Multi-feature fused variable-scale moving target positioning and tracking method and device

Abstract

The invention discloses a multi-feature-fused variable-scale moving target positioning and tracking method and device, and relates to the technical field of image processing, the variable-scale moving target positioning and tracking method comprises the following steps: obtaining feature points of a workpiece image, and establishing a feature point matching relationship between adjacent workpiece image frames; determining a preliminary workpiece contour area, and dividing workpiece categories based on the preliminary workpiece contour area to obtain a final workpiece contour area; determining contour parameters; and dynamically positioning and tracking the workpiece by combining the pose vector of the workpiece and the three-dimensional point cloud of the workpiece. Through a multi-scale feature fusion algorithm, richer and more detailed feature points can be extracted from the workpiece image, the capability of capturing fine features of the workpiece is enhanced, and a more accurate feature point matching relation can be established between adjacent image frames, so that the overall positioning precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a variable-scale moving target positioning and tracking method and equipment integrating multiple features. Background Art

[0002] In complex industrial environments, the continuous movement of identified workpieces on the assembly line poses a challenge to real-time tracking and positioning. In order to effectively achieve this goal, it is particularly important to study the feature extraction technology of variable-scale moving targets that integrates multiple features. This technology needs to be able to adapt to the changing environment and the dynamics of the target to ensure accurate identification and tracking of the workpiece. The dynamic tracking system should be able to automatically adjust its parameters to adapt to targets of different scales to optimize the tracking process. This includes using deep learning networks to automatically identify the most effective features and adjust the network structure in real time to adapt to changes in target size. Ultimately, this system will be able to provide continuous and accurate workpiece location data, support automated operation and monitoring on the assembly line, and greatly improve production efficiency and safety.

[0003] In the prior art, it is not convenient to extract rich and detailed feature points from the workpiece image, which reduces the ability to capture subtle features of the workpiece, and it is not convenient to establish a more accurate feature point matching relationship between adjacent image frames, thereby reducing the overall positioning accuracy. It is not convenient to perform preliminary detection of the workpiece contour, nor is it convenient to optimize the accuracy of contour parameters, and the ability to handle abnormal values ​​is reduced. At the same time, it is not convenient to estimate the pose vector of the workpiece, and it is not convenient to use the pose vector and three-dimensional point cloud data to achieve dynamic tracking. It cannot effectively cope with the high-speed movement and position changes of the workpiece on the production line, and cannot ensure the continuity and real-time nature of tracking.

[0004] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention proposes a variable-scale moving target positioning and tracking method and equipment that integrates multiple features, which solves the problems mentioned in the above background technology, such as it is not convenient to extract rich and detailed feature points from the workpiece image, reduces the ability to capture subtle features of the workpiece, and is not convenient to establish a more accurate feature point matching relationship between adjacent image frames, thereby reducing the overall positioning accuracy, and is not convenient for preliminary detection of the workpiece contour, nor for optimizing the accuracy of contour parameters, and reduces the ability to handle abnormal values. At the same time, it is not convenient to estimate the pose vector of the workpiece, and it is not convenient to use the pose vector and three-dimensional point cloud data to achieve dynamic tracking, and it cannot effectively cope with the high-speed movement and position changes of the workpiece on the production line, and cannot ensure the continuity and real-time of tracking.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: According to one aspect of the present invention, a variable-scale moving target positioning and tracking method integrating multiple features is provided, and the variable-scale moving target positioning and tracking method comprises the following steps: S1. Acquire an image of the workpiece to be identified, extract features from the workpiece image using a multi-scale feature fusion algorithm, obtain feature points of the workpiece image, and establish a feature point matching relationship between adjacent workpiece image frames; S2. Perform preliminary detection on the contour of the workpiece image using Hough transform, determine a preliminary workpiece contour area, and classify the workpiece categories based on the preliminary workpiece contour area to obtain a final workpiece contour area; S3, using the double point removal least squares method to regress the feature points in the final workpiece contour area to determine the contour parameters; S4, obtaining the three-dimensional point cloud and image features of the workpiece, estimating the pose vector of the workpiece using a deep convolutional neural network, and dynamically positioning and tracking the workpiece by combining the pose vector of the workpiece with the three-dimensional point cloud of the workpiece; Acquiring an image of the workpiece to be marked, extracting features from the workpiece image using a multi-scale feature fusion algorithm, obtaining feature points of the workpiece image, and establishing a feature point matching relationship between adjacent workpiece image frames includes the following steps: S11, extracting feature points from the acquired workpiece image using a multi-scale feature fusion algorithm; S12, extracting stable feature points from the workpiece image, and calculating a descriptor of each feature point to characterize image information around the feature point; S13, matching feature points in two adjacent workpiece images, and realizing feature point matching relationship by calculating the distance and similarity between two descriptors.

