Posture recognition algorithm for any object under monocular camera and application system

Abstract

The invention provides a posture recognition algorithm for any object under a monocular camera and an application system, and the algorithm comprises the steps: S1, constructing a target three-dimensional model, carrying out the multi-view annular shooting image collection of a target, and generating a dense grid model through feature extraction, matching, posture calculation and a multi-view geometric method; s2, generating an image depth map, and predicting depth information of a target in a motion process based on a monocular image sequence; s3, extracting a target image mask, and generating a target area mask graph through an image encoder, a prompt encoder and a mask decoder; and S4, executing attitude estimation, performing attitude initialization, correction and screening by combining the three-dimensional model, the depth map and the mask map, and outputting a six-degree-of-freedom attitude result of the target. According to the method, the target is subjected to annular shooting modeling through the method based on multi-view geometry, the three-dimensional model of the target is generated, attitude estimation is achieved in combination with the image mask and the depth map, the generalization ability of an attitude estimation algorithm in an actual scene is improved, and the application range of the attitude estimation algorithm in the actual scene is widened.

Description

Technical Field

[0001] The present invention relates to the technical field of target posture recognition, and in particular to a posture recognition algorithm and application system for arbitrary objects using a monocular camera. Background Art

[0002] Object pose estimation is one of the core research tasks in the field of computer vision. Its goal is to recover the position and orientation (i.e., translation and rotation information) of an object in three-dimensional space from an image or video sequence. This task has wide application value in scenarios such as augmented reality, intelligent manufacturing, autonomous driving, and medical image navigation.

[0003] Traditional pose estimation methods typically rely on hand-crafted image feature extraction algorithms, such as SIFT and HOG, combined with the PnP (Perspective-n-Point) model to estimate the pose of an object in three-dimensional space. These methods can achieve high accuracy in simple structures or controlled lighting conditions, but they are unstable and have poor robustness in real-world scenarios such as complex backgrounds, occlusion interference, scale changes, or textureless objects.

[0004] With the development of deep learning, neural network-based end-to-end pose estimation methods have gradually emerged in recent years. They mainly fall into two categories: one is based on 2D keypoint detection and 3D regression, typified by OpenPose and PoseNet; the other directly predicts six-degrees-of-freedom (6DoF) poses through deep networks, such as PVNet, GDR-Net, and DenseFusion. However, most of these methods are based on category-specific model training, requiring a large number of labeled samples for supervised learning. These methods have weak generalization and are difficult to adapt to the rapid deployment and recognition of arbitrary objects in real scenes.

[0005] Traditional technology 1 [Application No.: 202411576347.0] discloses a 3D pose estimation method based on a monocular camera. By constructing a mixed dataset of synthetic and real objects, training the YOLOPose model to extract 2D key points, and fitting the GHUM three-dimensional human model, although this method has a certain accuracy in human pose recognition, it cannot estimate the pose of targets of other categories, let alone targets of unknown categories. At the same time, this method still requires special model training before it can be used for human pose estimation.

[0006] Traditional technology 2 [Application No.: 202411221949.4] proposed a "monocular six-degree-of-freedom pose estimation method for space targets", which obtains the relative pose between adjacent frames through image mask and depth estimation;

[0007] However, this method is only applicable to spacecraft objects moving in space, and cannot estimate the posture of targets of other categories, let alone targets of unknown categories. At the same time, this method only uses image input and estimated depth data, and does not apply information such as the target's three-dimensional data, resulting in low accuracy.

[0008] To this end, a posture recognition algorithm and application system for arbitrary objects using a monocular camera are proposed. Summary of the Invention

[0009] In view of this, the present invention provides a posture recognition algorithm and application system for arbitrary objects under a monocular camera to solve or alleviate the technical problems existing in the prior art and at least provide a beneficial option.

[0010] The technical solution of the present invention is implemented as follows: a posture recognition algorithm for any object under a monocular camera, the algorithm comprising the following steps:

[0011] Step S1, constructing a 3D model of the target: capturing circular images of the target and constructing a dense mesh model through feature extraction, image registration, camera pose calculation, and point cloud reconstruction;

[0012] Step S2, generating an image depth map: performing depth estimation on the image sequence during target motion, and completing scale transformation and three-dimensional mapping;

[0013] Step S3, extracting the target image mask: interactively clicking in the target image, and generating a target area mask map through image coding and prompt information;

[0014] Step S4, perform pose estimation: combine the target 3D model, image depth map and target area mask map to construct an initial pose hypothesis set, perform pose initialization, pose correction and pose screening, and output the target's six-degree-of-freedom pose result.

