Pose estimation method based on transformer online target tracking

Abstract

The application discloses a pose estimation method based on online target tracking of a Transformer, and belongs to the technical field of computer vision, and comprises the following steps: acquiring a continuous frame image sequence of a non-cooperative target, and constructing a query feature vector; extracting a corresponding feature vector of each frame image through a Transformer encoder, introducing spatial position coding information for position coding, and obtaining enhanced multi-scale features; taking the query feature vector as a query input, taking the multi-scale features as key and value inputs, utilizing a deformable attention mechanism of a Transformer decoder to perform feature interaction and fusion, and acquiring a decoding feature vector; based on the decoding feature vector, predicting coordinates and categories of key points of the non-cooperative target in an image; and according to the key point coordinates and a pre-defined three-dimensional key point model, calculating the position and pose of the non-cooperative target relative to a camera. The application can realize high-precision prediction of the pose of the non-cooperative target.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a pose estimation method for online target tracking based on Transformer. Background Technology

[0002] Space situational awareness is fundamental to space activities, enabling the normal operation of on-orbit services through perception, control, and decision-making capabilities. A major challenge in on-orbit services lies in accurately acquiring the position and orientation information of space targets relative to the spacecraft being maintained. This is particularly crucial when dealing with targets that cannot autonomously provide position and attitude information, such as malfunctioning spacecraft or disintegrating space debris.

[0003] In space, the vast majority of targets are non-cooperative. These targets lack cooperative markers, active devices, and may even use low-observable methods to evade observation, making attitude measurement quite difficult.

[0004] Traditional pose estimation methods typically construct a feature description of a specific target object based on hand-designed features, extracting geometric features from the target surface. A correspondence is established between 2D image features and 3D model features, and then the 6D pose of the target is solved using perspective geometry principles. However, methods based on hand-designed features are limited by their inherent characteristics and difficult to extend. Furthermore, they often face the risk of failure in extreme environments (such as space scenarios with limited lighting).

[0005] With the emergence of deep learning, it has become the mainstream method for spatial target pose measurement, mainly including two categories: direct methods and indirect methods. Direct methods do not require establishing a correspondence between 2D pixel planes and 3D world coordinates. They extract features through deep neural networks, directly obtaining the 6D pose of non-cooperative targets in an end-to-end manner. However, direct methods often have low accuracy in regressing pose parameters. Rotation information is represented using Euler angles, quaternions, or rotation matrices, but singularities or non-commutativity exist, making it difficult to optimize the loss function and for the model to converge to the optimal solution. Indirect methods first regress the absolute coordinates or heatmaps of keypoints of the non-cooperative target, and then use perspective geometry principles to solve for the pose. The pose accuracy depends on the accuracy of the keypoint coordinates in keypoint regression; deviations in keypoint localization affect subsequent PnP (Positioning and Positioning) calculations. When one or more predefined keypoints are occluded, keypoint regression will fail, severely impacting the pose estimation of non-cooperative targets. Summary of the Invention

[0006] To address some or all of the technical problems existing in the prior art, this invention provides a pose estimation method for online target tracking based on Transformer.

[0007] The technical solution of the present invention is as follows:

[0008] A pose estimation method based on Transformer-based online target tracking is provided, including:

[0009] Obtain a continuous frame image sequence of a non-cooperative target and construct a query feature vector;

[0010] The Transformer encoder extracts the corresponding feature vectors for each frame of the image and introduces spatial location coding information for location coding to obtain enhanced multi-scale features.

[0011] The query feature vector of the current frame is used as the query input of the Transformer decoder, and the enhanced multi-scale features of the current frame image are used as key and value inputs. The Transformer decoder uses a deformable attention mechanism to perform feature interaction and fusion to obtain a decoded feature vector containing spatial semantic information.

[0012] Based on the decoded feature vector, the coordinates and category of the key points of the non-cooperative target in the current frame image are predicted by a fully connected network;

[0013] Based on the key point coordinates in the current frame image and their corresponding predefined 3D key point models, the position and pose of the non-cooperative target relative to the camera in the current frame are calculated using the n-point perspective algorithm and the random sampling consistency algorithm.