[0007] Furthermore, extracting feature points from the acquired workpiece image using a multi-scale feature fusion algorithm includes the following steps: S111, performing multi-scale transformation on the workpiece image by using a Gaussian difference pyramid, and generating workpiece images in different scale spaces; S112, detecting local extreme points as candidate feature points by comparing adjacent pixels in each scale space workpiece image; S113, calculating the gradient and direction of the surrounding pixels for each candidate feature point, generating a gradient direction histogram, and determining the main direction of the feature point with the peak value of the histogram; S114, constructing a local area with each feature point as the center, and calculating the gradient direction and amplitude in the local area, and combining them into a descriptor vector as a descriptor of the feature point to perform feature matching.

[0008] Further, the contour of the workpiece image is preliminarily detected by using Hough transform, a preliminary workpiece contour area is determined, and the workpiece categories are divided based on the preliminary workpiece contour area to obtain the final workpiece contour area, which includes the following steps: S21, converting the workpiece color image into a grayscale image, and applying Gaussian filtering to remove image noise; S22, extracting edges in the grayscale image using an edge detection algorithm; S23, setting the parameter space according to the shape of the workpiece to be detected; S24, increasing the count of all parameters corresponding to each edge point on the grayscale image in the parameter space to form a cumulative array, and searching for a local maximum value in the cumulative array; S25, obtaining a peak value from the accumulation array as a parameter of the detected workpiece shape, and using the parameter of the workpiece shape to draw a contour on the original workpiece image as a preliminary workpiece contour area; S26. Based on the preliminary workpiece contour area, the workpiece categories are divided using a division algorithm to obtain a final workpiece contour area.

[0009] Further, based on the preliminary workpiece contour area, the workpiece category is divided by using a division algorithm to obtain the final workpiece contour area, including the following steps: S261, setting the scaling factor and cross factor of the partitioning algorithm, and determining the maximum number of iterations and the number of chaotic searches; S262, classifying the workpiece according to the characteristics of the workpiece image and initializing candidate solutions for the workpiece contour; S263, randomly selecting a workpiece contour candidate solution from the workpiece contour candidate solutions according to each workpiece category, and generating a new variant workpiece contour candidate solution; S264, generating a new candidate solution for the workpiece contour by fusing different candidate solutions for the workpiece contour through cross factors; S265, comparing the newly generated candidate workpiece contour solution with the current optimal workpiece contour solution, and retaining the optimal workpiece contour solution; S266, repeating steps S263 to S265 until a predetermined number of iterations is reached; S267, performing a chaotic search around the best workpiece contour solution of each category, and selecting the workpiece contour with the highest fitness from the chaotic search as the final workpiece contour region.

[0010] Furthermore, the formula for generating new candidate solutions for variant workpiece contours is: ; In the formula, A i ( m +1) indicates the next iteration m+1 at the time i Candidate solutions for variant workpiece contours; G Expressed as a scaling factor; C r1 ( m )、 C r2 ( m )and C r2 ( m ) are respectively represented as m A candidate solution of the workpiece contour randomly selected from the candidate solutions of the workpiece contour at the time; r 1. r 2 and r 3 are random variables.

[0011] Furthermore, the feature points in the final workpiece contour area are regressed using the double point removal least squares method to determine the contour parameters, including the following steps: S31, obtaining a feature point data set within the final workpiece contour area; S32, using the least squares method to perform a preliminary linear regression analysis on all feature points to obtain an initial contour parameter estimate; S33, removing two feature points in the feature point data set one by one, and re-performing least squares regression using the remaining feature points; S34, calculating the difference between the contour parameters obtained by least square regression after removing two feature points each time and the initial contour parameter estimation, and recording the difference value; S35, repeating step S33 and step S34 until a preset number of iterations is reached; S36. Analyze the difference values ​​of all records, and select a set of regression results with the smallest difference values ​​as the estimation of the final profile parameters.

[0012] Further, analyzing the difference values ​​of all records and selecting a set of regression results with the smallest difference values ​​as the estimation of the final profile parameters includes the following steps: S361, obtain all historical regression result data sets, and initialize the regression result data sets and related constraints; S362, accepting a new regression result data update instruction and selecting a data processing mode according to analysis requirements; S363, if the priority of minimizing the difference value is selected, the constraint boundary of the data set is calculated by a specific formula, and if other priorities are selected, the boundary of the original data set is kept unchanged; S364, calculating the deviation of each set of regression result data and the deviation of the overall regression result data set, and selecting data points that need to be adjusted first; S365, performing regression analysis optimization on the selected data points using an optimization algorithm, and recording the difference value of each optimization; S366. Analyze the difference values ​​of all records, and select a set of regression results with the smallest difference value as the final profile parameters by comparing the calculated difference values.

