjacent posture categories through a customized loss function.
[0011] Furthermore, the customized loss function introduces a directional difference weight factor to impose a reverse penalty on category differences exceeding a critical value to reduce large-angle errors.
[0012] Furthermore, in step S1, the classification network outputs multiple discrete categories, each discrete category corresponds to a direction interval, wherein a specific category corresponds to a predetermined docking area of the target body, and the classification result is used to guide the execution body from the current direction interval toward the predetermined docking area of the target body.
[0013] Furthermore, step S2 specifically includes:
[0014] Using preset markers or natural feature points of the target as a reference, spatial features are obtained through a key point extractor;
[0015] The nonlinear optimization function is constructed using the projection error minimization principle, and the posture parameters are updated through iterative solution.
[0016] Furthermore, the nonlinear optimization function is constructed by minimizing the reprojection error of spatial feature points, and an iterative optimization algorithm is used to solve the optimal pose parameters:
[0017] Project the three-dimensional coordinates of the key points onto the image plane to generate theoretical projection points;
[0018] Calculate the Euclidean distance error between the theoretical projection point and the actual detection key point;
[0019] The pose parameters are iteratively adjusted by the gradient descent method to make the accumulated reprojection error converge to the minimum value.
[0020] Furthermore, the key point extractor adopts ORB feature descriptor, and the nonlinear optimization converges to an exact solution after multiple iterations.
[0021] Furthermore, the preset mark in step S2 is an AR code, and the natural feature points are edge feature points of the target body; the natural feature points serve as alternative benchmarks when AR code recognition fails.
[0022] Further, in step S2, the executor estimates the distance to the target body based on the target frame size output by the target detection network in real time, and triggers the close-range pose calibration operation when the target frame size is greater than a predetermined threshold.
[0023] A computer program product comprising a computer program which, when executed by a processor, implements the multi-agent collaborative system in the phased lightweight pose estimation method.
[0024] The present application has the following beneficial effects:
[0025] The present application proposes a multi-agent collaborative system in the phased lightweight pose estimation method, through the two-stage collaborative mechanism of long-distance pose range estimation and close-range pose calibration, the accuracy and real-time performance of the executor (such as a robot) in the multi-agent collaborative system for target body (such as an unmanned vehicle) pose recognition are significantly improved: in the long-distance stage, a lightweight classification network is used to replace the traditional regression algorithm, and the target body pose is divided into discrete categories to output a rough pose range, which greatly reduces the computational overhead; in the close-range stage, key point detection is performed based on a preset marker or natural feature point, and a non-linear optimization algorithm is used to iteratively compensate for the cumulative error of motion, effectively overcoming the recognition deviation in a dynamic environment, and achieving high-precision pose calibration. The solution of the present application not only overcomes the dependence of the prior art on a depth camera and intensive computing power, but also realizes the full-process coverage from long-distance preliminary estimation to close-range accurate recognition on a low-resource mobile platform through lightweight design, while achieving multiple optimizations of deployment convenience, recognition accuracy and system economy, ensuring that the executor and the target body can efficiently complete the collaborative work task in a complex outdoor scene.
[0026] Other beneficial effects of the embodiments of the present application will be further described below. BRIEF DESCRIPTION OF DRAWINGS
[0027] Figure 1 The present application is a phased lightweight pose estimation method.
[0028] Figure 2 The present application is a rough pose 8 classification diagram.
[0029] Figure 3 The present application is a first-stage network structure diagram.
[0030] Figure 4 The present application is a key point detection-based pose calibration algorithm block diagram.
[0031] Figure 5 The present application is a key point detection-based pose calibration result diagram.
[0032] Figure 6This is an overall flow chart of the staged lightweight posture estimation method in the multi-agent collaborative system of the present invention. DETAILED DESCRIPTION
[0033] The following is a detailed description of the embodiments of the present invention. It should be emphasized that the following description is only exemplary and is not intended to limit the scope of the present invention and its application.
[0034] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present invention, "plurality" means two or more, unless otherwise specifically defined.
