A method for extracting a missile body axis based on segmentation enhancement and robust fitting
Patent Information
- Application Number
- CN202610978553.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-02
- Publication Date
- 2026-09-18
AI Technical Summary
[0005]然而，现有基于边缘检测或轮廓特征的轴线提取方法对目标成像质量依赖较高，在光照变化、云雾干扰、图像退化、复杂背景以及局部遮挡等条件下，容易出现箭体边缘模糊、局部结构缺失、目标区域分割不稳定等问题，上述问题会导致候选轴线中点偏移、异常直线干扰增加以及最终轴线拟合结果不稳定，使得现有方法难以兼顾复杂工况下箭体目标区域的精细提取与主轴方向的鲁棒估计，进而难以满足复杂回收工况下箭体轴线高精度提取的应用需求
本发明通过箭体目标感知与分割增强中的多尺度候选区域生成、动态卷积特征增强、边缘加权监督约束，以及基于动态宽度采样与鲁棒拟合的轴线解算中的直线评分与RANSAC鲁棒拟合，考虑复杂背景干扰、弱边缘特征表征、目标尺度变化以及实时处理需求等因素，采用候选区域生成、动态卷积驱动的精细分割、动态宽度采样以及候选中点生成、直线评分与鲁棒拟合机制，实时性高、处理效率高。
Smart Images
Figure CN122780321A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of visual measurement and image processing technology for high-speed moving targets, and in particular to a method for extracting the axis of an arrow body based on segmentation enhancement and robust fitting. Background Technology
[0002] During the recovery of reusable rockets, the rocket body is subjected to dynamic conditions involving high-speed descent, continuous attitude changes, and complex environmental disturbances. The vision system needs to stably perceive and accurately represent the target structure within a short period of time. Compared to general static targets or low-speed moving targets, rocket imaging during the recovery phase often involves significant scale changes, complex background interference, decreased local contrast, and degradation of edge information. Therefore, higher requirements are placed on the accuracy, stability, and real-time performance of target structural feature extraction.
[0003] In the aforementioned visual measurement process, the rocket's axis is a crucial structural feature characterizing the target's spatial geometry and main orientation. On one hand, the axis reflects the overall extension direction of the rocket's main body region, serving as a key geometric expression for describing the contour and structural consistency of a slender target. On the other hand, the axis extraction result is a fundamental constraint for subsequent attitude analysis, motion state assessment, key structure localization, and related visual measurement processing. Therefore, if the rocket's axis extraction suffers from offsets, jitter, or false detections, it will not only reduce the accuracy of the target's main orientation characterization but also further affect the stability and reliability of subsequent geometric analysis and measurement results.
[0004] In existing technologies, high-speed cameras are typically used to acquire continuous image sequences during the rocket's descent, and these sequences are then combined with methods such as target detection, image segmentation, edge extraction, or line fitting to extract structural features of the rocket. High-speed cameras offer advantages such as non-contact operation, high frame rate, and high measurement accuracy, making them suitable for dynamic observation of high-speed moving targets. Furthermore, image processing and visual measurement methods can acquire the rocket's outline and main structural information without contacting the target, thus possessing high application value in complex recovery scenarios.
[0005] However, existing axis extraction methods based on edge detection or contour features are highly dependent on the target imaging quality. Under conditions such as changes in illumination, cloud and fog interference, image degradation, complex backgrounds, and local occlusion, problems such as blurred rocket body edges, missing local structures, and unstable target region segmentation are prone to occur. These problems can lead to the offset of the candidate axis midpoint, increased interference from abnormal straight lines, and unstable final axis fitting results. This makes it difficult for existing methods to simultaneously achieve both fine extraction of the rocket body target region and robust estimation of the main axis direction under complex working conditions, and thus it is difficult to meet the application requirements of high-precision rocket body axis extraction under complex recovery conditions. Summary of the Invention
[0006] The purpose of this invention is to provide a method for extracting the axis of an arrow body based on segmentation enhancement and robust fitting.