[0013] Further, the selected data points are optimized using an optimization algorithm for regression analysis, and the difference value of each optimization is recorded, including the following steps: S3651. Randomly initialize several groups of parameters of the regression model, each group of parameters represents a potential regression model solution; S3652, calculating the fitness of each set of regression model parameters, and selecting the regression model with the minimum difference value as the core regression model; S3653, adjusting the non-core regression model parameter set, using different adjustment strategies for the core regression model parameter set, and determining the number of new parameters to be generated according to their respective difference values; S3654. Generate a new parameter set based on the current regression model parameters and perform boundary detection; S3655, applying Gaussian mutation to selected regression model parameters to generate additional parameter sets; S3656. From all current regression model parameters, select a set of parameters with the smallest difference value to continue iteration, and record the best difference value; S3657. Check whether the predetermined number of iterations has been reached. If reached, stop the iteration and record the minimum difference value as the regression result.

[0014] Furthermore, obtaining the three-dimensional point cloud and image features of the workpiece, estimating the pose vector of the workpiece using a deep convolutional neural network, and dynamically positioning and tracking the workpiece by combining the pose vector of the workpiece with the three-dimensional point cloud of the workpiece include the following steps: S41. Extract features from workpiece images using a deep residual network and design a pose estimation network. S42, transforming the output of the feature extraction part through a fully connected layer to regress a pose vector; S43, dividing the output of the posture regression part into a position vector and a posture vector, as outputs of the estimated postures of the multiple classes of objects; S44, collecting the three-dimensional point cloud and image data of the workpiece in real time in a dynamic environment, and using the trained deep residual network to continuously update the pose estimation of the workpiece; S45. Combine the real-time updated pose vector and 3D point cloud to implement accurate dynamic positioning and tracking of workpieces.

[0015] According to another aspect of the present invention, a variable-scale moving target positioning and tracking device integrating multiple features is also provided, and the variable-scale moving target positioning and tracking device integrating multiple features includes: An image processing module is used to obtain an image of the workpiece to be marked, extract features from the workpiece image using a multi-scale feature fusion algorithm, obtain feature points of the workpiece image, and establish a feature point matching relationship between adjacent workpiece image frames; The workpiece detection module is used to perform preliminary detection on the contour of the workpiece image using Hough transform, determine the preliminary workpiece contour area, and classify the workpiece categories based on the preliminary workpiece contour area to obtain the final workpiece contour area; The data analysis and regression model module is used to regress the feature points in the final workpiece contour area using the double point removal least squares method to determine the contour parameters; The target positioning and tracking module is used to obtain the three-dimensional point cloud and image features of the workpiece, estimate the workpiece's pose vector using a deep convolutional neural network, and dynamically position and track the workpiece by combining the workpiece's pose vector with the workpiece's three-dimensional point cloud; Among them, the image processing module is connected through the workpiece detection module and the data analysis and regression model module, and the data analysis and regression model module is connected with the target positioning and tracking module.

[0016] The beneficial effects of the present invention are: 1. The present invention can extract richer and more detailed feature points from the workpiece image through a multi-scale feature fusion algorithm, enhance the ability to capture subtle features of the workpiece, and help to establish a more accurate feature point matching relationship between adjacent image frames, thereby improving the overall positioning accuracy. The application of Hough transform in contour detection can provide a stable preliminary contour area for workpiece classification and tracking, which helps to maintain efficient operation under conditions of visual noise or partial occlusion. The double-point removal least squares rule further optimizes the accuracy of contour parameters, and enhances the system's ability to handle outliers by eliminating abnormal data points. The deep convolutional neural network is combined to estimate the pose vector of the workpiece, and the pose vector and three-dimensional point cloud data are used to achieve dynamic tracking, which can effectively cope with the high-speed movement and position changes of the workpiece on the production line, ensuring the continuity and real-time tracking.

[0017] 2. The present invention can capture the details and structure of the image at different scale levels through a multi-scale feature fusion algorithm. This method is extremely effective for detecting feature points that remain stable under scale transformation. This capability is particularly suitable for scenarios where the size, shape or orientation of the workpiece changes rapidly. The descriptor constructed by calculating the gradient and direction of the pixels around each feature point not only provides the location information of the point, but also accurately describes its local image features. The distance and similarity between the descriptors are used to match the feature points, making the matching process more robust, especially when the image is slightly distorted or occluded due to the movement of the workpiece. It can then effectively connect the same workpiece in consecutive image frames, providing stable input for dynamic tracking.

[0018] 3. The present invention enhances the accurate recognition of workpiece contours by utilizing Hough transform combined with multi-step edge detection and shape parameter extraction, especially when the image quality is disturbed by noise. Grayscale conversion and Gaussian filtering help reduce processing errors and ensure the accuracy of edge detection. The partitioning algorithm used in the classification process further optimizes the category recognition of workpieces, thereby ensuring that each type of workpiece can be correctly classified. The chaotic search technology is used to optimize the workpiece contour after the classification is determined, which not only improves the accuracy of tracking, but also increases the adaptability to dynamic changes. By integrating different solutions and adjusting algorithm parameters such as scaling factors and cross factors, it can flexibly respond to the diversity and complexity of workpiece visual presentation, thereby maintaining efficient performance in a changing industrial environment.

