Hierarchical compression space target key point detection method and system

Abstract

The invention relates to a hierarchical compression space target key point detection method and system, belongs to the technical field of space situation awareness, and solves the technical problem that an existing method is low in space target key point detection efficiency and cannot meet the real-time performance of a complex scene. Comprising the steps of obtaining an original visible light image including a space target and performing preprocessing to obtain a preprocessed visible light image; performing connected domain analysis extraction on the pre-processed visible light image to obtain an envelope point set of the outer contour of the space target; and screening the envelope point set step by step and performing hierarchical compression to obtain geometric key points of the space target for space target positioning and attitude estimation. And real-time accurate detection of the key points of the space target is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of space situation awareness, and in particular, to a method and system for detecting key points of space targets with hierarchical compression. Background Art

[0002] The detection of feature points in space target images is often applied to various space vision tasks. Since the visible light image of a space target is a matrix containing the brightness, shape, color, and texture information of the target, directly processing the entire image will introduce a large amount of redundant information, which not only reduces the processing speed but also easily introduces errors. As one of the contents of feature extraction, the feature point detection technology has the characteristics of fast response, simple deployment, and strong universality, so it is widely used in on-orbit tasks. The space target feature point detection technology is a prerequisite for various tasks such as pose estimation, target navigation, and target positioning, and also has important significance in the field of image processing. As an important basis for target recognition, image registration, target positioning, image stitching, and three-dimensional reconstruction, the accuracy of feature point detection directly affects the effect of subsequent operations. Therefore, how to accurately detect target feature points is the key to the autonomous recognition of space non-cooperative targets.

[0003] The existing research on the detection of space target feature points is usually divided into two types. One is the method based on artificial intelligence, that is, using a neural network to train a large amount of data to achieve the purpose of feature point detection. The neural network has strong learning ability and good generalization ability. It can learn complex and abstract feature representations through a large amount of data to adapt to various complex scene changes. However, the data requirements and resource consumption of this method are relatively large, and it is usually difficult to give full play to its advantages when considering resource limitations. The other is the feature point detection based on traditional methods, which has the advantages of strong interpretability and high computational efficiency. In the space-based situation awareness scenario, space targets usually have diversity in scale and shape. For non-cooperative targets, their specific morphological information is unknown. At the same time, space targets need to consider the changes in illumination and viewing angle, and the observed content usually cannot maintain an ideal detection state. Therefore, a series of preprocessing means are required to achieve better feature extraction. In the detection of space target feature points, factors such as sensor noise and environmental noise may have a great impact on the recognition of feature points in the image. At the same time, for the on-orbit observation scenario, the movement of space targets needs to be considered, which will cause the feature points in the image to change continuously and regularly between adjacent frames. We use the information of sequence images to assist in extracting feature points.

[0004] The space target feature point detection method first needs to have real-time performance, and secondly, the demand for resources should be as small as possible. At the same time, the morphology of non-cooperative targets is unknown, and a method with strong robustness is needed to detect targets with different three-dimensional structures. For the space scenario, it is necessary to cope with the changes brought by noise interference and illumination. Summary of the Invention

[0005] In view of the above analysis, embodiments of the present invention aim to provide a hierarchical compression method and system for detecting key points of spatial targets, so as to solve the technical problem that the existing methods are inefficient in detecting key points of spatial targets and cannot meet the real-time requirements of complex scenarios.

[0006] The purpose of the present invention is mainly achieved through the following technical solutions:

[0007] The present invention provides a hierarchical compression method for detecting key points of spatial targets, including:

[0008] Obtain the original visible light image including the spatial target and perform preprocessing to obtain the preprocessed visible light image;

[0009] Perform connected component analysis extraction on the preprocessed visible light image to obtain an envelope point set of the outer contour of the spatial target;

[0010] Perform hierarchical compression on the envelope point set through step-by-step screening to obtain geometric key points of the spatial target for spatial target positioning and attitude estimation.

[0011] Further, the step-by-step screening includes the first, second, and third rounds of elimination. Among them,

[0012] In the first round of elimination, based on all the envelope points in the envelope point set, the number of 8-neighborhood bright points of each envelope point is extracted by using an adaptive-step hole convolution neighborhood feature, and the redundant points located on the straight line or hypotenuse are determined and eliminated. The remaining envelope points after the first round of elimination are candidate key points; among them, the bright point is a point with a brightness of 1;

[0013] In the second round of elimination, multi-scale dynamic threshold screening of the 24-neighborhood is performed based on the candidate key points. When the number of 24-neighborhood bright points of the envelope point exceeds the threshold range dynamically set based on its 8-neighborhood bright points, it is determined as a burr point and deleted. The remaining envelope points after the second round of elimination are effective key points;

[0014] In the third round of elimination, based on the effective key points, by calculating the distance and confidence difference between each effective key point and its adjacent effective key points, adjacent point clustering and merging of adjacent points are performed to retain the true key points.

[0015] Further, the first round of elimination specifically includes:

[0016] Based on the envelope point set, obtain the area of the spatial target; adaptively set the neighborhood traversal step size step based on the area of the spatial target;

[0017] step = k·S target

[0018] Where k is the scaling factor, S target is the spatial target area;

[0019] Taking each envelope point in the envelope point set as the center, traverse the neighborhood in a jump manner according to the step length step to generate a binary feature descriptor corresponding to each envelope point;

[0020] Based on the binary feature descriptor, the number of bright spots N8(x, y) in the neighborhood of the corresponding envelope point 8 is counted. If N8(x, y)=5, the corresponding envelope point is determined to be a redundant point on a straight edge or a hypotenuse and is removed; the envelope points with N8(x, y)≠5 are retained as candidate key points.

[0021] Furthermore, the second round of elimination specifically includes:

[0022] Based on the binary feature descriptor, calculate the number of bright spots N in the 24 neighborhood of each envelope point in the candidate key point 24 (x,y);

[0023] Dynamically set N based on N8(x,y) of each envelope point in the candidate key points 24 The dynamic decision threshold [T min (N 24 (x,y)), T max (N 24 (x,y))], where T min (N 24 (x,y))、T max (N 24 (x, y)) are the minimum and maximum thresholds for the number of points with brightness 1 in the 24 neighborhood;

[0024] If T min (N 24 (x,y))≤N 24 (x,y)≤T max (N 24 (x,y)), the envelope point is determined to be a burr point and is removed; otherwise, the envelope point is determined to be a valid key point.