[0007] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Step S1, Enhanced Target Perception and Segmentation of the Arrow Body: Obtain the sequence of arrow body images to be processed; Perform candidate region generation processing on the arrow body target in the arrow body sequence image and output the arrow body candidate region; Dynamic convolutional feature enhancement processing is performed on the image features within the candidate region of the arrow body, and enhanced features are output. The enhanced features are subjected to edge-weighted supervision constraint processing to output the arrow body segmentation mask; Step S2, axis solution based on dynamic width sampling and robust fitting: Based on the rocket body segmentation mask, dynamic width sampling is performed along the rocket body region to generate a candidate axis midpoint set; Based on the set of midpoints of the candidate axes, the candidate straight line model is evaluated using a straight line scoring function, and the RANSAC algorithm is used for fitting to output the estimated axis of the rocket body.
[0008] Furthermore, the candidate region generation process includes: using a target detection strategy based on multi-scale feature maps to output a set of candidate boxes for the input image, the set of candidate boxes including multiple candidate boxes and the confidence level of each candidate box; and selecting candidate boxes with a confidence level greater than a preset threshold as candidate regions for the arrow body.
[0009] Furthermore, the dynamic convolutional feature enhancement processing includes: constructing K parallel convolutional kernels; and based on the input features... Generate the weights corresponding to each convolution kernel. ; The outputs of the K parallel convolutional kernels are aggregated according to the weights. Dynamic convolution kernel is represented as , The output features are represented as , The weights satisfy
[0010] in, Indicates the first One convolutional kernel, This represents the attention weights generated from the input features.
[0011] Furthermore, the edge-weighted supervision constraint processing includes: training the segmentation network using Gaussian-weighted binary cross-entropy loss and Dice loss; The Gaussian weighted binary cross-entropy loss can be expressed as:
[0012] Dice loss can be expressed as:
[0013] in, Indicates the weight of the edge region. Indicates the true label, Indicates the predicted probability. For the smoothing term, the total loss is the weighted sum of all loss terms.
[0014] Furthermore, the dynamic width sampling includes: Multiple sampling sections are set, and each sampling section is distributed along the longitudinal direction of the arrow body region; Obtain the x-coordinates of the intersection points of each sampling section and the left and right boundaries of the rocket body; The midpoint coordinates are calculated based on the x-coordinates of the intersection points of the same sampling section and the left and right boundaries, and candidate midpoints are generated. Summarize the candidate midpoints corresponding to each sampling section to generate a candidate midpoint set.
[0015] Furthermore, the evaluation of the candidate line model using the line scoring function includes: the candidate line model is represented by a slope-intercept equation; the distance from each point in the candidate midpoint set to the candidate line is calculated; the number of candidate midpoints with a distance less than a preset threshold is counted and used as the score value of the candidate line; wherein, the larger the score value, the higher the support of the candidate line model.
[0016] Furthermore, the fitting using the RANSAC algorithm includes: The model of the straight line to be fitted is represented by implicit equations; Calculate the distance from each point in the candidate midpoint set to the candidate line; Candidate midpoints with a distance less than a preset threshold are identified as interior points; Iterative sampling is used to select the straight line model with the most interior points as the estimated result of the rocket's axis.
[0017] The beneficial effects of this invention are: This invention is a method for extracting the axis of an arrow body based on segmentation enhancement and robust fitting. Compared with the prior art, this invention has the following technical advantages: This invention utilizes multi-scale candidate region generation, dynamic convolution feature enhancement, edge-weighted supervision constraints, and line scoring and RANSAC robust fitting in the axis solution based on dynamic width sampling and robust fitting in rocket target perception and segmentation enhancement. It takes into account factors such as complex background interference, weak edge feature representation, target scale changes, and real-time processing requirements. It adopts a candidate region generation, dynamic convolution-driven fine segmentation, dynamic width sampling, candidate midpoint generation, line scoring, and robust fitting mechanism, which has high real-time performance and high processing efficiency. Attached Figure Description
[0018] Figure 1 This is a flowchart of the working steps of a method for extracting the axis of an arrow body based on segmentation enhancement and robust fitting proposed in this invention; Detailed Implementation The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.
[0019] The present invention provides a method for extracting the axis of an arrow body based on segmentation enhancement and robust fitting, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: (1) Enhanced target perception and segmentation of rocket body In the process of extracting the rocket body axis, it is first necessary to accurately extract the target region of the rocket body from the sequence of images during the recovery phase. However, due to the characteristics of the rocket body during the recovery and descent, such as high movement speed, significant changes in imaging scale, slender target structure and narrow edges, and susceptibility to factors such as changes in illumination, cloud and fog interference, complex background and image degradation, traditional segmentation methods are prone to problems such as incomplete response of the target region, weakened edge structure and unstable local contour extraction, thereby reducing the accuracy and reliability of subsequent axis calculation.