[0035] This paper proposes a phased, lightweight pose estimation method for multi-agent collaborative systems, aiming to address the problem of lightweight, purely visual pose estimation in multi-agent collaborative systems. Initial pose estimation is performed at long range, followed by high-precision pose recognition at close range. This innovative two-stage solution, from long range to close range, achieves both long-range pose estimation and close-range pose correction, ensuring that the actuator can accurately dock with the target (e.g., but not limited to, a robot successfully completing the action of docking with an unmanned vehicle).
[0036] See Figure 1 and Figure 6 The embodiment of the present invention provides a staged lightweight posture estimation method in a multi-agent collaborative system, comprising the following steps:
[0037] Step S1, long-distance pose range estimation: The actuator (e.g., a robot) captures a target (e.g., an unmanned vehicle) using its onboard visual sensor. Based on the captured image, the target is identified using a lightweight object detection network. A classification network is used to classify the target's pose into discrete categories, thereby obtaining a rough pose range for the target.
[0038] In some embodiments, in step S1, the classification network employs channel attention and spatial attention mechanisms to enhance feature extraction capabilities, and employs a customized loss function to enhance the discrimination of non-adjacent pose categories. Preferably, the customized loss function incorporates a directional difference weighting factor, imposing a negative penalty on category differences exceeding a critical value to reduce large angular errors.
[0039] In some embodiments, in step S1, the classification network outputs multiple discrete categories, each discrete category corresponds to a direction interval (for example, a total of 8 direction intervals), where a specific category corresponds to a predetermined docking area of the target body, and the classification result is used to guide the execution body from the current direction interval toward the predetermined docking area of the target body.
[0040] Step S2, close-range posture calibration: When the executor detects that the distance between it and the target object is less than a predetermined range, a rough posture estimation value of the target object is obtained through a marker detection algorithm based on preset markers or natural feature points of the target object in the captured image, and the rough posture estimation value is used as the initial posture of the target object. The posture of the target object is iteratively updated through key point detection combined with a nonlinear optimization algorithm, thereby obtaining an accurate posture estimation value of the target object.
[0041] Among them, dynamic posture tracking is carried out through a two-stage collaborative mechanism, and the rough posture range of the target body is obtained in the long-distance stage, replacing continuous posture regression to reduce computing overhead and save computing resources. In the close-range stage, the accumulated motion error is compensated to achieve real-time lightweight posture tracking until the docking collaboration task between the executor and the target body is completed (such as the robot completes the boarding action).
[0042] In some embodiments, step S2 specifically includes: using preset markers or natural feature points of the target body (such as the body of an unmanned vehicle) (such as the edge features of the body) as a reference, obtaining spatial features through a key point extractor; constructing a nonlinear optimization function using the projection error minimization principle, and updating the posture parameters through iterative solution.
[0043] In some embodiments, the nonlinear optimization function is constructed by minimizing the reprojection error of spatial feature points, and an iterative optimization algorithm is used to solve the optimal posture parameters: the three-dimensional coordinates of the key points are projected onto the image plane to generate theoretical projection points; the Euclidean distance error between the theoretical projection points and the actual detected key points is calculated; the posture parameters are iteratively adjusted through the gradient descent method so that the cumulative reprojection error converges to a minimum value.
[0044] In some embodiments, the key point extractor uses an ORB feature descriptor, and the nonlinear optimization converges to an exact solution after multiple iterations.
[0045] In some embodiments, the preset mark in step S2 is an AR (augmented reality) code, and the natural feature points are edge feature points of the target object (such as the body of an unmanned vehicle); the natural feature points serve as alternative benchmarks when AR code recognition fails.
[0046] In some embodiments, in step S2, the execution body estimates the distance to the target body in real time based on the target frame size output by the target detection network, and triggers a close-range posture calibration operation when the target frame size is greater than a predetermined threshold.