[0014] The query feature vector is updated using the key point coordinates in the current frame image, and the updated query feature vector is used as the query feature vector for the next frame. Based on the query feature vector of the next frame, the position and attitude of the non-cooperative target relative to the camera in the next frame are calculated.

[0015] Furthermore, in some embodiments, the step of acquiring a continuous frame image sequence of a non-cooperative target and constructing a query feature vector includes:

[0016] Acquire a sequence of consecutive frame images of a non-cooperative target;

[0017] Each frame in a continuous frame image sequence is input into a keypoint regression network to obtain the query feature vector;

[0018] The consecutive frame image sequence is represented as follows:

[0019] ;

[0020] The query feature vector is represented as follows:

[0021] ;

[0022] in, Represents a sequence of consecutive frames of images. This represents the first frame of the image. This represents the second frame image. Indicates the first Frame image, Represents the real number field. Indicates the number of time frames. Indicates the image height. This indicates the image width, and 3 represents the number of color channels. The dimension is The four-dimensional real tensor, This represents the query feature vector. This indicates the first query point, derived from the time information of the first frame image. and features of the first frame image composition, This indicates the second query point, derived from the time information of the second frame image. Second frame image features composition, Indicates the first The query point is the first one. Frame image time information and the Frame Image Features composition.

[0023] Furthermore, in some embodiments, the step of extracting corresponding feature vectors for each frame of image using a Transformer encoder and introducing spatial location coding information for location coding to obtain enhanced multi-scale features includes:

[0024] Convolutional feature extraction is performed on each frame of the image to obtain feature vectors at different levels corresponding to each frame of the image;

[0025] Based on feature vectors of different levels, the highest-level feature vector is flattened and positional encoding is performed using spatial positional encoding information to obtain preliminary multi-scale features;

[0026] By using feature pyramids to perform bidirectional fusion of preliminary multi-scale features from top to bottom and bottom to top, enhanced multi-scale features are obtained.

[0027] Furthermore, in some implementations, position encoding is performed in the following manner:

[0028] ;

[0029] in, Indicates the first The highest-level feature vector corresponding to the frame image Represents spatial location encoding information, Indicates the first The single-scale feature is obtained by position encoding the highest-level feature vector corresponding to the frame image.

[0030] Furthermore, in some embodiments, the deformable attention mechanism is represented as:

[0031] ;

[0032] in, This represents a deformable attention mechanism function. This represents the query feature vector. Indicates the first element in the query feature vector. Normalized coordinates of the reference point corresponding to each element This represents the multi-scale features of the input. Representing the first in multi-scale features Layer features, The number of feature layers representing multi-scale features. Indicates the number of attention heads. Indicates the output projection matrix. Indicates the number of sampling points. Indicates the first The first layer of features The sampling point of the first sampling point Attention weights for each attention head Represents the value projection matrix, Represents the coordinate mapping function. Indicates the first The first layer of features The sampling point of the first sampling point The sampling offset of each attention head.

[0033] Furthermore, in some embodiments, the Transformer decoder employs a multi-layer iterative decoding structure.

[0034] Furthermore, in some embodiments, the fully connected network includes a classification module and a regression module, wherein the classification module predicts the category of key points through a linear projection layer, and the regression model predicts the coordinates of key points through a multilayer perceptron.

[0035] Furthermore, in some implementations, the step of calculating the position and pose of the non-cooperative target relative to the camera in the current frame, based on the keypoint coordinates in the current frame image and their corresponding predefined 3D keypoint models, using an n-point perspective algorithm and a random sampling consistency algorithm, includes:

[0036] Construct 2D-3D corresponding point pairs using the key point coordinates in the current frame image and a predefined 3D key point model;

[0037] The initial pose is calculated using an n-point perspective algorithm based on the corresponding 2D-3D point pairs.

[0038] The random sampling consensus algorithm is used to remove abnormal matches in the corresponding point pairs, and the initial pose is iteratively optimized to obtain the position and pose of the non-cooperative target relative to the camera in the current frame.