[0020] Therefore, it is necessary to establish a target perception and segmentation enhancement method for rocket bodies that takes into account both fine segmentation capabilities and real-time processing performance, so as to improve the integrity, stability and accuracy of target region extraction under complex working conditions.
[0021] To this end, this section first narrows the scope of subsequent fine processing by generating candidate regions, reducing the interference of complex backgrounds and redundant regions on the target extraction process; then, it enhances the network's adaptive representation ability of the arrow body edge and main structure by feature enhancement for slender target structures; and further enhances the integrity and stability of the segmentation results under weak edge and complex background conditions by edge region supervision constraints.
[0022] To address the aforementioned issues, this invention proposes a method for enhancing rocket target perception and segmentation. This method comprehensively considers factors such as complex background interference, weak edge feature representation, target scale variations, and real-time processing requirements. The specific steps are as follows: ① Candidate region generation A target detection strategy based on multi-scale feature maps is used to coarsely locate the arrow body target in the input image to obtain candidate regions required for subsequent fine processing. Considering the significant scale changes, complex background interference, and low local contrast of the arrow body during descent and recovery, directly performing fine segmentation on the entire image would easily increase computational redundancy and introduce irrelevant regions that interfere with the target structure response. Therefore, preliminary constraints on the arrow body target are first imposed through candidate region generation. Let the candidate box set output by the detection network be:
[0023] in, Indicates the first One candidate box, This indicates the corresponding confidence level. Indicates the number of candidate boxes. Select those that meet the criteria. The candidate bounding box is used as the candidate region for the arrow body, where The preset threshold is used. This allows for narrowing the scope of subsequent processing while ensuring the integrity of the target response, reducing interference from complex backgrounds and redundant regions in the rocket body target extraction process, improving the targeting and real-time processing efficiency of rocket body segmentation, and providing a reliable input basis for subsequent fine segmentation and structural feature enhancement.
[0024] ② Weighted dynamic convolution feature enhancement for slender targets After obtaining candidate regions for the arrow body, fine segmentation of the candidate region image is performed. Considering that the subsequent segmentation results need to simultaneously reflect the integrity of the main body region and the continuity of the edge contours, this implementation introduces a dynamic convolution mechanism in the feature extraction process of the segmentation network to improve the network's ability to represent the structural information of the arrow body target. Specifically, by constructing multiple parallel convolution kernels and generating corresponding weight coefficients based on the input features, the output results of each convolution kernel are adaptively aggregated, thereby achieving dynamic response and effective extraction of features from different regions.
[0025] Let the input features be Dynamic convolution is The convolution kernel is obtained by weighted aggregation of parallel convolution kernels, and its convolution kernel can be represented as:
[0026] in, Indicates the first One convolutional kernel, The attention weights generated from the input features correspond to the following output features:
[0027] in, satisfy:
[0028] In this way, the network can adaptively adjust the response weights of different convolutional kernels according to the arrow's size, background texture, and lighting changes in the input image, thereby improving its ability to extract the edges of slender targets, local contours, and details of the main structure.
[0029] Furthermore, this mechanism can enhance the effective structural response under complex background, low contrast and weak edge conditions, reduce the dependence of fixed feature extraction methods on a single imaging mode, and provide a more stable feature basis for subsequent edge supervision constraints and high-quality segmentation output.
[0030] ③ Edge-weighted supervision and constraint After completing candidate region generation and feature enhancement, to further improve the boundary integrity and contour stability of the arrow body segmentation results, this implementation introduces edge-weighted supervision constraints, assigning higher supervision weights to pixels adjacent to the edges. Considering the narrow edges of the arrow body target, the high requirements for contour continuity, and its susceptibility to complex backgrounds and weak edge conditions, implementing stronger supervision constraints on the contour region can enhance the network's ability to represent boundary positions and local detail information.
[0031] The Gaussian weighted binary cross-entropy loss can be expressed as:
[0032] Dice loss can be expressed as:
[0033] in, Indicates the weight of the edge region. Indicates the true label, Indicates the predicted probability. This is a smoothing term.