[0047] Compared with the traditional posture estimation method through regression, which consumes more computing power, the staged lightweight posture estimation method of the present invention is designed in two stages: posture range estimation and posture calibration. When the precise posture of the target body is not needed at a long distance, only the approximate posture range is estimated by the classification method. When the distance is close to a certain extent, the initial posture is obtained based on the marker detection algorithm, and the precise posture of the target body is obtained through iterative calibration through key point detection and nonlinear optimization algorithm. Through the innovative two-stage collaborative mechanism, the precise posture of the target body is obtained while the consumption of computing power resources is greatly reduced, and real-time lightweight posture tracking is realized until the task of docking between the executor and the target body is completed, thereby improving the accuracy and real-time performance of posture recognition, and is particularly suitable for scenarios with limited computing power.
[0048] The following further describes specific embodiments of the present invention, its algorithm examples and experimental verification.
[0049] A phased lightweight posture estimation method in a multi-agent collaborative system. By designing a two-stage solution from long distance to short distance, long-distance posture estimation and short-distance posture correction are achieved respectively. Taking the robot boarding task as an example, this method ensures that the robot can successfully board the unmanned vehicle. The complete process is as follows Figure 1 shown.
[0050] Phase 1: When the robot is far away from the unmanned vehicle, it uses a lightweight object detection network (YOLO-tiny) to search for the unmanned vehicle in its field of view. After the unmanned vehicle is found, its image is classified into eight categories, corresponding to eight coarse pose ranges, using a lightweight classification algorithm.
[0051] Figure 2 This is a schematic diagram of the rough 8-classification of postures, showing the service robot's rough posture classification scheme for the unmanned vehicle (Ego-Vehicle). The relative posture is discretized into 8 direction intervals (indicated by the arrows labeled 0-7 in the figure). The classification network outputs the approximate position of the unmanned vehicle, where category 0 corresponds to the preset docking area facing the vehicle door.
[0052] When designing the network loss function, the parameter Θ is introduced to strengthen the distinction between non-adjacent postures and reduce the possibility of large errors, thereby providing more accurate posture estimation preparation for subsequent boarding tasks.
[0053]
[0054] The robot's goal is to reach an area with a relative posture category of 0 relative to the unmanned vehicle, which also corresponds to a position directly opposite the vehicle's side door. In the early, longer-distance phase, the robot ensures it approaches the vehicle in the correct general direction. During this process, the size of the target box output by the object detection output is used to roughly estimate the real-time distance between the robot and the vehicle. When the distance closes within a certain range (approximately 2-3 meters), the algorithm enters the second phase.
[0055] Figure 3 Figure 1 shows the first-stage network structure. The first-stage pose classification network consists of an object detection module and subsequent processing steps. The network uses channel attention and spatial attention mechanisms in parallel to enhance feature extraction capabilities. Finally, a fully connected layer (FC) outputs the category to which the current pose belongs, achieving lightweight real-time classification.
[0056] Phase 2: In the first phase, the robot is ensured to face the door side when approaching the unmanned vehicle. When the robot is close to the unmanned vehicle, the AR code detection algorithm is activated. However, due to the movement of the robot and the unmanned vehicle, there is a large error in the AR code detection result. This solution draws on the idea of bundle gradient descent in the ORB SLAM algorithm, combines key point detection with the BundleAdjustment algorithm, and iterates the code to detect the key points.
[0057] Update the unmanned vehicle's posture. Using the rough posture obtained by AR code detection as the initial value, a more accurate posture estimate is obtained through optimization.
[0058]
[0059] Figure 4 This is a block diagram of the posture calibration algorithm based on key point detection in an embodiment of the present invention, which fully depicts the calculation process of close-range posture calibration: first, the initial posture of the external input (derived from the preset marker or natural feature point detection result) is used as the optimization starting point, and the current frame two-dimensional image and the known 3D-2D key point spatial correspondence are synchronously input; then the key point extractor detects and describes the feature points from the two-dimensional image to generate a 2D feature set, and uses the initial posture to project the 3D key points to the 2D image plane to generate the theoretical projection point coordinates, and calculates the position deviation between the theoretical projection point and the actual detected feature point; then the PnP algorithm is used to construct the reprojection error minimization problem, and iterative optimization is implemented through the filtering control module: the posture parameters are adjusted based on the bundle adjustment principle, the projection error is recalculated and it is iterated until the error converges; finally, the optimized high-precision posture estimate is output to meet the accuracy requirements of the collaborative task.