[0039] The main advantages of the technical solution of this invention are as follows:

[0040] The pose estimation method for online target tracking based on Transformer of this invention acquires a continuous frame image sequence of a non-cooperative target, and performs feature extraction and feature association analysis on the frame images of the non-cooperative target based on temporal information and Transformer architecture. Iterative processing is performed by combining query feature vectors and multi-scale features of the image to obtain decoded feature vectors. Based on the decoded feature vectors, key points of the non-cooperative target in the image are predicted through a fully connected network. Based on the coordinates of the key points, the pose information of the non-cooperative target relative to the camera is calculated and determined using the n-point perspective algorithm and the random sampling consensus algorithm. The regression process is further refined, which significantly improves the localization accuracy of key points, thereby achieving high-precision prediction of the pose of non-cooperative targets. This solves the problem of inaccurate key point regression localization when predefined key points are occluded under complex environmental conditions such as drastic changes in lighting and partial occlusion in traditional technologies, and improves the accuracy of pose estimation for non-cooperative targets. Attached Figure Description

[0041] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and constitute a part of this invention, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0042] Figure 1 This is a flowchart illustrating a pose estimation method for online target tracking based on Transformer, provided in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0044] The technical solutions provided by the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0045] refer to Figure 1This invention provides a pose estimation method for online target tracking based on Transformer, which includes the following steps:

[0046] Step 1: Obtain a continuous frame image sequence of non-cooperative targets and construct a query feature vector;

[0047] Step 2: Extract the corresponding feature vectors for each frame of image using the Transformer encoder, and introduce spatial location coding information for location coding to obtain enhanced multi-scale features;

[0048] Step 3: Use the query feature vector of the current frame as the query input of the Transformer decoder, and use the enhanced multi-scale features of the current frame image as the key and value input. The Transformer decoder uses the deformable attention mechanism to perform feature interaction and fusion to obtain the decoded feature vector containing spatial semantic information.

[0049] Step 4: Based on the decoded feature vector, predict the coordinates and category of key points of non-cooperative targets in the current frame image using a fully connected network;

[0050] Step 5: Based on the key point coordinates in the current frame image and their corresponding predefined 3D key point models, use the n-point perspective algorithm (PnP algorithm) and the random sampling consensus algorithm (RANSAC algorithm) to calculate the position and pose of the non-cooperative target relative to the camera in the current frame.

[0051] Step 6: Update the query feature vector using the key point coordinates in the current frame image, and use the updated query feature vector as the query feature vector for the next frame. Based on the query feature vector of the next frame, calculate the position and pose of the non-cooperative target relative to the camera in the next frame.

[0052] In this embodiment of the invention, the query feature vector is updated using the key point coordinates in the current frame image, and the updated query feature vector is used as the query feature vector for the next frame. Based on the query feature vector of the next frame, steps 3 to 5 are repeated to obtain the position and orientation of the non-cooperative target relative to the camera in the next frame. By repeating the above process, the position and orientation of the non-cooperative target relative to the camera in each frame are obtained.

[0053] In this embodiment of the invention, the predefined three-dimensional key point model refers to a set of physical feature points with known three-dimensional coordinates on the surface of a non-cooperative target, which is used to establish a geometric correspondence with two-dimensional key points in the image, thereby solving the pose.

[0054] In this embodiment of the invention, when updating the query feature vector using the coordinates of key points in the current frame image, the key points in the current frame image are used to replace the query points in the corresponding frame image, thereby updating the query feature vector.

[0055] In this embodiment of the invention, the Transformer encoder and Transformer decoder can adopt existing structures.

[0056] In this embodiment of the invention, key points of non-cooperative targets are located and tracked based on temporal information and the Transformer architecture. The spatial position of key points is gradually optimized through iterative methods. By focusing on the local image features around the key points and their own historical state features, the reliability of temporal sequence feature matching and the consistency of behavior can be guaranteed.