[0019] 4. The present invention provides a robust baseline by using the least squares method for preliminary regression. By removing potential abnormal feature points one by one and recalculating the regression, the influence of noise and outliers can be significantly reduced, and the accuracy and robustness of parameter estimation can be improved. By continuously analyzing the difference values ​​and selecting the regression results with the smallest difference values, the model is guaranteed to perform optimally in the face of variable and complex data. This method is particularly suitable for dynamically changing industrial environments, ensuring that the estimation of profile parameters can adapt to actual physical changes. By finely managing each data point and iterative process, it can more effectively process large amounts of data, reduce the waste of computing resources, and thus optimize operating efficiency and output quality.

[0020] 5. The present invention extracts high-quality features from the workpiece image through a deep residual network, and accurately calculates the pose vector through a pose estimation network, thereby achieving efficient dynamic tracking. In a dynamic environment, it can collect and process three-dimensional point clouds and image data in real time, and continuously update the pose estimation of the workpiece. This not only improves the system's ability to adapt to environmental changes, but also ensures that the continuity and accuracy of tracking can be maintained under various working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0022] Figure 1 is a flow chart of a variable-scale moving target positioning and tracking method integrating multiple features according to an embodiment of the present invention; Figure 2 It is a principle block diagram of a variable-scale moving target positioning and tracking device integrating multiple features according to an embodiment of the present invention.

[0023] In the figure: 1. Image processing module; 2. Workpiece detection module; 3. Data analysis and regression model module; 4. Target positioning and tracking module. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0025] In the description of the present invention, unless otherwise specified, the meaning of "plurality" is two or more. In addition, the terms "first", "second", "third", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0026] According to an embodiment of the present invention, a method and equipment for positioning and tracking a variable-scale moving target by integrating multiple features are provided.

[0027] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to a variable-scale moving target positioning and tracking method integrating multiple features according to an embodiment of the present invention, the variable-scale moving target positioning and tracking method comprises the following steps: S1. Acquire an image of the workpiece to be identified, extract features from the workpiece image using a multi-scale feature fusion algorithm, obtain feature points of the workpiece image, and establish a feature point matching relationship between adjacent workpiece image frames; S2. Performing preliminary detection on the contour of the workpiece image using Hough transform, determining a preliminary workpiece contour area, and classifying the workpiece categories based on the preliminary workpiece contour area to obtain a final workpiece contour area; S3, using the double point removal least squares method to regress the feature points in the final workpiece contour area to determine the contour parameters; S4, obtaining the three-dimensional point cloud and image features of the workpiece, estimating the pose vector of the workpiece using a deep convolutional neural network, and dynamically positioning and tracking the workpiece by combining the pose vector of the workpiece with the three-dimensional point cloud of the workpiece; It should be explained that the workpiece is continuously photographed by the camera to obtain images. Using multi-scale feature fusion algorithms, such as the scale-invariant feature transformation algorithm, each workpiece image is processed to extract key feature points. The scale-invariant feature transformation algorithm provides a scale- and rotation-invariant feature description by detecting local extreme points in different scale spaces and calculating their directions and gradients. It is suitable for maintaining the stability of features at multiple scales. The edges and contours of each workpiece image are detected by Hough transform. After the preliminary contour is determined, the workpiece is classified using a shape-based classification algorithm (such as support vector machine, chaotic differential evolution algorithm, etc.) to determine which type of workpiece it is. After the preliminary classification, the feature points in the contour area of ​​the workpiece belonging to the same category are subjected to regression analysis using the double-point removal least squares method to determine more accurate contour parameters. In the assembly line, the 3D scanner obtains the 3D point cloud data of the workpiece in real time. Combined with the pose vector estimated by the deep convolutional neural network, the position and direction of the workpiece on the assembly line can be tracked in real time.

[0028] Acquiring an image of the workpiece to be marked, extracting features from the workpiece image using a multi-scale feature fusion algorithm, obtaining feature points of the workpiece image, and establishing a feature point matching relationship between adjacent workpiece image frames includes the following steps: S11, extracting feature points from the acquired workpiece image using a multi-scale feature fusion algorithm; S12, extracting stable feature points from the workpiece image, and calculating a descriptor of each feature point to characterize image information around the feature point; S13, matching feature points in two adjacent workpiece images, and realizing feature point matching relationship by calculating the distance and similarity between two descriptors.