[0015] Further preferably, in step S1, constructing the target three-dimensional model includes: using an ORB feature extraction algorithm to detect key points from the ring-shot image and generate descriptors, matching feature points between adjacent images by descriptors, using a RANSAC method to perform geometric consistency screening, and obtaining rotation and translation parameters of the image pair by solving a basic matrix based on feature matching and combining camera intrinsic parameters;

[0016] The sparse point cloud is reconstructed using the Structure from Motion method, and then a dense point cloud is generated using the multi-view stereo matching algorithm. Finally, a triangular mesh is constructed through Delaunay triangulation, and a mapping model is generated by combining the original image texture.

[0017] Further preferably, in step S2, generating the image depth map includes: using a monocular image depth estimation network to generate a relative depth map for each input frame of the image, and performing scale calibration in combination with known structures or calibration reference points in the selected frame, back-projecting each pixel position to a three-dimensional space coordinate system through camera intrinsic parameters, completing image-to-space point cloud mapping for subsequent target position initialization.

[0018] Further preferably, in step S3, the extracting of the target image mask includes: using the image encoder of the visual base model to extract high-dimensional semantic features of the entire image, inputting the foreground points and background points obtained by user interaction into the prompt encoder to generate spatial guidance information, and jointly inputting the image encoding and the prompt encoding into the mask decoder to output a target area mask map with consistent resolution, and the mask map is directly associated with the image cropping area required for posture estimation in step S4.

[0019] Further preferably, during the posture initialization process in step S4, a set of pixels corresponding to the depth map is extracted in the target area identified by the mask map, and the target depth position is obtained by median calculation. The initial translation vector of the target in the world coordinate system is obtained by combining the back projection of the intrinsic parameters of the monocular camera. The initial estimation of the rotation matrix adopts uniform sampling of multiple observation directions on an icosphere constructed with the target center as the sphere center to form a rotation hypothesis set as a posture initialization candidate.

[0020] Further preferably, each set of initial posture parameters in the initial hypothesis set described in step S4 is applied to the target three-dimensional model to perform image rendering, generate a corresponding simulated image, and then perform texture comparison with the masked cropped area in the real observation image, construct a posture loss function based on image error, and perform gradient backpropagation optimization on the rotation and translation parameters through implicit neural network representation to obtain a high-confidence posture estimation set after convergence.

[0021] Further preferably, in the posture screening process in step S4, target image rendering is performed on each group of posture results output by the posture correction, and its image semantic coding features are extracted and compared one by one with the coding features of the mask area of ​​the input image. The posture set is sorted by the indicators of contour overlap, regional alignment and texture matching, and finally the posture with the highest score is output as the current frame estimation result.

[0022] The present invention also provides an application system for a posture recognition algorithm for any object under a monocular camera, the system comprising:

[0023] A three-dimensional model building module, used to perform the target three-dimensional modeling operation in step S1;

[0024] A depth map generation module, configured to execute the image depth map construction operation in step S2;

[0025] An image mask generation module, configured to perform the target area mask extraction operation in step S3;

[0026] A posture estimation module, comprising an initialization submodule, a correction submodule, and a screening submodule, for executing the posture estimation process in step S4;

[0027] Among them, each module in the system corresponds to the algorithm step, and the control program run by the processor calls the modules for collaborative execution to complete the posture recognition task of any target.

[0028] Further preferably, the system completes the execution result caching of the three-dimensional model construction module in the initialization phase, the depth map generation module and the image mask generation module support dynamic processing of real-time input image sequences, and the posture estimation module performs multiple sets of parameter optimization on the initial posture hypothesis set constructed in step S4 based on a parallel computing architecture, and outputs a high-confidence posture estimation result.