[0025] Furthermore, the third round of elimination specifically includes:

[0026] Calculate the distance and confidence of each pair of adjacent valid key points;

[0027] Set distance threshold and confidence difference threshold;

[0028] Use the DBSCAN algorithm to classify points smaller than the distance threshold into one category, and obtain multiple clusters;

[0029] Retain the point with the highest confidence in each cluster;

[0030] For the points in each cluster that are less than the distance threshold and less than the confidence difference threshold, merge them into a real key point, and obtain the real key point and its position coordinates (x, y).

[0031] Further, based on the N8(x, y) of each envelope point in the candidate key points, dynamically set the N 24 (x, y) dynamic determination threshold as follows:

[0032]

[0033] T min (N 24 (x, y)) and T max (N 24 (x, y)) are rounded integers.

[0034] Further, preprocess the original visible light image including the spatial target, including:

[0035] Remove background noise by subtracting the background mean in blocks to obtain an image with background noise removed;

[0036] Perform binary processing on the image with background noise removed to enhance the contrast between the spatial target and the background, and obtain a binary image as the preprocessed visible light image of the spatial target.

[0037] Further, perform connected component analysis extraction on the preprocessed visible light image to obtain an envelope point set of the outer contour of the spatial target, including:

[0038] Perform morphological closing operation connected component analysis extraction on the preprocessed visible light image, perform dilation first and then erosion operations, and obtain the preprocessed visible light image including the spatial target area as follows:

[0039]

[0040] Among them, I close is the contour image of the spatial target after the closing operation, I Binary is the preprocessed visible light image, S is the structural element defined based on the shape and size characteristics of the spatial target, ⊕ is the dilation operation, is the erosion operation;

[0041] For the points in I close with non-zero gray values, judge whether there are points with pixel values different from their own in the 8-neighborhood of the points with non-zero gray values; if so, the points with non-zero gray values are the envelope points of the spatial target;

[0042] All the envelope points of the space target form an envelope point set of the outer contour of the space target.

[0043] Further, every n rows of the original visible light image including the space target are taken as a block, and the original visible light image including the space target is divided into multiple blocks;

[0044] Calculate the average gray value of each block;

[0045]

[0046] where mean i is the gray value of all pixels in the i-th block, I(i,j,k) is the pixel value of the j-th row and k-th column in the i-th block, and W is the width of the original visible light image including the space target;

[0047] Subtract the average gray value of each block from the pixel value of each pixel point in each block to remove background noise.

[0048] A hierarchical compression space target key point detection system includes the following modules:

[0049] An image acquisition and preprocessing module, which is used to acquire the original visible light image of the space target and perform preprocessing to obtain the preprocessed visible light image of the space target;

[0050] A target contour envelope extraction module, which is used to perform connected domain analysis and extraction on the preprocessed visible light image of the space target to obtain an envelope point set of the outer contour of the space target;

[0051] A geometric key point hierarchical compression screening module, which is used to screen the envelope point set of the outer contour of the space target by using a hierarchical compression method to obtain the geometric key points of the space target for space target positioning and attitude estimation.

[0052] Compared with the prior art, the present invention can at least achieve one of the following beneficial effects:

[0053] 1. Through the hierarchical compression method, the present invention screens envelope points step by step, significantly reducing the calculation amount, improving the detection efficiency, and meeting the real-time requirement. The preprocessing and feature extraction steps are efficient and can complete key point detection in a short time, which is suitable for real-time update and tracking in a rapidly changing space environment and dynamic scenes;

[0054] 2. The present invention adopts adaptive stepwise dilated convolutional neighborhood feature extraction and multi-scale dynamic threshold screening, which can effectively cope with targets of different scales and shapes, enhancing the robustness of the algorithm. Through multi-scale analysis and dynamic adjustment of the judgment range, it adapts to the detection of space targets under different illumination and noise conditions and has a wide application prospect;

[0055] 3. Through a three-round elimination mechanism, the present invention gradually eliminates the points on the straight edges and oblique edges, burr points, and duplicate redundant points, and finally extracts high-precision geometric key points, improving the accuracy of feature point detection. Through confidence evaluation, it ensures that the retained geometric key points have high confidence, further improving the detection accuracy and stability;

[0056] 4. Compared with the artificial intelligence-based methods, the present application has less data requirements and resource consumption, and is more suitable for applications in resource-constrained environments, such as on-orbit observation scenarios. The preprocessing and feature extraction steps are efficient, and can complete key point detection with limited computing resources, featuring fast response, simple deployment, and strong universality, and being easy to deploy and use in practical applications;

[0057] 5. The method of the present application is applicable to a variety of space target detection scenarios, including space-based observation data, simulation data, laboratory simulation data, and public data sets, with strong universality. Through multiple rounds of screening and confidence evaluation, it ensures the stability and reliability of key point detection, reduces false detections and missed detections, and improves the reliability of detection.

[0058] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the subsequent specification, and some advantages can be made obvious from the specification, or understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained from the content specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The drawings are only for the purpose of showing specific embodiments, and are not considered as limiting the present invention. Throughout the drawings, the same reference signs represent the same components.

[0060] Figure 1 It is a flowchart of a hierarchical compression-based space target key point detection method in an embodiment of the present invention;

[0061] Figure 2 It is a schematic diagram of a visible light simulation image of a space target in an embodiment of the present invention;

[0062] Figure 3 It is a schematic diagram of target envelope point detection in an embodiment of the present invention;

[0063] Figure 4 It is a schematic diagram of an 8-connected domain example of straight line points in an embodiment of the present invention;

[0064] Figure 5 It is a schematic diagram of a skip traversal with a step of 2 in an embodiment of the present invention;

[0065] Figure 6Schematic diagram of the distribution of connected components of potential key points in the embodiment of the present invention;

[0066] Figure 7 Schematic diagram of the first-round key point detection in the embodiment of the present invention;

[0067] Figure 8 Schematic diagram of the details of the first-round detection in the embodiment of the present invention;

[0068] Figure 9 Schematic diagram of the expansion of the search range of the second connected component in the embodiment of the present invention;

[0069] Figure 10 Schematic diagram of the distribution of connected components under the influence of noise in the embodiment of the present invention;

[0070] Figure 11 Schematic diagram of the second-round key point detection in the embodiment of the present invention;

[0071] Figure 12 Schematic diagram of the details of the key point detection in the embodiment of the present invention;

[0072] Figure 13 Schematic diagram of the third-round key point detection in the embodiment of the present invention;

[0073] Figure 14 Schematic diagram of a hierarchical compression-based spatial target key point detection system in the embodiment of the present invention. Detailed implementation manners

[0074] Next, the preferred embodiments of the present invention will be specifically described with reference to the accompanying drawings. The accompanying drawings form a part of the present application and are used together with the embodiments of the present invention to explain the principle of the present invention, rather than to limit the scope of the present invention.