[0057] The pose estimation method for online target tracking based on Transformer provided in this invention acquires a continuous frame image sequence of a non-cooperative target. Based on temporal information and the Transformer architecture, it performs feature extraction and feature association analysis on the frame images of the non-cooperative target. Iterative processing is then performed using query feature vectors and multi-scale features of the image to obtain decoded feature vectors. Based on these decoded feature vectors, a fully connected network predicts the key points of the non-cooperative target in the image. Based on the key point coordinates, an n-point perspective algorithm and a random sampling consensus algorithm are used to calculate and determine the pose information of the non-cooperative target relative to the camera. Compared with traditional techniques, the pose estimation method for online target tracking based on Transformer provided in this invention comprehensively utilizes temporal information, focuses on the contextual and local features of key points, and iteratively optimizes the spatial position of key points, further refining the regression process and significantly improving the localization accuracy of key points. This achieves high-precision prediction of the pose of non-cooperative targets, solving the problem of inaccurate key point regression localization in traditional techniques when predefined key points are occluded under complex environmental conditions such as drastic changes in lighting and partial occlusion, thus improving the accuracy of pose estimation for non-cooperative targets.

[0058] Furthermore, in this embodiment of the invention, in step 1, obtaining a continuous frame image sequence of a non-cooperative target and constructing a query feature vector includes:

[0059] Acquire a sequence of consecutive frame images of a non-cooperative target;

[0060] Each frame in a continuous frame image sequence is input into a keypoint regression network to obtain the query feature vector;

[0061] The sequence of consecutive frame images is represented as follows:

[0062] ;

[0063] The query feature vector is represented as:

[0064] ;

[0065] in, Represents a sequence of consecutive frames of images. This represents the first frame of the image. This represents the second frame image. Indicates the first Frame image, Represents the real number field. Indicates the number of time frames. Indicates the image height. This indicates the image width, and 3 represents the number of color channels. The dimension is The four-dimensional real tensor, This represents the query feature vector. This indicates the first query point, derived from the time information of the first frame image. and features of the first frame image composition, This indicates the second query point, derived from the time information of the second frame image. Second frame image features composition, Indicates the first The query point is the first one. Frame image time information and the Frame Image Features composition.

[0066] In this embodiment of the invention, the keypoint regression network is a neural network used to predict the location of predefined keypoints (or their feature representations) of a target object from a single frame image. Its output (such as image features) is combined with a timestamp to form a query feature vector for the Transformer decoder to track and match.

[0067] In this embodiment of the invention, by constructing query points containing temporal and feature information for each frame of image, a structured input rich in historical context is provided to the Transformer, thereby significantly improving the continuity and robustness of target tracking in complex scenes (such as occlusion and deformation).

[0068] Furthermore, in this embodiment of the invention, in step 2, the Transformer encoder extracts the corresponding feature vector for each frame of image and introduces spatial location coding information for location coding to obtain enhanced multi-scale features, including:

[0069] Convolutional feature extraction is performed on each frame of the image to obtain feature vectors at different levels corresponding to each frame of the image;

[0070] Based on feature vectors of different levels, the highest-level feature vector is flattened and positional encoding is performed using spatial positional encoding information to obtain preliminary multi-scale features;

[0071] By using feature pyramids to perform bidirectional fusion of preliminary multi-scale features from top to bottom and bottom to top, enhanced multi-scale features are obtained.

[0072] In this embodiment of the invention, position encoding is performed in the following manner:

[0073] ;

[0074] in, Indicates the first The highest-level feature vector corresponding to the frame image Represents spatial location encoding information, Indicates the first The single-scale feature is obtained by position encoding the highest-level feature vector corresponding to the frame image.

[0075] In this embodiment of the invention, the initial multi-scale features are composed of feature vectors at different levels, and the highest-level feature vector is a position-encoded feature vector.

[0076] In this embodiment of the invention, by combining spatial location coding and multi-scale feature bidirectional fusion, the feature map's ability to represent the spatial structure and multi-scale appearance of the target is significantly enhanced, providing a more robust and richer visual foundation for subsequent decoding and tracking.