[0029] The embodiment of the present invention adopts the above technical solution, which has the following advantages:

[0030] 1. The present invention uses a multi-view geometry-based method to perform circumferential modeling of the target, generate a three-dimensional model of the target, and implement pose estimation by combining image masks and depth maps. Unlike traditional deep learning methods, this method does not rely on large-scale training data for target objects or specific scenes, nor does it require model fine-tuning or transfer learning. It can be applied to any category of target objects without prior knowledge, significantly improving the generalization ability and applicability of pose estimation algorithms in actual scenarios, overcoming the limitation of traditional methods that can only recognize trained objects.

[0031] 2. Compared with traditional methods that require deploying multiple cameras or using complex sensor equipment such as structured light and depth cameras to collect multi-view motion data, the technical solution provided by the present invention only requires using a single ordinary monocular camera to collect image sequences of the target during movement to complete the posture estimation process. This method significantly reduces the difficulty of system deployment and data acquisition costs, and is suitable for use in resource-constrained, rapid deployment or large-scale application scenarios.

[0032] The above summary is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments and features described above, further aspects, embodiments and features of the present invention will be readily apparent by reference to the accompanying drawings and the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0034] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0035] Hereinafter, only certain exemplary embodiments are briefly described. As will be appreciated by those skilled in the art, the described embodiments may be modified in various ways without departing from the spirit or scope of the present invention. Therefore, the drawings and description are to be considered as illustrative in nature and not restrictive.

[0036] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0037] like Figure 1 As shown, an embodiment of the present invention provides a posture recognition algorithm for any object under a monocular camera, and the algorithm includes the following steps:

[0038] Step S1: Constructing the target 3D model

[0039] This step is to collect circular images of the target in a static state, and generate a complete three-dimensional surface model of the target through image processing and reconstruction. Specifically,

[0040] Image acquisition: A fixed-focal-length monocular camera is used to capture images from multiple angles around the target. The angle distribution is evenly distributed across 360°, and the viewing angle covers no less than 20 frames. All images maintain consistent exposure and resolution during acquisition. The ORB (Oriented FAST and Rotated BRIEF) algorithm is used to extract corner points and descriptor information for each image. The ORB algorithm is both scale- and rotation-invariant and has fast computational speed, making it suitable for high-frequency industrial recognition scenarios. Feature points between adjacent images are roughly matched using descriptor distances, and the RANSAC algorithm is introduced for geometric consistency filtering to eliminate mismatched point pairs and improve registration accuracy.

[0041] By matching the feature point pairs and the camera intrinsic parameter matrix, the fundamental matrix (Fundamental Matrix) between the image pairs is calculated, and then the essential matrix (Essential Matrix) is solved. The camera extrinsic parameters (rotation matrix and translation vector) are obtained by SVD decomposition, realizing a unified coordinate representation of multi-view images. Based on the acquired poses and feature points, all images are uniformly input into the Structure from Motion (SfM) process to reconstruct a sparse point cloud. Subsequently, the Multi-view Stereo (MVS) technology is used to expand it into dense point cloud data. The point cloud surface is reconstructed using Poisson or Delaunay triangulation to obtain a continuous target triangular mesh model.

[0042] The RGB color data of the corresponding mesh area is extracted from the original image and mapped to the triangular surface to form a texture map, so that the model has both geometric integrity and retains the target texture information.

[0043] Output: The constructed target 3D model contains geometric structure and color texture, which is used as a 3D reference in the subsequent posture recognition process.

[0044] Step S2: Generate image depth map

[0045] This step is performed during the target's motion, using the monocular image sequence to restore the depth information of each frame in the scene for posture initialization and positioning. Specifically,

[0046] A monocular camera is used to capture image sequences of the target during continuous motion, with a frame rate of ≥10fps to ensure smooth posture changes and good data continuity. The image frames are input to the depth estimation model, such as a depth estimation network built based on a U-Net structure or a Transformer structure, and a relative depth map is output. Before generating the depth map, a mask or image classification module is used to identify the target object area from the input image, and the prediction range is limited during the depth estimation process to reduce the impact of background interference on depth accuracy. Each pixel in the depth prediction result will be combined with the target area index and mapped to a unified three-dimensional coordinate system through the back projection algorithm and the camera intrinsic parameters to generate a dense point cloud structure.