[0034] Furthermore, by incorporating the aforementioned loss terms into the total loss function, the total loss can be expressed as:
[0035] By applying the aforementioned edge-weighted supervision constraints, we can obtain arrow segmentation results with more complete boundaries and more stable structural responses, providing a foundation for subsequent axis midpoint generation and robust axis extraction.
[0036] (2) Axis solution based on dynamic width sampling and robust fitting After obtaining the segmentation mask of the rocket body, to adapt to the changes in imaging scale of the rocket body under different observation distances and attitudes, this implementation method performs dynamic width sampling based on the segmented region and generates candidate axis midpoints. Let the first... The x-coordinates of the intersection points of each sampling section and the left and right boundaries of the rocket body are respectively and Then the corresponding candidate midpoint can be represented as:
[0037] in, Indicates the first The longitudinal position coordinates of each sampling section. By performing dynamic width sampling along the arrow body region, a set of candidate midpoints for axis fitting can be generated.
[0038] After obtaining the candidate midpoint set, a straight line scoring function is used to evaluate the candidate axes to obtain the optimal axis model. Let the candidate straight line model be... Then its scoring function can be expressed as:
[0039] in, Denotes the set of candidate midpoints. Indicates the coordinates of the midpoint. and Let represent the slope and intercept of the candidate line, respectively. Indicates the distance threshold. This is an indicator function. When the distance from the candidate midpoint to the line is less than a threshold... When a point is found to be a valid support point, the number of support points is used to measure the quality of the current linear model. The scoring function uses a slope-intercept linear model for evaluation. In the subsequent RANSAC robust fitting, to facilitate the calculation of the distance from the candidate midpoint to the line, the line is further written as an implicit equation. The two representations are equivalent.
[0040] Based on this, to suppress the influence of local noise, edge missegmentation, and outlier midpoints, RANSAC is further used for robust fitting of candidate midpoints. Let the linear model to be fitted be... Then the first Candidate midpoints The distance to this line can be expressed as:
[0041] when When the point is identified as an interior point, the line model with the most interior points is selected as the final rocket axis estimation result through iterative sampling. By combining the candidate midpoint generated by dynamic width sampling, the line scoring function, and RANSAC robust fitting, stable extraction of the rocket axis can be achieved under complex backgrounds and incomplete local edges.
[0042] In this embodiment, the test data consists of a sequence of 1,000 images captured by a high-speed camera during the simulated recovery phase of a reusable rocket. The image resolution is 1920×1080 pixels. The rocket body is elongated in the image, with its length accounting for 30% to 70% of the image height and its width ranging from 10 to 30 pixels.
[0043] The comparison method is an axis extraction method based on Canny edge detection and least squares line fitting.
[0044] From a sequence of 1000 images, 100 frames containing background interference such as clouds, fog, lighting changes, and buildings, trees, etc., were selected as test samples. Using the manually labeled true axis of the rocket body as a reference, the axis angle error and position offset error were calculated.
[0045] On 100 frames of test images with complex backgrounds, the average angle error of the axis extracted by the comparative method was 2.37 degrees, with a standard deviation of 1.84 degrees; the average angle error of the axis extracted by our method was 0.52 degrees, with a standard deviation of 0.31 degrees. The average positional offset of the axis extracted by the comparative method was 8.62 pixels, with a standard deviation of 6.91 pixels; the average positional offset of the axis extracted by our method was 1.73 pixels, with a standard deviation of 0.95 pixels. The effective detection rate of the comparative method was 73.0%, while the effective detection rate of our method was 97.0%.
[0046] From a sequence of 1000 images, 200 frames were selected containing continuously varying arrow dimensions, with the arrow length's proportion of the image height varying from 30% to 70%. On these 200 scale-variation test images, the comparative method showed an axis angle drift range of -4.21 degrees to 3.87 degrees, with a standard deviation of 1.96 degrees; the proposed method showed an axis angle drift range of -0.68 degrees to 0.59 degrees, with a standard deviation of 0.28 degrees. The comparative method showed a positional offset drift range of -15.3 pixels to 12.8 pixels, with a standard deviation of 7.23 pixels; the proposed method showed a positional offset drift range of -2.1 pixels to 2.4 pixels, with a standard deviation of 1.12 pixels.