[0077] Furthermore, in this embodiment of the invention, the deformable attention mechanism is represented as follows:

[0078] ;

[0079] in, This represents a deformable attention mechanism function. This represents the query feature vector. Indicates the first element in the query feature vector. Normalized coordinates of the reference point corresponding to each element This represents the multi-scale features of the input. Representing the first in multi-scale features Layer features, The number of feature layers representing multi-scale features. Indicates the number of attention heads. Indicates the output projection matrix. Indicates the number of sampling points. Indicates the first The first layer of features The sampling point of the first sampling point Attention weights for each attention head Represents the value projection matrix, Represents the coordinate mapping function. Indicates the first The first layer of features The sampling point of the first sampling point The sampling offset of each attention head.

[0080] In this embodiment of the invention, the deformable attention mechanism significantly enhances the adaptability to target deformation, occlusion and scale changes by dynamically adjusting the sampling position and fusing multi-scale context, thereby improving the accuracy and robustness of key point localization.

[0081] Furthermore, in this embodiment of the invention, the Transformer decoder employs a multi-layer iterative decoding structure.

[0082] In this embodiment of the invention, the Transformer decoder is based on a multi-layer iterative decoding structure, which achieves iterative optimization of key points from coarse to fine by layering offsets.

[0083] Specifically, the iterative optimization of key points by the Transformer decoder can be represented as:

[0084] ;

[0085] in, Indicates the first Key information output by the layer decoder Indicates the first Key information output by the layer decoder Indicates the first The offset predicted by the layer decoder.

[0086] Furthermore, in this embodiment of the invention, the fully connected network includes a classification module and a regression module. The classification module predicts the category of the key point through a linear projection layer, and the regression model predicts the coordinates of the key point through a multilayer perceptron.

[0087] Further, in this embodiment of the invention, in step 5, based on the keypoint coordinates in the current frame image and their corresponding predefined 3D keypoint models, the position and pose of the non-cooperative target relative to the camera in the current frame are calculated using an n-point perspective algorithm and a random sampling consistency algorithm, including:

[0088] Construct 2D-3D corresponding point pairs using the key point coordinates in the current frame image and a predefined 3D key point model;

[0089] The initial pose is calculated using an n-point perspective algorithm based on the corresponding 2D-3D point pairs.

[0090] The random sampling consensus algorithm is used to remove abnormal matches in the corresponding point pairs, and the initial pose is iteratively optimized to obtain the position and pose of the non-cooperative target relative to the camera in the current frame.

[0091] In this embodiment of the invention, considering that the n-point perspective algorithm and the random sampling consensus algorithm are conventional algorithms in the field, the n-point perspective algorithm and the random sampling consensus algorithm will not be described in detail here.

[0092] In this embodiment of the invention, by combining the fast geometric calculation of the n-point perspective algorithm with the robust elimination of outlier matching by RANSAC, high-precision and high-stability pose estimation of non-cooperative targets can be achieved under complex imaging conditions.

[0093] It should be noted that, in this document, relational terms such as “first” and “second” are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A pose estimation method for online target tracking based on Transformer, characterized in that, include: Obtain a continuous frame image sequence of a non-cooperative target and construct a query feature vector; The Transformer encoder extracts the corresponding feature vectors for each frame of the image and introduces spatial location coding information for location coding to obtain enhanced multi-scale features. The query feature vector of the current frame is used as the query input of the Transformer decoder, and the enhanced multi-scale features of the current frame image are used as key and value inputs. The Transformer decoder uses a deformable attention mechanism to perform feature interaction and fusion to obtain a decoded feature vector containing spatial semantic information. Based on the decoded feature vector, the coordinates and category of the key points of the non-cooperative target in the current frame image are predicted by a fully connected network; Based on the key point coordinates in the current frame image and their corresponding predefined 3D key point models, the position and pose of the non-cooperative target relative to the camera in the current frame are calculated using the n-point perspective algorithm and the random sampling consistency algorithm. The query feature vector is updated using the key point coordinates in the current frame image, and the updated query feature vector is used as the query feature vector for the next frame. Based on the query feature vector of the next frame, the position and attitude of the non-cooperative target relative to the camera in the next frame are calculated. The step of obtaining a continuous frame image sequence of a non-cooperative target and constructing a query feature vector includes: Acquire a sequence of consecutive frame images of a non-cooperative target; Each frame of a continuous frame image sequence is input into a keypoint regression network to obtain a query feature vector. The keypoint regression network is a neural network used to predict the location of predefined keypoints of a target object from a single frame image. The output of the keypoint regression network is combined with a timestamp to form the query feature vector used by the Transformer decoder for tracking and matching. The consecutive frame image sequence is represented as follows: ； The query feature vector is represented as follows: ； in, Represents a sequence of consecutive frames of images. This represents the first frame of the image. This represents the second frame image. Indicates the first Frame image, Represents the real number field. Indicates the number of time frames. Indicates the image height. This indicates the image width, and 3 represents the number of color channels. The dimension is The four-dimensional real tensor, This represents the query feature vector. This indicates the first query point, derived from the time information of the first frame image. and features of the first frame image composition, This indicates the second query point, derived from the time information of the second frame image. Second frame image features composition, Indicates the first The query point is the first one. Frame image time information and the Frame Image Features composition.