[0060] Example 1:
[0061] AR code is installed on the left side of the unmanned vehicle, the robot is equipped with a monocular camera, YOLO-tiny is used as a detector, a pose classification network is combined for long-distance pose range estimation, and after entering a close distance, pose calibration is completed through AR code recognition combined with a BundleAdjustment algorithm. The scheme realizes the fast approach and accurate boarding of the robot to the unmanned vehicle in an actual scene.
[0062] Example 2:
[0063] AR code is replaced by natural feature points (such as vehicle edge features) as a pose calibration reference, and a two-stage strategy of classification + optimization is still used for pose estimation. In an environment where light or shielding is not conducive to AR code recognition, the scheme can be used as an alternative deployment form to ensure the robustness of pose recognition and has good engineering adaptability.
[0064] Figure 5 The pose calibration result based on key point detection of the embodiment of the application is shown. Figure 5 The left and right images in the middle respectively represent the positions of 3D key points projected on 2D images according to the poses before and after calibration. Figure 5 As can be seen from the middle, after the pose calibration process of step S2, the obtained pose is closer to the real pose.
[0065] Finally, the actual test result shows that accurate poses can be obtained within 5 cycles, so as to realize lightweight real-time dynamic pose estimation until the robot boards.
[0066] In summary, the present invention proposes a phased lightweight posture estimation method in a multi-agent collaborative system. Through the phased posture estimation and calibration mechanism, the accuracy and real-time performance of the posture recognition of the target body (such as an unmanned vehicle) by the executor (such as a robot) in the multi-agent collaborative system are significantly improved. Compared with the existing regression algorithms that rely on depth cameras or intensive computing resources, the present invention innovatively achieves a technological breakthrough through a phased design: first, the traditional regression method is used to estimate the posture, which consumes a lot of computing power. In scenarios where precise posture is not required at a long distance, a lightweight classification network is used to divide the posture of the target body into discrete categories and output a rough posture range, which greatly saves computing resources; when the distance between the executor and the target body is close to a predetermined range, the posture calibration is performed based on the AR code, and the posture parameters are iteratively updated through key point detection and nonlinear optimization algorithms, and finally the posture estimation value that meets the accuracy requirements of the collaborative operation is obtained, which effectively compensates for the AR code recognition error and ensures that the executor can accurately dock with the target body (such as the robot successfully completes the boarding action). This phased collaborative mechanism not only achieves accurate posture estimation throughout the entire process, but also effectively overcomes the interference of accumulated motion errors in dynamic environments, enabling the algorithm to maintain high-precision dynamic tracking under low network parameter conditions. It is particularly suitable for mobile platforms with limited computing power, forming the triple advantages of deployment convenience, recognition accuracy and system economy, ensuring the efficient and reliable completion of collaborative tasks between the executor and the target body.
[0067] Experimental results show that the present invention can achieve high-precision dynamic posture estimation with low network parameters. It is suitable for mobile robot platforms with limited computing power and has significant advantages such as convenient deployment, accurate recognition and low cost.
[0068] An embodiment of the present invention further provides a storage medium for storing a computer program, which at least performs the above method when executed.
[0069] An embodiment of the present invention further provides a control device, comprising a processor and a storage medium for storing a computer program; wherein the processor is configured to execute at least the method described above when executing the computer program.
[0070] An embodiment of the present invention further provides a processor, which executes a computer program and at least performs the method described above.
[0071] The storage medium can be implemented by any type of nonvolatile storage device, or a combination thereof. The nonvolatile memory can be a Read Only Memory (ROM), a Programmable Read-Only Memory (PROM), an Erasable Programmable Read-Only Memory (EPROM), an Electrically Erasable Programmable Read-Only Memory (EEPROM), a Ferromagnetic Random Access Memory (FRAM), a Flash Memory, a magnetic surface storage, an optical disc or a Compact Disc Read-Only Memory (CD-ROM). The magnetic surface storage can be a magnetic disc memory or a magnetic tape memory. The storage medium described in the embodiments of the present application is intended to include, but is not limited to, these and any other suitable type of memory.