[0047] Since monocular estimation only has relative depth information, the mapping from relative depth to actual scale is achieved by identifying known physical reference objects in the image or using the model size in the 3D modeling stage. The depth value of each pixel in the depth map is combined with the camera intrinsic parameters for back-projection to obtain the corresponding dense point cloud in the frame image, that is, to establish the mapping relationship between the image pixel plane and the 3D coordinates.

[0048] Output: Depth map (Z channel) and dense point cloud structure corresponding to each frame image.

[0049] Step S3: Extract target image mask

[0050] This step is used to segment the target image area before pose estimation, which serves as an important basis for cropping area and image alignment calculation. Specifically,

[0051] The original image is input into an image encoder (optionally a SAM, DINO, or Mask2Former model) to obtain a global semantic representation. Through user interaction or rule recognition, the coordinates of several pixels inside and outside the target area are selected as foreground and background points. The prompt encoder converts these into guidance vectors for mask inference. The user manually selects foreground and background points in the target area image, and the system records the spatial coordinates of the selected pixels to form a list of selected coordinates. This list is used to drive the prompt encoder to generate local guidance information, forming a hint vector. The image encoding result and the hint encoding result are input into the mask decoder. After multi-scale fusion, the mask map of the target area is output. The mask map is a binary image with the same resolution as the input image, where the foreground target area is 1 and the rest of the area is 0.

[0052] Output: Object mask image that can be used for image cropping and pose optimization.

[0053] Step S4: Perform pose estimation

[0054] This step is to use the target 3D model, depth map and mask map to complete the pose estimation of the target in each frame of the image sequence. Specifically,

[0055] Extract all valid pixels in the depth map within the masked area, calculate the median to obtain the target depth center, and then back-project the pixels to world coordinates based on the camera's intrinsic parameters to obtain the initial translation vector. Construct an icosphere at the target center and uniformly sample several pose directions on its surface. Each direction and the camera's viewpoint form a set of rotation matrices as the initial rotation hypothesis set. Use the mask to crop the original image to generate a cropped image, which is paired with the simulated rendering for pose error calculation. The cropped image can be considered the target's valid region in the observed image and is used to calculate the degree of pixel alignment.

[0056] Each set of initial poses is input into the 3D model, and perspective rendering is performed to obtain a simulated image (RGB rendering). The simulated image is texture-compared with the cropped image, and a loss function is constructed based on the contour overlap and texture error between the images. An implicit neural network or a deep pose regression network is used for gradient backpropagation, and the rotation and translation parameters are iteratively updated until the loss converges.

[0057] The encoder extracts image features from each set of optimized pose rendering images and compares them with the mask region features of the real image. All pose candidates are scored based on multiple indicators such as contour overlap, region alignment rate, texture residual, etc., and the scores are weighted and summarized. The pose result with the highest score is output as the final estimated value of this frame image and stored in the pose sequence.

[0058] In addition, the present invention also provides an application system for a posture recognition algorithm for arbitrary objects using a monocular camera. The system includes the following modules, which are interconnected via a control bus or a logic call interface and work with a processor and memory to implement the algorithm's operating process:

[0059] The 3D model construction module is used to implement the target 3D modeling operations in step S1, including target image acquisition, ORB feature extraction, feature point matching, RANSAC geometric screening, basic matrix and essential matrix calculation, Structure from Motion point cloud recovery, multi-view stereo matching (MVS) reconstruction of dense point cloud, and Delaunay triangulation to achieve mesh reconstruction and texture map generation.

[0060] Preferably, the three-dimensional model construction module is pre-executed during the system initialization phase, and the output target three-dimensional model is cached in the memory for subsequent recognition process calls, thereby avoiding repeated calculation and modeling steps in each recognition task and improving system response efficiency.

[0061] The depth map generation module, corresponding to algorithm step S2, is used to generate a depth map and corresponding point cloud for each frame of the image sequence during the target motion process. The depth map generation module integrates a monocular depth estimation network internally, generates a relative depth map by inputting a sequence of images, and performs a back-projection operation based on the camera's intrinsic parameters to map the pixels to a three-dimensional coordinate system and construct a dense point cloud structure. In actual operation, this module can process continuous input images in real time to adapt to the continuous changes in target motion.