[0075] Embodiment 1:

[0076] In this embodiment, a hierarchical compression-based spatial target key point detection method can quickly detect key points for the original visible light image including the spatial target, and can adapt to visible light images including spatial targets with different shapes, scales, angles, and illuminations, having the advantages of high detection efficiency and strong robustness. It can provide technical support and reference for the design of key point detection and pose estimation schemes for spatial targets, and solve the technical problems of related projects.

[0077] A specific embodiment of the present invention discloses a hierarchical compression-based spatial target key point detection method, as Figure 1 shown, including the following steps:

[0078] Step S1, obtain the original visible light image including the spatial target and perform preprocessing to obtain the preprocessed visible light image;

[0079] Step S2: Perform connected component analysis and extraction on the preprocessed visible light image to obtain the set of envelope points of the outer contour of the space target.

[0080] Step S3: Gradually screen and hierarchically compress the set of envelope points to obtain the geometric key points of the space target for space target positioning and attitude estimation.

[0081] Step S1 includes steps S11 - S12.

[0082] Step S11: Obtain the original visible light image data including the space target.

[0083] The original visible light image data including the space target can be obtained through the following several methods:

[0084] (1) Space - based observation data: High - resolution visible light images including space targets are obtained through optical cameras, infrared sensors, or radars mounted on satellites. This type of data is suitable for monitoring space debris, near - Earth orbit satellites, and deep - space targets.

[0085] (2) Simulation data: Generated through physical modeling and computer simulation, combined with the actually observed visible light images including space targets for algorithm training and testing.

[0086] (3) Laboratory simulation data: Obtained through microgravity environment experiments or target model photography, used to test the algorithm's recognition ability for specific space targets.

[0087] (4) Public data sets: Exemplarily, such as the space visible light image databases released by institutions such as the International Space Station, NASA (National Aeronautics and Space Administration), and ESA (European Space Agency). These data provide standardized and high - quality visible light image samples.

[0088] Exemplarily, this embodiment is based on the visible light image of a single - wing L - type space target, and uses the key - point detection method based on hierarchical compression to Figure 2 perform key - point detection on the shown visible light image including the space target. This visible light image is a simulation image with a size of 2048 * 2048. The image contains two parts: the common background of space - based visible light observation and the space target. Among them, the background includes stars and noise, and the target selects an L - type imitating the "Starlink" satellite.

[0089] Step S12: Preprocess the obtained original visible light image including the space target to obtain the preprocessed visible light image.

[0090] The original visible light image of a space target is usually a grayscale image, lacking RGB information. At the same time, it is affected by factors such as illumination shadows and noise, resulting in a large amount of redundant information in the image, while the number of effective geometric key points of the space target is usually small. Traditional key point detection algorithms such as Harris or SIFT may be affected by noise and redundant points, resulting in incorrect detection results or redundant detections, and cannot effectively support pose solution. To solve this problem, this application preprocesses the original visible light image including the space target to suppress noise, separate the background, and extract the target area.

[0091] Preprocess the original visible light image including the space target, including:

[0092] Remove background noise by subtracting the background mean in blocks to obtain an image with background noise removed;

[0093] Perform binarization processing on the image with background noise removed to enhance the contrast between the space target and the background, and obtain a binarized image as the preprocessed visible light image of the space target.

[0094] The actually observed visible light image including the space target is inevitably affected by noise, and the noise introduced by the detector is the main source of noise. To reduce the impact of noise on subsequent image processing, first perform denoising processing on the original visible light image including the space target. Considering the timeliness and feasibility of the algorithm for on-orbit scenarios, the method of subtracting the background is used to remove noise.

[0095] Preprocess the original visible light image including the space target by subtracting the background mean in blocks and performing binarization processing to enhance the contrast between the target and the background, specifically as follows:

[0096] First, use the method of subtracting the background mean in blocks. By dividing the original visible light image including the space target into blocks, calculating the background mean of each block area and subtracting it, the interference of large-range background noise can be effectively removed, thereby highlighting the characteristic information of the space target area.

[0097] Considering the possibility that the visible light image taken by a LEO (Low Earth Orbit) satellite is affected by earthshine, apply the method of background adaptive removal to reduce the impact of the background on target extraction, and remove earthshine by subtracting the mean in blocks. Calculate the average gray value of every n rows of the image, and subtract this mean from each row of pixel values to remove background noise.

[0098] Take every n rows of the original visible light image including the space target as a block, and the original visible light image including the space target is divided into multiple blocks;

[0099] Calculate the average gray value of each block;

[0100]

[0101] where mean i is the gray value of all pixels within the i-th block, I(i, j, k) is the pixel value at the j-th row and k-th column in the i-th block, and W is the width of the original visible light image including the spatial target;

[0102] Subtract the average gray value of each block from the pixel value of each pixel point in the block to remove background noise.

[0103] In this example, the size of the processed image is 2048 * 2048. According to Figure 2 the influence degree of the earth - atmosphere light and the target size, exemplarily, set n to be 4. For each 4 - row block B i in the image, calculate the average gray value of all pixels within the block:

[0104]

[0105] Subtract the average gray value of the block from each pixel value in the i-th block to obtain the pixel value after background removal. Subtracting the background average value block by block aims to eliminate background interference, highlight the target itself, and some noise is also eliminated.

[0106] I'(i, j, k) = I(i, j, k) - mean i Formula (3)

[0107] where I'(i, j, k) represents the pixel value after background adaptive removal processing.

[0108] Subtracting the average gray value of each block from the pixel value of each pixel point in the block mainly plays the role of removing background interference, enhancing target features, adapting to local changes, and reducing noise, thereby improving the detection accuracy and robustness of target detection.