[0029] It should be explained that the factory takes multiple continuous images of the chip (workpiece) through a camera. Since there may be complex textures or tiny identification marks on the chip, a multi-scale feature fusion algorithm (such as the scale-invariant feature transformation algorithm, the accelerated robust feature algorithm or the ORB algorithm) is used to extract feature points in the chip image. These feature points can be edges, marks or other significant areas on the chip. The role of the multi-scale feature fusion algorithm is that it can extract stable feature points from different scale spaces of the image. For example, the scale-invariant feature transformation algorithm decomposes the image into multiple scale spaces through the Gaussian difference pyramid and finds local extreme points in these spaces as feature points. This method is robust to rotation, scaling and illumination changes, and is very suitable for chip image processing in dynamic industrial environments. Each feature point extracted from the chip image will calculate a descriptor, which is used to characterize the image structure and information around the feature point. The role of the descriptor is to provide uniqueness for the feature point, that is, even if the chip image is slightly deformed or shot at different angles, these feature points can still be identified in other images. The descriptor is a code used to represent the local information of the feature point as a vector. For example, the scale-invariant feature transform algorithm descriptor encodes the local features of the feature point by calculating the gradient direction histogram of the area around the feature point. These descriptors are crucial for matching feature points and eliminating noise. Assume that in two consecutive image frames, the first frame and the second frame contain the same chip. The feature points are matched by comparing the feature point descriptors in these frames. Usually, the matching process determines their similarity by calculating the Euclidean distance between two descriptors. The higher the similarity, the more it means that the two feature points come from the same location. Through the matching relationship of feature points, the system can track the movement trajectory of the chip on the assembly line. Feature point matching is achieved by calculating the distance between descriptors. The most common method is to use Euclidean distance or other similarity metrics. For example, the nearest neighbor algorithm can quickly match two feature points. In order to eliminate false matches, a "bidirectional matching" strategy is usually used, that is, to ensure that feature points A and B are the best match for each other.

[0030] Preferably, extracting feature points from the acquired workpiece image using a multi-scale feature fusion algorithm comprises the following steps: S111, performing multi-scale transformation on the workpiece image by using a Gaussian difference pyramid, and generating workpiece images in different scale spaces; Specifically, performing multi-scale transformation on the workpiece image by using a Gaussian difference pyramid and generating workpiece images of different scale spaces includes the following steps: S1111, performing Gaussian filtering on the workpiece image to obtain Gaussian blurred images of different scales; S1112, performing a downsampling operation on each Gaussian blurred image to obtain low-resolution images of different scales; S1113, performing a subtraction operation on each adjacent low-resolution image to obtain Gaussian difference images of different scales as spatial artifact images of different scales.

[0031] It should be explained that the camera obtains the original image of the workpiece. In order to process workpiece images of different scales, each image is first Gaussian filtered. The Gaussian filtering process smoothes the image and reduces noise through convolution operations. Images with different blur levels are generated by setting different standard deviations. The larger the standard deviation, the higher the blur level of the image. Each Gaussian filtered image is downsampled, that is, the resolution of the image is reduced. For example, every other pixel of the original image is sampled to generate a smaller image. This process can quickly generate multi-scale low-resolution images, representing the workpiece at different perspectives. Through downsampling operations, images of different resolutions can be generated to adapt to the different visual effects caused by the distance or size changes of the workpiece in the actual scene. Each adjacent low-resolution image is subtracted pixel by pixel to generate a Gaussian difference image. The role of this step is to emphasize the edges and feature areas in the workpiece image through differential operations and eliminate irrelevant information in the image.

[0032] S112, detecting local extreme points as candidate feature points by comparing adjacent pixels in each scale space workpiece image; S113, calculating the gradient and direction of the surrounding pixels for each candidate feature point, generating a gradient direction histogram, and determining the main direction of the feature point with the peak value of the histogram; S114, constructing a local area with each feature point as the center, and calculating the gradient direction and amplitude in the local area, and combining them into a descriptor vector as a descriptor of the feature point to perform feature matching.

[0033] It should be explained that the camera captures multiple images of the workpiece, which may have scale differences due to the distance or angle of the workpiece. Through the Gaussian difference pyramid, the workpiece image is Gaussian blurred multiple times, and an image pyramid is generated at different blur levels. The blur level of these images starts from the original image and gradually increases, so that the system can detect stable feature points at different scales. In the image of each scale, local extreme points are found by comparing each pixel with its neighboring pixels. Local extreme points refer to points that are larger or smaller than the pixel values ​​around them. These local extreme points are considered candidate feature points because they have good stability at different scales. After the candidate feature points are detected, the gradient and direction of each pixel in the neighborhood of the feature point are calculated to form a gradient direction histogram. The histogram counts the gradient strength in each direction in the area. In order to ensure rotation invariance, the system selects the peak value in the gradient direction histogram as the main direction of the feature point. For each feature point, a local area (usually 16×16 pixels in size) is constructed with it as the center. In this local area, the gradient direction and magnitude of each pixel are calculated, and this information is combined into a feature vector, called a descriptor. Each descriptor represents the local structural information of the feature point. In the subsequent image matching process, the descriptors are compared to determine whether the feature points in the two images are the same.