[0062] The image mask generation module, corresponding to algorithm step S3, is used to extract the foreground mask area of ​​the target in the image as the basis for image cropping in subsequent pose estimation. The image mask generation module integrates the image encoder, prompt encoder and mask decoder components. Through user selection or rule recognition, it collects the positions of foreground and background points in the image and generates a prompt vector. The global semantic information output by the image encoder is input into the mask decoder together with the prompt vector, and the target mask map is generated through fusion reasoning.

[0063] The posture estimation module, corresponding to step S4 of the algorithm, executes the core process of posture estimation. The posture estimation module includes the following submodules:

[0064] The initialization submodule is used to calculate the translation vector of the target based on the depth map and the mask area, and to construct a uniformly distributed initial rotation hypothesis set based on the icosphere;

[0065] The correction submodule applies the initial pose to the target 3D model for image rendering, performs texture comparison between the simulated image and the cropped image, constructs the pose residual loss, and optimizes the pose parameters through an implicit neural network.

[0066] Screening submodule: sorts all pose candidates based on image semantic feature similarity, contour overlap, texture alignment rate and other indicators, and selects the optimal pose as output.

[0067] In this embodiment, the posture estimation module adopts a parallel computing architecture during system operation, which can simultaneously optimize multiple initial posture hypotheses in parallel, improve the posture calculation speed and stability, and is particularly suitable for posture update tasks in high frame rate real-time scenarios.

[0068] The system is centrally scheduled by a control program run by the processor, calling each module in sequence according to steps S1 to S4 to collaboratively complete the full-process recognition task from input image to posture result. All modules transmit three-dimensional models, depth maps, mask maps and intermediate rendered images through a unified data interface to ensure data consistency and process traceability.

[0069] In this embodiment, the English abbreviations of the following technical terms are explained as follows:

[0070] ORB (Oriented FAST and Rotated BRIEF): is an image feature extraction algorithm that combines FAST corner detection and BRIEF descriptor. It is rotation and scale invariant and suitable for quickly extracting key points and their description information in images for matching between multi-view images.

[0071] RANSAC (Random Sample Consensus): a random sampling consensus algorithm used to robustly estimate model parameters in a sample set with errors or outliers. In this invention, it is used for geometric consistency screening in the image feature matching process to eliminate incorrect matching point pairs.

[0072] SfM (Structure from Motion): It is a method that gradually recovers the three-dimensional structure of the scene and the camera pose through the feature points and camera perspective information extracted from the image sequence, and is used to construct a sparse point cloud.

[0073] MVS (Multi-view Stereo): Multi-view stereo matching technology, based on sparse point clouds, further generates dense point clouds through stereo matching of multi-angle images, achieving more detailed modeling of three-dimensional structures.

[0074] U-Net: It is a typical convolutional neural network with an encoder-decoder structure, which is widely used in tasks such as image segmentation and depth estimation. In this invention, U-Net can be used as the basis for building a monocular depth estimation network.

[0075] RGB: Represents the three-channel color composition of an image, corresponding to the red (Red), green (Green), and blue (Blue) color channels. It is the basic image data format for image rendering and comparison.

[0076] Crop image: refers to the image sub-region obtained by cropping the target region mask from the original image. The image retains the pixel information of the target region and is used for posture comparison with the rendered image.

[0077] Icosphere (Icosahedron): A three-dimensional grid structure that is approximately spherical and consists of evenly distributed triangles. It is used to sample the rotation direction at equal intervals during the posture initialization phase.

[0078] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various modifications and substitutions within the technical scope disclosed in the present invention, and such modifications and substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A posture recognition algorithm for arbitrary objects using a monocular camera, characterized by: The algorithm comprises the following steps: Step S1, constructing a 3D model of the target: capturing circular images of the target and constructing a dense mesh model through feature extraction, image registration, camera pose calculation, and point cloud reconstruction; Step S2, generating an image depth map: performing depth estimation on the image sequence during target motion, and completing scale transformation and three-dimensional mapping; Step S3, extracting the target image mask: interactively clicking in the target image, and generating a target area mask map through image coding and prompt information; Step S4, perform pose estimation: combine the target 3D model, image depth map and target area mask map to construct an initial pose hypothesis set, perform pose initialization, pose correction and pose screening, and output the target's six-degree-of-freedom pose result.