[0109] Secondly, perform binarization processing on the image after removing background noise. The purpose is to convert the gray - scale image into a binary image with clear target boundaries, thereby enhancing the contrast between the target area and the background and providing a clear input for subsequent geometric feature extraction.

[0110] Use the gray - scale value threshold T to generate a binary image. The binarization segmentation formula is as follows:

[0111]

[0112] where I' is the image after background adaptive removal processing, I'(i, j) is the gray - scale value at the point (i, j) in the image; M is the gray - scale threshold for binarization segmentation, IB (i,j) is the binarized image. The selection of the threshold M generally uses the mean and variance distribution parameters of the star map background and adopts an adaptive threshold.

[0113] Set the adaptive binarization segmentation threshold M as follows:

[0114] M = μ + ασ Formula (5)

[0115] Where, μ is the gray mean of all pixels in the preprocessed spatial target image, σ is the standard deviation of the pixel gray values, and the constant coefficient α is used to adjust the sensitivity of M;

[0116] The value of α needs to consider the signal-to-noise ratio of the detected target and the requirements of the detection rate and false alarm rate of single-frame rough extraction, and then select a suitable value. Assume that the star map background noise and the gray value distribution probability of the target can be approximately represented by Gaussian distribution.

[0117] Exemplarily, μ = 20, σ = 576, α = 0.12.

[0118] The size of the original visible light image including the spatial target is the same as that of the preprocessed visible light image. The subsequent compression extraction algorithm only processes the image near the envelope points of the spatial target.

[0119] After the above preprocessing steps, the finally obtained image can significantly reduce the background interference and noise influence. The boundary of the target area is smoother and more complete, laying a good foundation for the subsequent detection of geometric key points of the spatial target.

[0120] The function of step S1 is to provide a clear and accurate image basis for the subsequent key point detection by preprocessing the original visible light image including the spatial target to remove background noise and enhance binarization contrast, so as to reduce redundant information and interference and improve the accuracy and reliability of key point detection.

[0121] Step S2, specifically.

[0122] Use morphological closing operation to remove the holes in the spatial target area and ensure the integrity of the target contour. Based on the target contour, use an edge detection algorithm to extract the target edge information, and screen the points on the edge through a hierarchical screening method. First, initially remove the points located in the middle of the line segment based on the curvature change, and retain the points located at both ends of the line segment as key points. Then, iteratively remove redundant points and noise points, and finally retain the effective key points that meet the geometric characteristics. To further improve the detection accuracy, optimize the screened key points to ensure the accurate position of each key point.

[0123] Apply morphological closing operation to eliminate holes, serrations, and unevenness at the edges in the target region. The closing operation can smooth the target boundary and fill small internal voids through the operations of dilation followed by erosion, thereby improving the integrity of the target region and the regularity of the boundary.

[0124] The purpose of morphological processing is to extract the spatial targets existing in the field of view from the deep space background to form connected components. Then, the closing operation is adopted, which performs the operations of dilation first and then erosion. On the premise of retaining the target morphology, the holes generated due to uneven illumination and noise are eliminated, making the spatial target a continuous connected domain.

[0125] Perform connected component analysis and extraction on the preprocessed visible light image to obtain the set of envelope points of the outer contour of the spatial target, including:

[0126] Perform morphological closing operation and connected component analysis extraction on the preprocessed visible light image, perform the operations of dilation first and then erosion, and obtain the preprocessed visible light image containing the spatial target region as follows:

[0127]

[0128] Among them, I close is the contour image of the spatial target after the closing operation, I Binary is the preprocessed visible light image, S is the structural element defined based on the shape and size characteristics of the spatial target, ⊕ is the dilation operation, is the erosion operation;

[0129] For the points in I close with non-zero gray values, judge whether there are points with pixel values different from its own pixel value in the 8-neighborhood of the point with non-zero gray value; if so, the point with non-zero gray value is the envelope point of the spatial target;

[0130] All the envelope points of the spatial target form the set of envelope points of the outer contour of the spatial target.

[0131] The physical meaning of formula (6) is that the closing operation can effectively remove small holes or small separated regions in the image. It is usually used to smooth the boundary of the spatial target, making the spatial target region more coherent, and is applicable to removing noise and filling small voids.

[0132] S is the structural element defined based on the shape and size of the spatial target, usually a rectangular or circular region.

[0133] Exemplarily, a 3×4 rectangular structural element is as follows:

[0134]

[0135] Since the visible light image including the space target contains the space target and stars, there are several connected regions after binarization and closing operation processing.

[0136] For each point (x, y) with a non-zero gray value in the visible light image including the space target, its 8-connected region includes this point and its adjacent 8 pixel points:

[0137] N8(x,y) = {(x - 1,y - 1),(x - 1,y),(x - 1,y + 1),(x,y - 1),(x,y + 1),(x + 1,y - 1),(x + 1,y),(x + 1,y + 1),}

[0138] Formula (8)

[0139] For the points in each connected region, check whether there are pixel values different from the pixel value of the point (x, y) in its 8-neighborhood. If so, this point is the target envelope point:

[0140] If there exists {I(x',y') ≠ I(x,y)|(x',y') ∈ N8(x,y)}, then (x, y) is the space target envelope point.

[0141] Where I(x, y) and I(x', y') are the pixel values of the point (x, y) and its 8-connected neighborhood point (x', y') respectively. Judge each point in each connected region. If the point meets the requirements of the outer contour envelope point, it is stored in the set as a contour point.

[0142] The main purpose of this part is to extract the outer contour points of the space target, no longer considering the entire image or the entire target.

[0143] Such as Figure 3 Shown is the detection result of the space target envelope point after image preprocessing and extraction of the outer contour of the space target.

[0144] The function of step S2 is to obtain the envelope point set of the outer contour of the space target through morphological closing operation and connected region analysis on the preprocessed visible light image. The contour of the target is repaired through the closing operation to ensure the coherence and integrity of the target area. Finally, extracting the outer contour envelope points can accurately describe the shape of the target, providing a reliable basis for subsequent geometric key point detection, thereby improving the accuracy and efficiency of space target detection.

[0145] Step S3, specifically.