2. The pose estimation method for online target tracking based on Transformer according to claim 1, characterized in that, The process involves extracting corresponding feature vectors from each frame of the image using a Transformer encoder and incorporating spatial location coding information for location coding to obtain enhanced multi-scale features, including: Convolutional feature extraction is performed on each frame of the image to obtain feature vectors at different levels corresponding to each frame of the image; Based on feature vectors of different levels, the highest-level feature vector is flattened and positional encoding is performed using spatial positional encoding information to obtain preliminary multi-scale features; By using feature pyramids to perform bidirectional fusion of preliminary multi-scale features from top to bottom and bottom to top, enhanced multi-scale features are obtained.

3. The pose estimation method for online target tracking based on Transformer according to claim 2, characterized in that, Location encoding is performed in the following manner: ； in, Indicates the first The highest-level feature vector corresponding to the frame image Represents spatial location encoding information, Indicates the first The single-scale feature is obtained by position encoding the highest-level feature vector corresponding to the frame image.

4. The pose estimation method for online target tracking based on Transformer according to claim 1, characterized in that, The deformable attention mechanism is represented as follows: ； in, This represents a deformable attention mechanism function. This represents the query feature vector. Indicates the first element in the query feature vector. Normalized coordinates of the reference point corresponding to each element This represents the multi-scale features of the input. Representing the first in multi-scale features Layer features, The number of feature layers representing multi-scale features. Indicates the number of attention heads. Indicates the output projection matrix. Indicates the number of sampling points. Indicates the first The first layer of features The sampling point of the first sampling point Attention weights for each attention head Represents the value projection matrix, Represents the coordinate mapping function. Indicates the first The first layer of features The sampling point of the first sampling point The sampling offset of each attention head.

5. The pose estimation method for online target tracking based on Transformer according to claim 1 or 4, characterized in that, The Transformer decoder employs a multi-layer iterative decoding structure.

6. The pose estimation method for online target tracking based on Transformer according to claim 1, characterized in that, The fully connected network includes a classification module and a regression module. The classification module predicts the category of key points through a linear projection layer, and the regression model predicts the coordinates of key points through a multilayer perceptron.

7. The pose estimation method for online target tracking based on Transformer according to claim 1, characterized in that, The step of calculating the position and pose of the non-cooperative target relative to the camera in the current frame, based on the keypoint coordinates in the current frame image and their corresponding predefined 3D keypoint models, using an n-point perspective algorithm and a random sampling consistency algorithm, includes: Construct 2D-3D corresponding point pairs using the key point coordinates in the current frame image and a predefined 3D key point model; The initial pose is calculated using an n-point perspective algorithm based on the corresponding 2D-3D point pairs. The random sampling consensus algorithm is used to remove abnormal matches in the corresponding point pairs, and the initial pose is iteratively optimized to obtain the position and pose of the non-cooperative target relative to the camera in the current frame.

Images