2. The posture recognition algorithm for arbitrary objects using a monocular camera according to claim 1, characterized in that: In step S1, constructing the target three-dimensional model includes: using the ORB feature extraction algorithm to detect key points from the ring-shot image and generate descriptors, matching the feature points between adjacent images through descriptors, using the RANSAC method to perform geometric consistency screening, and based on the feature matching, combining the camera intrinsic parameters to obtain the rotation and translation parameters of the image pair by solving the basic matrix; The sparse point cloud is reconstructed using the Structure from Motion method, and then a dense point cloud is generated using the multi-view stereo matching algorithm. Finally, a triangular mesh is constructed through Delaunay triangulation, and a mapping model is generated by combining the original image texture.

3. The posture recognition algorithm for arbitrary objects using a monocular camera according to claim 1, characterized in that: In step S2, generating the image depth map includes: using a monocular image depth estimation network to generate a relative depth map for each input frame of the image, and performing scale calibration in combination with known structures or calibration reference points in the selected frame, back-projecting each pixel position into a three-dimensional space coordinate system through camera intrinsic parameters, completing image-to-space point cloud mapping for subsequent target position initialization.

4. The posture recognition algorithm for arbitrary objects using a monocular camera according to claim 1, characterized in that: In step S3, the extraction of the target image mask includes: using the image encoder of the visual base model to extract high-dimensional semantic features of the entire image, inputting the foreground points and background points obtained by user interaction into the prompt encoder to generate spatial guidance information, and jointly inputting the image encoding and the prompt encoding into the mask decoder to output a target area mask map with consistent resolution. The mask map is directly associated with the image cropping area required for posture estimation in step S4.

5. The posture recognition algorithm for arbitrary objects using a monocular camera according to claim 1, characterized in that: During the posture initialization process in step S4, a set of pixels corresponding to the depth map is extracted within the target area identified by the mask map, and the target depth position is obtained by median calculation. The initial translation vector of the target in the world coordinate system is obtained by combining the back projection of the intrinsic parameters of the monocular camera. The initial estimation of the rotation matrix adopts uniform sampling of multiple observation directions on the icosphere constructed with the target center as the sphere center to form a rotation hypothesis set as a posture initialization candidate.

6. The posture recognition algorithm for arbitrary objects using a monocular camera according to claim 5, characterized in that: Each set of initial pose parameters in the initial hypothesis set described in step S4 is applied to the target 3D model to perform image rendering to generate the corresponding simulated image, which is then texture-compared with the masked cropped area in the real observed image. A pose loss function based on image error is constructed, and gradient backpropagation optimization is performed on the rotation and translation parameters through implicit neural network representation to obtain a high-confidence pose estimation set after convergence.

7. The posture recognition algorithm for arbitrary objects using a monocular camera according to claim 6, characterized in that: During the posture screening process in step S4, target image rendering is performed on each group of posture results output by the posture correction, and its image semantic coding features are extracted and compared one by one with the coding features of the mask area of ​​the input image. The posture set is sorted by the indicators of contour overlap, regional alignment and texture matching, and finally the posture with the highest score is output as the current frame estimation result.

8. The application system of the posture recognition algorithm for arbitrary objects using a monocular camera according to any one of claims 1 to 7, characterized in that: The application system includes: A three-dimensional model building module, used to perform the target three-dimensional modeling operation in step S1; A depth map generation module, configured to execute the image depth map construction operation in step S2; An image mask generation module, configured to perform the target area mask extraction operation in step S3; A posture estimation module, comprising an initialization submodule, a correction submodule, and a screening submodule, for executing the posture estimation process in step S4; Among them, each module in the application system corresponds to the algorithm steps, and the control program run by the processor calls the above modules for collaborative execution to complete the posture recognition task of any target.

9. The application system of the posture recognition algorithm for arbitrary objects using a monocular camera according to claim 8, characterized in that: The application system completes the execution result caching of the three-dimensional model construction module during the initialization phase. The depth map generation module and the image mask generation module support dynamic processing of real-time input image sequences. The posture estimation module optimizes multiple sets of parameters for the initial posture hypothesis set constructed in step S4 based on a parallel computing architecture and outputs a high-confidence posture estimation result.

Images