[0072] In several embodiments provided by the present application, it should be understood that the disclosed system and method can be implemented in other manners. The described device embodiments are merely schematic, and the division of the units is merely a logical function division. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling or direct coupling or communication connection between the components can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.
[0073] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place or distributed on a plurality of network units; some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.
[0074] In addition, each functional unit in the embodiments of the present application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in the form of hardware, or in the form of hardware plus software functional units.
[0075] Those skilled in the art will understand that all or part of the steps of the above-mentioned method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc. Various media that can store program codes.
[0076] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.
[0077] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.
[0078] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.
[0079] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.
[0080] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. Those skilled in the art will recognize that, without departing from the scope of the present invention, several equivalent substitutions or obvious variations can be made, and the performance or use of the same should be considered to fall within the scope of protection of the present invention.
Claims
1. A staged lightweight pose estimation method for robots in a multi-agent collaborative system, characterized in that The following steps are involved: S1. Long-distance pose range estimation: The actuator captures the target using its onboard visual sensor, identifies the target based on the captured image using a lightweight object detection network, and uses a classification network to classify the target's pose into discrete categories to obtain a rough pose range of the target. S2. Close-range pose calibration: When the actuator detects that the distance between it and the target is less than a predetermined range, a rough pose estimate of the target is obtained by a marker detection algorithm based on preset markers or natural feature points of the target in the captured image. The rough pose estimate is used as the initial pose of the target, and the pose of the target is iteratively updated by key point detection combined with a nonlinear optimization algorithm to obtain a precise pose estimate of the target. Among them, dynamic posture tracking is performed through a two-stage collaborative mechanism, a rough posture range is obtained in the long-distance stage, and the motion cumulative error is compensated in the close-distance stage, thereby realizing real-time lightweight posture tracking until the docking collaboration task between the execution body and the target body is completed.
2. The method according to claim 1, wherein In step S1, the classification network uses channel attention and spatial attention mechanisms to enhance feature extraction capabilities, and strengthens the discrimination of non-adjacent posture categories through a customized loss function.
3. The method according to claim 2, wherein The customized loss function introduces a directional difference weight factor and imposes a reverse penalty on category differences exceeding a critical value to reduce large-angle errors.
4. The method according to any one of claims 1 to 3, wherein In step S1, the classification network outputs multiple discrete categories, each discrete category corresponds to a direction interval, wherein a specific category corresponds to a predetermined docking area of the target body, and the classification result is used to guide the execution body from the current direction interval toward the predetermined docking area of the target body.
5. The method according to any one of claims 1 to 4, characterized in that Step S2 specifically includes: Using preset markers or natural feature points of the target as a reference, spatial features are obtained through a key point extractor; The nonlinear optimization function is constructed using the projection error minimization principle, and the posture parameters are updated through iterative solution.
6. The method according to claim 5, wherein The nonlinear optimization function is constructed by minimizing the reprojection error of spatial feature points, and an iterative optimization algorithm is used to solve the optimal pose parameters: Project the three-dimensional coordinates of the key points onto the image plane to generate theoretical projection points; Calculate the Euclidean distance error between the theoretical projection point and the actual detection key point; The pose parameters are iteratively adjusted by the gradient descent method to make the accumulated reprojection error converge to the minimum value.
7. The method according to claim 5 or 6, wherein: The key point extractor adopts ORB feature descriptor, and the nonlinear optimization converges to an exact solution after multiple iterations.
8. The method according to any one of claims 1 to 7, wherein: In step S2, the preset mark is an AR code, and the natural feature points are edge feature points of the target body; the natural feature points serve as alternative benchmarks when AR code recognition fails.
9. The method according to any one of claims 1 to 8, wherein In step S2, the execution body estimates the distance to the target body in real time based on the target frame size output by the target detection network, and triggers a close-range posture calibration operation when the target frame size is greater than a predetermined threshold.
10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the staged lightweight posture estimation method in a multi-agent collaborative system as described in any one of claims 1 to 9.