[0034] Preferably, using Hough transform to preliminarily detect the contour of the workpiece image, determine the preliminary workpiece contour area, and classify the workpiece categories based on the preliminary workpiece contour area to obtain the final workpiece contour area includes the following steps: S21, converting the workpiece color image into a grayscale image, and applying Gaussian filtering to remove image noise; S22, extracting edges in the grayscale image using an edge detection algorithm; S23, setting the parameter space according to the shape of the workpiece to be detected; S24, increasing the count of all parameters corresponding to each edge point on the grayscale image in the parameter space to form a cumulative array, and searching for a local maximum value in the cumulative array; S25, obtaining a peak value from the accumulation array as a parameter of the detected workpiece shape, and using the parameter of the workpiece shape to draw a contour on the original workpiece image as a preliminary workpiece contour area; S26. Based on the preliminary workpiece contour area, the workpiece categories are divided using a division algorithm to obtain a final workpiece contour area.

[0035] It should be explained that after the camera acquires the color image of the workpiece, it first converts it into a grayscale image. The grayscale image removes the color information and retains the brightness information of the image, making the edges and shapes clearer. Subsequently, the system performs Gaussian filtering on the grayscale image to remove high-frequency noise in the image, such as surface texture, illumination changes, etc., to ensure the accuracy of subsequent contour detection. The grayscale image is processed using a classic edge detection algorithm (such as Canny edge detection) to extract the edges in the image. Canny edge detection calculates the image gradient and identifies the area with significant brightness changes in the image to form the edge information of the workpiece. In order to detect circular workpieces, the parameter space of the Hough transform is set to define the parameters of the circle (such as the center position and radius). The Hough transform converts between the image space and the parameter space, records the parameter value of each possible circle, and at each edge point, the Hough transform algorithm increases the count for each possible circular parameter value in the parameter space. As all edge points vote, the cumulative array in the parameter space forms a peak, representing the existence of a circle. In the cumulative array, the system finds the local maximum value, and the parameters corresponding to the maximum value are the center and radius of the circular workpiece. According to these parameters, a circular outline is drawn on the original image, and the preliminary outline area of ​​the workpiece is marked. After obtaining the preliminary outline of the workpiece, the workpiece is classified according to the size and shape parameters of the workpiece using a classification algorithm based on geometric features (such as K-means clustering, support vector machine, chaotic differential evolution algorithm, etc.). For example, a larger circle may be a bearing, while a smaller circle may be a washer. Finally, the category of the workpiece is determined based on the classification results, and the outline area is refined.

[0036] Preferably, based on the preliminary workpiece contour area, the workpiece categories are divided by using a division algorithm to obtain the final workpiece contour area, including the following steps: S261, setting the scaling factor and cross factor of the partitioning algorithm, and determining the maximum number of iterations and the number of chaotic searches; S262, classifying the workpiece according to the characteristics of the workpiece image and initializing candidate solutions for the workpiece contour; S263, randomly selecting a workpiece contour candidate solution from the workpiece contour candidate solutions according to each workpiece category, and generating a new variant workpiece contour candidate solution; S264, generating a new candidate solution for the workpiece contour by fusing different candidate solutions for the workpiece contour through cross factors; S265, comparing the newly generated candidate workpiece contour solution with the current optimal workpiece contour solution, and retaining the optimal workpiece contour solution; S266, repeating steps S263 to S265 until a predetermined number of iterations is reached; S267, performing a chaotic search around the best workpiece contour solution of each category, and selecting the workpiece contour with the highest fitness from the chaotic search as the final workpiece contour region.

[0037] Specifically, the partitioning algorithm is based on the improved chaotic differential evolution algorithm. In order to accelerate the global convergence speed, the Logistic chaotic mapping is introduced, and the chaotic differential evolution algorithm is realized by chaotic search in the solution space centered on the best individual using the ergodicity of chaotic motion.

[0038] It should be explained that the parameters are initialized for the partitioning algorithm. The scaling factor is used to control the degree of variation of the candidate solutions, and the crossover factor controls the degree of fusion of different candidate solutions. Data example: scaling factor = 0.5, crossover factor = 0.9, maximum number of iterations = 100, number of chaotic searches = 50. Based on the preliminary classification of the workpiece in the image, the contour candidate solution is initialized. Data example: The workpiece image is divided into three categories: circular workpieces, square workpieces, and irregular-shaped workpieces. For each category, a preliminary contour solution is extracted. For example, the initial contour of a circular workpiece can be an estimate of the radius size, a square workpiece can be an estimate of the side length, and an irregular shape is based on the edge detection algorithm, and a new workpiece contour solution is generated through mutation operation. Data example: For a circular workpiece, the initial candidate solutions with radii of 20mm, 25mm, and 30mm are selected, and the new radius solution = 25 + scaling factor × (30-20) is generated by mutation, and the new radius = 25 + 0.5 × 10 = 30 is obtained. Similarly, for square workpieces, the side length can be mutated to generate new candidate solutions, and the contour candidate solutions of different categories can be mixed by crossover operation to generate new contour candidate solutions. Data example: Use crossover factor = 0.9 to fuse different parameters from the existing contour solution with a probability of 90%. For circular workpieces, different radii can be fused, and square workpieces may fuse contour solutions of different side lengths. The fitness of the newly generated contour solution is evaluated, and the solution with the highest fitness is retained as the current optimal solution. Data example: The fitness function can be based on the matching degree between the contour solution and the actual workpiece image contour. The higher the matching degree, the higher the fitness. Assuming that the fitness of the current optimal solution is 0.85 and the fitness of the new solution is 0.88, the new solution is retained as the optimal solution, and the contour solution is gradually optimized through iteration. Data example: Repeat the mutation and crossover operations, each time generate a new contour solution and compare it with the current optimal solution. After 100 iterations, the fitness reaches 0.92. Use the chaos search algorithm to perform local search near the optimal solution to further optimize the contour solution. Data example: Perform local search near the optimal solution radius = 30, and generate contour solutions with radii of 29mm, 30mm, and 31mm through chaos search. After evaluating the fitness, it is found that 30.5mm has the highest fitness, and it is finally selected as the final workpiece contour solution.