[0146] This step performs hierarchical screening on the envelope point set of the target contour (i.e., the edge of the spatial target), and finally finds the precise geometric key points of the spatial target. The final output is the geometric key points in the target image. These points are located at the significant geometric positions of the target, with high accuracy and stability, and can effectively support the pose calculation task of the spatial target. Compared with traditional detection methods, it has better robustness and accuracy in dealing with noise and redundant points, and can provide strong support in complex spatial target detection and positioning tasks.

[0147] Hierarchical compression is performed by gradually screening the envelope point set of the outer contour of the spatial target. The gradual screening of the envelope point set is as follows:

[0148] The gradual screening includes the first, second, and third rounds of elimination. Among them,

[0149] In the first round of elimination, based on all the envelope points in the envelope point set, the 8-neighborhood bright point number statistics are extracted using the adaptive-step hollow convolutional neighborhood feature, and the redundant points located on the straight line or hypotenuse are determined and eliminated. The remaining envelope points after the first round of elimination are candidate key points; among them, the bright points are points with a brightness of 1.

[0150] In the second round of elimination, multi-scale dynamic threshold screening of the 24-neighborhood is performed based on the candidate key points. When the number of bright points in the 24-neighborhood of the envelope point exceeds the threshold range dynamically set based on the number of bright points in its 8-neighborhood, it is determined as a burr point and deleted. The remaining envelope points after the second round of elimination are effective key points;

[0151] In the third round of elimination, based on the effective key points, by calculating the distance and confidence difference between each effective key point and its adjacent effective key points, adjacent point clustering and merging of adjacent points are performed, and the true key points are retained.

[0152] The geometric key point extraction method based on hierarchical compression is a fast detection method designed considering saving computing resources. The method of multiple rounds of elimination is adopted to screen out suitable geometric key points according to the custom feature hierarchy.

[0153] After closing operation and connected component processing, the coordinates of each envelope point in the envelope point set of the spatial target and the area information of the spatial target are obtained. The envelope points of all connected components are extracted from the segmented binary image, and the envelope point set C is recorded. At this time, the number of envelope points of the target is large and most of them are located on the straight edges of the target. The envelope points located on the straight edges of the target have two 8-connected domain distribution situations, as Figure 4 shown, and two situations are focused on. As Figure 6 shown, for the 8-connected domain distribution situation, the points marked A are potential candidate key points.

[0154] The first round of elimination specifically includes:

[0155] Obtain the area of the spatial target based on the set of envelope points; adaptively set the neighborhood traversal step size step based on the area of the spatial target;

[0156] step = k·S target Formula (9)

[0157] where k is a scaling factor and S target is the area of the spatial target;

[0158] Taking each envelope point in the set of envelope points as the center, traversing the neighborhood in a jumping manner according to the step size step, and generating binary feature descriptors corresponding to each envelope point;

[0159] Based on the binary feature descriptors, count the number of bright points N8(x, y) in the 8-neighborhood of the corresponding envelope point. If N8(x, y) = 5, then determine that the corresponding envelope point is a redundant point on the straight edge or the oblique edge and eliminate it; retain the envelope points with N8(x, y) ≠ 5 as candidate key points.

[0160] Exemplarily, when k is 1, it means traversing each surrounding pixel one by one according to the target position.

[0161] Such as Figure 7 shown, it is the result of the candidate key points after the first round of elimination screening detection. As Figure 8 shown, it is the detection details of the candidate key points after the first round of elimination screening detection.

[0162] The first round of elimination mainly targets the envelope points on the vertical edges and oblique edges of the target, aiming to greatly reduce the detection range of geometric key points.

[0163] Since there is still noise in the binarized image, in order to avoid the adverse effects of the uneliminated noise on feature extraction, the idea of "dilated convolution" in the neural network is adopted, and the step is set to perform "jumping extraction" on the information around the envelope points.

[0164] Set different steps for spatial targets of different sizes, traverse the information around the envelope points in a jumping manner, so as to generate feature descriptors of the same size for discrimination. Exemplarily, the feature descriptor of a certain corner point is [1, 0, 1, 1, 1, 0, 1, 0, 1], where 0 and 1 represent the brightness values of pixels, 1 is a bright point, and 0 is a dark point.

[0165] After setting step, perform a jumping traversal on the neighborhood near the envelope points. Exemplarily, set S target = 2. As Figure 5 shown, step is S target= 2 Since the key geometric vertex information of the outer contour can be obtained from the neighborhood of the target envelope point, jump feature extraction is performed on the processed binary image to form the envelope point neighborhood. First, the 8-neighborhood is extracted, and the possible situations within the 8-neighborhood are divided, and the points located on the straight line are removed.

[0166] Exemplarily, for a corner point, the pixel points of the jump traversal with a step size of 2 include: (x - 2, y - 2), (x - 2, y), (x - 2, y + 2), (x, y - 2), (x, y + 2), (x + 2, y - 2), (x + 2, y), (x + 2, y + 2).

[0167] Since this step is after image binarization and closing operation, the image at this time is a binary image, which only contains 0 and 1 information and is easy to calculate. Therefore, only the brightness and darkness of the envelope point and the surrounding points are considered.

[0168]

[0169] Where N bright,8 is the number of points with a brightness of 1 within the 8-neighborhood of the envelope point in the binary image, and I bin (x', y') is the pixel value of the point (x', y') in the binary image of the preprocessed visible light image. It is stipulated that in the binary image, the value of the bright point is 1 and the value of the dark point is 0.

[0170] Studying the pixel-level scale of space targets, the 8-neighborhoods of points located on straight lines or hypotenuses usually show the following two types of distributions, that is, there are 3 pixels as the background around this point and the others are the target itself. As Figure 4 , in the figure, for easy observation, gray represents the space target and white represents the background. It can be seen that if the outer contour point of the space target is located on the straight line edge of the space target edge and is not the corner point at both ends of the straight line edge, there are usually two types of 8-neighborhoods at this time. It is stipulated that if N bright,8 = 5, then this point is removed. If N bright,8 ≠ 5, then record the coordinates of this point and N bright,8 as the basis for the second round of elimination and screening.