[0047] The average single-frame processing time of this method and the comparison method was tested on 1000 frames of images. The average single-frame processing time of the comparison method was 45.2 milliseconds, with a processing frame rate of 22.1 frames / second; the average single-frame processing time of this method was 52.8 milliseconds, with a processing frame rate of 18.9 frames / second.
[0048] The experimental results above show that the proposed method can achieve high-precision extraction of the rocket body axis under complex background conditions and scale variations, and the processing frame rate can meet the real-time measurement requirements of the rocket body recovery process.
[0049] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for extracting the axis of an arrow body based on segmentation enhancement and robust fitting, characterized in that, Includes the following steps: Step S1, Enhanced Target Perception and Segmentation of the Arrow Body: Obtain the sequence of arrow body images to be processed; Perform candidate region generation processing on the arrow body target in the arrow body sequence image and output the arrow body candidate region; Dynamic convolutional feature enhancement processing is performed on the image features within the candidate region of the arrow body, and enhanced features are output. The enhanced features are subjected to edge-weighted supervision constraint processing to output the arrow body segmentation mask; Step S2, axis solution based on dynamic width sampling and robust fitting: Based on the rocket body segmentation mask, dynamic width sampling is performed along the rocket body region to generate a candidate axis midpoint set; Based on the set of midpoints of the candidate axes, the candidate straight line model is evaluated using a straight line scoring function, and the RANSAC algorithm is used for fitting to output the estimated axis of the rocket body.
2. The method according to claim 1, characterized in that, The candidate region generation process includes: using a target detection strategy based on multi-scale feature maps to output a set of candidate boxes for the input image, the set of candidate boxes including multiple candidate boxes and the confidence level of each candidate box; and selecting candidate boxes with a confidence level greater than a preset threshold as candidate regions for the arrow body.
3. The method for extracting the arrow body axis based on segmentation enhancement and robust fitting according to claim 1, characterized in that, The dynamic convolutional feature enhancement process includes: constructing K parallel convolutional kernels; and based on the input features... Generate the weights corresponding to each convolution kernel. ; The outputs of the K parallel convolutional kernels are aggregated according to the weights. Dynamic convolution kernel is represented as ; The output features are represented as ; The weights satisfy ; in, Indicates the first One convolutional kernel, This represents the attention weights generated from the input features.
4. The method for extracting the arrow body axis based on segmentation enhancement and robust fitting according to claim 1, characterized in that, The edge-weighted supervision constraint processing includes: training the segmentation network using Gaussian-weighted binary cross-entropy loss and Dice loss; The Gaussian weighted binary cross-entropy loss can be expressed as: ； Dice loss can be expressed as: ； in, Indicates the weight of the edge region. Indicates the true label, Indicates the predicted probability. For the smoothing term, the total loss is the weighted sum of all loss terms.
5. The method for extracting the arrow body axis based on segmentation enhancement and robust fitting according to claim 1, characterized in that, The dynamic width sampling includes: Multiple sampling sections are set, and each sampling section is distributed along the longitudinal direction of the arrow body region; Obtain the x-coordinates of the intersection points of each sampling section and the left and right boundaries of the rocket body; The midpoint coordinates are calculated based on the x-coordinates of the intersection points of the same sampling section and the left and right boundaries, and candidate midpoints are generated. Summarize the candidate midpoints corresponding to each sampling section to generate a candidate midpoint set.
6. The method for extracting the arrow body axis based on segmentation enhancement and robust fitting according to claim 1, characterized in that, The evaluation of the candidate line model using the line scoring function includes: the candidate line model is represented by a slope-intercept equation; the distance from each point in the candidate midpoint set to the candidate line is calculated; the number of candidate midpoints whose distance is less than a preset threshold is counted and used as the score value of the candidate line; wherein, the larger the score value, the higher the support of the candidate line model.
7. The method for extracting the arrow body axis based on segmentation enhancement and robust fitting according to claim 1, characterized in that, The fitting process using the RANSAC algorithm includes: The model of the straight line to be fitted is represented by implicit equations; Calculate the distance from each point in the candidate midpoint set to the candidate line; Candidate midpoints with a distance less than a preset threshold are identified as interior points; Iterative sampling is used to select the straight line model with the most interior points as the estimated result of the rocket's axis.