[0039] Preferably, the formula for generating a new candidate solution of the variant workpiece contour is: ; In the formula, A i ( m +1) indicates the next iteration m +1 at the time i Candidate solutions for variant workpiece contours; G Expressed as a scaling factor, G belongs to (0, 1); C r1 ( m )、 C r2 ( m )and C r2 ( m ) are respectively represented as m A candidate solution of the workpiece contour randomly selected from the candidate solutions of the workpiece contour at the time; r 1. r 2 and r 3 are respectively represented as random variables (which are random positive integers and satisfy r 1≠ r 2≠ r 3≠ i ).

[0040] Preferably, regressing the feature points in the final workpiece contour area using the double point removal least squares method to determine the contour parameters includes the following steps: S31, obtaining a feature point data set within the final workpiece contour area; S32, using the least squares method to perform a preliminary linear regression analysis on all feature points to obtain an initial contour parameter estimate; S33, removing two feature points in the feature point data set one by one, and re-performing least squares regression using the remaining feature points; S34, calculating the difference between the contour parameters obtained by least square regression after removing two feature points each time and the initial contour parameter estimation, and recording the difference value; S35, repeating step S33 and step S34 until a preset number of iterations is reached; S36. Analyze the difference values ​​of all records, and select a set of regression results with the smallest difference values ​​as the estimation of the final profile parameters.

[0041] It should be explained that after the previous contour detection and shape segmentation, the contour area of ​​the gear has been identified and extracted. All feature points (such as edge points and key contour points) in the contour area are combined into a data set for subsequent contour fitting, and all extracted feature points are input into the least squares model for linear regression. Assuming that the contour of the gear is approximately circular or elliptical, a suitable curve model (such as circle or ellipse fitting) is used to perform preliminary regression on the data to obtain the initial contour parameter estimation. After the preliminary regression is completed, the system begins to remove feature points one by one. Each time two feature points are removed from the feature point data set, the remaining feature points are input into the least squares model again for regression. The two removed feature points may be outliers or points with a large deviation from the regression result. Each time two feature points are removed, a new regression result is obtained. The difference between the contour parameters of the regression result and the initial regression result is calculated, and the difference value is recorded. The smaller the difference value, the greater the impact of the removed feature points on the regression result, which may be noise or anomalies. The process of removing feature points, re-regressing, and calculating the difference value is repeated according to the preset number of iterations (for example, 500 times). Through multiple iterations, the system can gradually eliminate the impact of noise and outliers on the regression results, further optimize the fitting curve, and after all iterations are completed, compare the difference values ​​obtained from each regression, and select the one with the smallest difference value as the optimal regression result. The profile parameters of this result are considered to be the final accurate estimate. The profile of the gear is drawn based on these parameters. The regression result with the smallest difference value represents the most accurate fit after removing noise and outliers. By selecting the regression result with the smallest difference value, the accuracy of the profile parameters can be ensured.

[0042] Preferably, analyzing the difference values ​​of all records and selecting a set of regression results with the smallest difference values ​​as the estimation of the final profile parameters comprises the following steps: S361, obtain all historical regression result data sets, and initialize the regression result data sets and related constraints; S362, accepting a new regression result data update instruction and selecting a data processing mode according to analysis requirements; S363, if the priority of minimizing the difference value is selected, the constraint boundary of the data set is calculated by a specific formula, and if other priorities are selected, the boundary of the original data set is kept unchanged; S364, calculating the deviation of each set of regression result data and the deviation of the overall regression result data set, and selecting data points that need to be adjusted first; S365, performing regression analysis optimization on the selected data points using an optimization algorithm, and recording the difference value of each optimization; S366. Analyze the difference values ​​of all records, and select a set of regression results with the smallest difference value as the final profile parameters by comparing the calculated difference values.