[0171] The function of the first-round elimination is to screen these points after obtaining the set of outer contour envelope points of the spatial target. Since the number of points is usually large while the number of real geometric key points (corner points) is relatively small, it is necessary to screen these points. Considering the morphological characteristics of the spatial target, especially in the case of straight or oblique edges, the points on the straight or oblique edges are usually dense outside their two ends, but these points do not have the characteristics of geometric key points. To effectively eliminate these non-key points, the first-round elimination traverses the 8-neighborhood of the envelope points, and regards the points with 5 bright points in the neighborhood as the points on the straight line or the oblique line and eliminates them. In this way, most of the redundant points can be removed in the preliminary screening, further reducing the computational complexity and focusing on the more representative geometric key points.

[0172] The second-round elimination specifically includes:

[0173] Based on the binary feature descriptor, calculate the number N of bright points in the 24-neighborhood of each envelope point in the candidate key points 24 (x, y);

[0174] Dynamically set N based on N8(x, y) of each envelope point in the candidate key points 24 (x, y) of the dynamic decision threshold [T min (N 24 (x, y)), T max (N 24 (x, y))] range, where T min (N 24 (x, y)), T max (N 24 (x, y)) are respectively the minimum and maximum thresholds of the number of points with brightness 1 in the 24-neighborhood;

[0175] If T min (N 24 (x, y)) ≤ N 24 (x, y) ≤ T max (N 24 (x, y)), then determine that this envelope point is a burr point and eliminate it; otherwise, determine that this envelope point is a valid key point.

[0176] The number of bright points (points with binary image value 1) in the 24-neighborhood of a certain point, and the total number of bright points in the 24-neighborhood is N 24 (x, y).

[0177] As Figure 11 shown, it is the result of the valid key points screened and detected by the second-round elimination. As Figure 12 shown, it is the detection details of the valid key points screened and detected by the second-round elimination.

[0178] Due to problems such as small imaging of space targets in images or camera shooting jitter, target trailing or uneven image edges will occur. The jagged uneven edges will also be detected as suspected candidate key points in the first-round elimination screening. Therefore, in the second-round elimination screening, the judgment range can be adaptively adjusted according to the target size, and the fringing of the image can be further removed based on the information of the 8-neighborhood of each contour point in the previous round, and the burr points can be eliminated.

[0179] Dynamically set N based on N8(x, y) of each envelope point in the candidate key points 24 (x, y) of the dynamic determination threshold, as follows:

[0180]

[0181] T min (N 24 (x, y)), T max (N 24 (x, y)) are integers after rounding.

[0182] Since the 8-connected domain contains less information and cannot cover a wide range of shape information, it is considered to use the remaining envelope points after the first-round elimination. On the basis of the original 8-connected domain search, the range is expanded to traverse the 24-neighborhood of each point, and more representative key points are further screened out from the candidate point set, and redundant points are eliminated through multi-scale analysis and local optimization. Expand the analysis range to the 24-neighborhood and calculate the total number of bright points of each point

[0183]

[0184] Among them, N bright,24 (x, y) is the number of points with a brightness of 1 in the 24-neighborhood of the envelope point in the binary image, and N 24 (x, y) is the 24-neighborhood obtained by jump traversal of the point (x, y) in the binary image, and I bin (x′, y′) is the brightness at the point (x′, y′) in the binary image.

[0185] N 24 (x, y) is smaller, indicating that the area where the point is located may be an isolated feature point, and the threshold T(N) is lower to retain more points. The larger N is, it means that the local area around the point is already close to a straight line or a curved surface, and the threshold T(N) is higher for further filtering.

[0186] Expanding the analysis range to the 24-neighborhood can capture more extensive local features and reduce the influence of single noise points on the results. Using the classification results of the 8-neighborhood eliminated in the first round can achieve more targeted screening and avoid blindly setting a unified threshold. Such as Figure 9As shown, the dark gray points are potential geometric key points to be judged. In the first round of compression, attention is paid to the 8-neighborhood formed after skip traversal processing, that is, the light gray part. When performing the second round of compression processing, it is necessary to expand 1 pixel outward in each direction and judge the 24-neighborhood.

[0187] Through the first and second rounds of elimination screening and two rounds of dynamic screening, hierarchical compression is achieved, gradually eliminating redundant points, and at the same time fully retaining the key points that contribute to the overall contour features of the target.

[0188] As Figure 10 shown, the purpose of this round of elimination is to eliminate as many false geometric key points caused by burrs as possible and identify the real geometric key points near the burr points. Multiple experiments can be carried out for this example to screen out appropriate thresholds. Since geometric key points appear as corner points of the target in the image, and according to common spatial targets, their geometric key points do not change much, so an empirical model can be established by exhaustive method through experiments to find appropriate thresholds, and these thresholds only need to be fine-tuned for different models in the same scenario. Exemplarily,

[0189]

[0190] N8(x,y) and N 24 (x,y) respectively represent the number of bright points of the current point in the 8-neighborhood and the 24-neighborhood.

[0191] For different N8(x,y) values, by setting corresponding thresholds for N 24 (x,y) to judge whether the point meets the conditions of geometric key points. For example, when N8(x,y) = 1, only the points where N 24 (x,y) is less than 17 are considered possible geometric key points; when N8(x,y) = 7, it is required that N 24 (x,y) is greater than 4 to meet the conditions.

[0192] The second-round elimination screening mechanism aims to remove noise points and retain those key points that make significant contributions to the target contour features. By combining and analyzing the number of pixels with a value of 1 in the 8-neighborhood and the 24-neighborhood, the robustness and accuracy of key point detection are effectively enhanced.

[0193] Since there may be multiple points detected as suspected geometric key points for the same geometric key point of a spatial target in a small-range pixel image, the third-round elimination adaptively adjusts the judgment range according to the target area size and the number of remaining geometric key points, and deletes multiple suspected geometric key points corresponding to the same real geometric key point.

[0194] The third-round elimination specifically includes:

[0195] Calculate the distance and confidence of each pair of adjacent valid key points;

[0196] Set a distance threshold and a confidence difference threshold;

[0197] Use the DBSCAN algorithm to group points with a distance less than the distance threshold into one class, obtaining multiple clustering clusters; retain the point with the highest confidence in each clustering cluster;

[0198] For points in each clustering cluster that are less than the distance threshold and less than the confidence difference threshold, merge them into one true key point to obtain the true key point and its position coordinates (x, y).

[0199] In the third round of elimination, first perform point matching within a small range for each point to find duplicate points in each feature point, and then screen and retain the point with the highest confidence based on the information of weighted summation of connected regions in the previous round for each duplicate point, and delete other duplicate points. The result is as Figure 13 shown, and the feature point detection is completed.

[0200] Perform adjacent point clustering. For points that are close (for example, within a preset distance threshold d threshold range), group them into one class. This process is similar to spatial clustering, and its purpose is to reduce the redundancy of multiple points in the same target area. The density clustering algorithm DBSCAN is used to automatically identify dense areas and classify them. For each cluster, only retain the point with the highest confidence as the final true key point;

[0201] In one class, screen and merge adjacent points. The merging criterion is: if the distances between multiple points in the neighborhood are less than the preset distance threshold d threshold , and their confidence difference C(p1, p2) is also small, less than the preset confidence threshold C threshold , then it is considered that these two points belong to the same target area and are merged into one geometric key point. The finally retained point is the point with the highest confidence in this class. The elimination condition is that if the distances between multiple points are very close but the confidence differences are large, preferentially retain the point with higher confidence and eliminate other low-confidence points. For each clustering cluster, finally retain the point with the highest confidence in the cluster as the geometric key point. By comparing all possible clusters and making a final choice based on the confidence difference value, ensure that only one most likely geometric key point in the same area is retained as the true key point.

[0202] Calculate the distance and relative confidence of adjacent key points. For each group of adjacent eligible valid key points, calculate the distance d and confidence C between them. Exemplarily, the Euclidean distance or Manhattan distance is selected to measure the distance between valid key points in the neighborhood.

[0203] Confidence Measure: Confidence can be scored based on the brightness features and spatial positions of each point. Points with high confidence indicate that the point is more likely to be a true geometric key point.

[0204] Exemplarily, the confidence C(p) of point p is calculated as follows:

[0205]

[0206] where B(p) and L(p) are the brightness feature and spatial position of point p respectively, λ1 and λ2 are the weight values of the brightness feature and spatial position respectively; I(p) is the brightness value of point p, α is the average brightness value of the area around point p, β is the standard deviation of the brightness values of the area around point p; center is the center of the spatial target; d(p,center) is the distance from point p to the center of the spatial target, and δ is the standard deviation of the Gaussian function.

[0207] Calculate the distance d(p1,p2) and confidence difference C(p1,p2) between two points p1 and p2 as follows:

[0208]

[0209] where (x1,x2) and (y1,y2) are the coordinates of p1 and p2 respectively, and C(p1) and C(p2) are the confidences of points p1 and p2 respectively.

[0210] If the distance d(p1,p2) between points p1 and p2 satisfies d(p1,p2) < d threshold , and the confidence difference C(p1,p2) < C threshold , the final true key point detection formula is as follows:

[0211]

[0212] Through the above process, the selection of the most representative geometric key points within the same spatial target area is achieved, avoiding misdeletion caused by noise or redundant points, and at the same time retaining the true key points that contribute to the overall morphological features of the spatial target. Through the calculation of confidence and the local neighborhood selection mechanism, the true geometric key points of the spatial target can be effectively identified and retained, ensuring the high precision and robustness of the spatial target detection.

[0213] After performing the above operations, all the geometric key points of the space target are successfully extracted. On this basis, the detection of geometric key points of space targets is of great significance in the fields of target recognition, positioning, and attitude estimation. As important features of the target shape, real geometric key points can effectively describe the geometric structure and external boundary of the target, and become the key basis for subsequent analysis tasks such as target edge detection and component recognition. By extracting the real geometric key points of the target, the local and global features of the target can be accurately captured, thereby improving the accuracy and robustness of target detection and recognition. Real geometric key points play a crucial role in target pose estimation. By matching and spatially reasoning the extracted real geometric key points, the positioning and attitude estimation of space targets can be achieved. This process not only provides a prerequisite for the precise positioning of space targets, but also provides reliable support for the real-time tracking and navigation of space targets in a dynamic environment. Especially in a complex observation environment, real geometric key points, as the input for pose calculation, can significantly improve the positioning accuracy and stability of space targets.

[0214] Embodiment 2:

[0215] Another embodiment of the present invention discloses a hierarchical compression space target key point detection system to implement a hierarchical compression space target key point detection method in Embodiment 1. The specific implementation manners of each module refer to the corresponding descriptions in Embodiment 1.

[0216] As Figure 14 shown, a hierarchical compression space target key point detection system in this embodiment includes the following modules:

[0217] An image acquisition and preprocessing module M1, configured to obtain the original visible light image of the space target and perform preprocessing to obtain the preprocessed visible light image of the space target;

[0218] A target contour envelope extraction module M2, configured to perform connected component analysis extraction on the preprocessed visible light image of the space target to obtain an envelope point set of the outer contour of the space target;

[0219] A geometric key point hierarchical compression screening module M3, configured to screen the envelope point set of the outer contour of the space target by using a hierarchical compression method to obtain the geometric key points of the space target for space target positioning and attitude estimation.

[0220] In summary, a hierarchical compression space target key point detection method according to an embodiment of the present invention has the following beneficial effects:

[0221] 1. Through the hierarchical compression method, the present invention gradually screens the envelope points, significantly reducing the computational amount, improving the detection efficiency, and meeting the real-time requirements. The preprocessing and feature extraction steps are efficient and can complete the key point detection in a short time, being applicable to real-time updates and tracking in a rapidly changing spatial environment and dynamic scenarios;

[0222] 2. The present invention adopts the dilated convolutional neighborhood feature extraction with adaptive stepping and multi-scale dynamic threshold screening, which can effectively cope with targets of different scales and shapes, enhancing the robustness of the algorithm. Through multi-scale analysis and dynamically adjusting the judgment range, it adapts to the spatial target detection under different illumination and noise conditions, having a wide range of application prospects;

[0223] 3. Through the three-round elimination mechanism, the present invention gradually eliminates the points on the straight edges and oblique edges, burr points, and duplicate redundant points, and finally extracts high-precision geometric key points, improving the accuracy of feature point detection. Through confidence evaluation, it ensures that the retained geometric key points have high confidence, further improving the detection accuracy and stability;

[0224] 4. Compared with the method based on artificial intelligence, the present application has less data requirements and resource consumption, and is more suitable for applications in resource-constrained environments, such as on-orbit observation scenarios. The preprocessing and feature extraction steps are efficient and can complete the key point detection with limited computing resources, having the characteristics of fast response, simple deployment, and strong universality, and being easy to be deployed and used in practical applications;

[0225] 5. The method of the present application is applicable to a variety of spatial target detection scenarios, including space-based observation data, simulation data, laboratory simulation data, and public data sets, having strong universality. Through multiple rounds of screening and confidence evaluation, it ensures the stability and reliability of key point detection, reduces false detection and missed detection, and improves the detection reliability.

[0226] Those skilled in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a disk, an optical disc, a read-only memory, or a random access memory, etc.

[0227] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A hierarchical compressed spatial target key point detection method, characterized in that: The steps include: Acquire an original visible light image including a space target and perform preprocessing to obtain a preprocessed visible light image; Performing connected domain analysis and extraction on the preprocessed visible light image to obtain an envelope point set of the outer contour of the space target; The envelope point set is screened and compressed level by level to obtain geometric key points of the space target for space target positioning and attitude estimation.

2. The method according to claim 1, characterized in that: The step-by-step screening includes first, second and third rounds of elimination, wherein: The first round of elimination is based on all envelope points in the envelope point set, and uses the adaptive step-by-step dilated convolutional neighborhood feature to extract the number of bright spots in the 8 neighborhoods of each envelope point, and determines the redundant points located on the straight line or the hypotenuse for elimination. The remaining envelope points after the first round of elimination are candidate key points; wherein the bright spots are points with a brightness of 1; The second round of elimination is based on the multi-scale dynamic threshold screening of 24 neighborhoods of the candidate key points. When the number of bright spots in the 24 neighborhoods of the envelope point exceeds the threshold range dynamically set based on the number of bright spots in its 8 neighborhoods, it is determined to be a burr point and deleted. The remaining envelope points after the second round of elimination are valid key points. The third round of elimination is based on the valid key points. By calculating the distance and confidence difference between each valid key point and adjacent valid key points, adjacent points are clustered and merged to retain the real key points.

3. The method according to claim 2, characterized in that: The first round of elimination specifically includes: Obtaining a spatial target area based on the envelope point set; adaptively setting a neighborhood traversal step length step based on the spatial target area; step=k·S target Where k is the scaling factor, S target is the spatial target area; Taking each envelope point in the envelope point set as the center, traverse the neighborhood in a jump manner according to the step length step to generate a binary feature descriptor corresponding to each envelope point; Based on the binary feature descriptor, the number of bright spots N8(x, y) in the neighborhood of the corresponding envelope point 8 is counted. If N8(x, y)=5, the corresponding envelope point is determined to be a redundant point on a straight edge or a hypotenuse and is removed; the envelope points with N8(x, y)≠5 are retained as candidate key points.

4. The method according to claim 3, characterized in that: The second round of elimination specifically includes: Based on the binary feature descriptor, calculate the number of bright spots N in the 24 neighborhood of each envelope point in the candidate key point 24 (x,y); Dynamically set N based on N8(x,y) of each envelope point in the candidate key points 24 The dynamic decision threshold [T min (N 24 (x,y)), T max (N 24 (x,y))], where T min (N 24 (x,y))、T max (N 24 (x, y)) are the minimum and maximum thresholds for the number of points with brightness 1 in the 24 neighborhood; If T min (N 24 (x,y))≤N 24 (x,y)≤T max (N 24 (x,y)), the envelope point is determined to be a burr point and is removed; otherwise, the envelope point is determined to be a valid key point.

5. The method according to claim 2, characterized in that: The third round of elimination specifically includes: Calculate the distance and confidence of each pair of adjacent valid key points; Set distance threshold and confidence difference threshold; Use the DBSCAN algorithm to classify points smaller than the distance threshold into one category, and obtain multiple clusters; Keep the point with the highest confidence in each cluster; For each cluster, the points that are smaller than the distance threshold and smaller than the confidence difference threshold are merged into one true key point to obtain the true key point and position coordinates (x, y).

6. The method according to claim 4, characterized in that: Dynamically set N based on N8(x,y) of each envelope point in the candidate key points 24 The dynamic decision threshold of (x,y) is as follows: T min (N 24 (x,y))、T max (N 24 (x,y)) is rounded to an integer.

7. The method according to any one of claims 1 to 6, characterized in that: Preprocessing the original visible light image including the space target includes: The background noise is removed by subtracting the background mean in blocks to obtain an image without background noise; The image with background noise removed is binarized to enhance the contrast between the space target and the background, so as to obtain a binarized image as the pre-processed visible light image of the space target.

8. The method according to claim 7, characterized in that: Performing connected domain analysis and extraction on the preprocessed visible light image to obtain an envelope point set of the outer contour of the space target, including: The preprocessed visible light image is subjected to morphological closed operation connected domain analysis and extraction, and an expansion and then corrosion operation is performed to obtain the preprocessed visible light image containing the spatial target area, as follows: Among them, I close is the contour image of the space target after the closing operation, I Binary is the preprocessed visible light image, S is the structural element defined based on the shape and size characteristics of the space target, ⊕ is the expansion operation, For corrosion operations; to I close For a point whose grayscale value is not 0, determine whether there is a point whose pixel value in the neighborhood of the point whose grayscale value is not 0 is different from its own pixel value; if there is, the point whose grayscale value is not 0 is the envelope point of the spatial target; All space target envelope points constitute the envelope point set of the space target outer contour.

9. The method according to claim 7, characterized in that: Every n lines of the original visible light image including the space target are taken as a block, and the original visible light image including the space target is divided into a plurality of blocks; Calculate the average gray value of each block; Among them, mean i is the grayscale value of all pixels in the i-th block, I(i,j,k) is the pixel value of the j-th row and k-th column in the i-th block, and W is the width of the original visible light image including the space target; The pixel value of each pixel in each block is subtracted from the average gray value of the block to remove background noise.

10. A hierarchical compressed space target key point detection system, characterized in that: Includes the following modules: The image acquisition and preprocessing module is used to acquire the original visible light image of the space target and perform preprocessing to obtain the preprocessed visible light image of the space target; A target contour envelope extraction module is used to perform connected domain analysis and extraction on the preprocessed space target visible light image to obtain an envelope point set of the space target outer contour; The geometric key point hierarchical compression screening module is used to screen the envelope point set of the outer contour of the space target using a hierarchical compression method to obtain the geometric key points of the space target for space target positioning and posture estimation.

Images