[0043] It should be explained that a set of historical regression result data is obtained from past contour fitting, which includes the parameter estimates of each set of regression curves during the fitting process and the deviation from the initial contour estimate. Next, a regression result data set is initialized and relevant constraints are set, such as the maximum allowable deviation or the limit range of specific contour parameters (such as roundness, smoothness, etc.), and regression analysis is run based on the latest contour feature point data to generate new regression result data. Accept instructions from the user or automatic system to choose how to process the new data: whether to minimize the difference value as the priority goal, or to keep the original data set boundaries unchanged to retain historical references. If the difference value minimization is selected as the priority, a specific formula is used to recalculate the boundaries of the regression data set, such as by Lagrange multiplier method or other optimization algorithms, to limit the regression results to the specified boundary range. Otherwise, the original constraints are maintained to allow the regression results to be consistent with the historical data. The deviation of each regression result is analyzed step by step, especially the difference between each set of regression parameters and the overall result. By analyzing the impact of each data point on the overall regression, we can select the points that have the greatest impact on the deviation and mark them as points that need to be adjusted first. For the selected priority adjustment data points, we use optimization algorithms (such as gradient descent, Newton's method, fireworks algorithm, etc.) to re-perform regression analysis and adjust the weight or position of the data points to reduce the deviation of the overall regression. The difference values ​​after each optimization are recorded for subsequent comparison and analysis. All recorded difference values ​​are analyzed and compared to select a set of regression results with the smallest difference value as the final profile parameter estimation. These profile parameters will be used to generate the precise profile of the body parts to ensure that the processing equipment can operate according to the optimized profile.

[0044] Preferably, performing regression analysis optimization on the selected data points using an optimization algorithm and recording the difference value of each optimization includes the following steps: S3651. Randomly initialize several groups of parameters of the regression model, each group of parameters represents a potential regression model solution; S3652, calculating the fitness of each set of regression model parameters, and selecting the regression model with the minimum difference value as the core regression model; S3653, adjusting the non-core regression model parameter set, using different adjustment strategies for the core regression model parameter set, and determining the number of new parameters to be generated according to their respective difference values; S3654. Generate a new parameter set based on the current regression model parameters and perform boundary detection; S3655, applying Gaussian mutation to selected regression model parameters to generate additional parameter sets; S3656. From all current regression model parameters, select a set of parameters with the smallest difference value to continue iteration, and record the best difference value; S3657. Check whether the predetermined number of iterations has been reached. If reached, stop the iteration and record the minimum difference value as the regression result.

[0045] Specifically, the optimization algorithm is the improved fireworks algorithm. In the basic fireworks algorithm, the number of explosion sparks and the size of the explosion radius generated by an individual are completely determined by the difference in fitness values ​​between this individual and other individuals. In the later stage of algorithm optimization, when the population is close to the optimal value, the numerator of the explosion operator is close to 0, that is, the explosion radius tends to 0, then the subsequent optimization process is basically invalid, and its local optimization ability has certain limitations. For this reason, some improvements have been made to the basic fireworks algorithm, that is, the generation of the optimal individual explosion radius in the fireworks population is dynamically changed, not just based on the fitness value change. Compared with the basic fireworks algorithm, the improved fireworks algorithm has greater local exploration ability, because its optimization radius is dynamically changed and gradually reduced, so the improved algorithm does not have the problem of optimization invalidation, and its local optimization accuracy is greatly improved. In load distribution, it can allocate the best load to each unit as much as possible.

[0046] It should be explained that first, multiple sets of regression model parameters are randomly generated, which may represent different blade profile shapes (such as different bends, curvatures, etc.). For example, 10 sets of initial parameters can be randomly generated to represent possible blade profile curves. For each set of regression model parameters, its fitness is calculated, which is usually measured by comparing the error between the regression model fitting result and the actual blade feature points. The higher the fitness, the smaller the error. The regression model with the highest fitness (the smallest difference value) is selected as the core regression model, and all non-core regression model parameter sets are adjusted, for example, by generating new parameters through simulated annealing, genetic algorithm crossover, mutation, etc. Different strategies are used for the core regression model parameters, for example, they remain unchanged or make a small adjustment. According to the size of the difference value, a new number of parameters is generated to ensure more exploration of the optimal parameter set. Based on the adjusted parameters, a new parameter set is generated, and it is checked whether these parameters are within the preset physical and engineering constraints (for example, the curvature of the blade cannot exceed the limit allowed by the material). If the parameters exceed the boundary, they are readjusted to a reasonable range, and Gaussian mutation is applied to the currently selected regression model parameters. A new set of parameters is generated by adding random small changes (Gaussian noise) around the parameters. These new parameter sets further enrich the diversity of model solutions. All current parameter sets are analyzed, and the set with the smallest difference value is selected as the starting point of a new round of iterations, and the best difference value in this iteration is recorded. This set of parameters represents the fitting result closest to the actual blade profile, and checks whether the current number of iterations reaches the predetermined limit (for example, 100 iterations). If the limit is reached, the iteration is stopped, and the recorded minimum difference value and its corresponding regression model parameters are used as the final regression result to determine the blade profile parameters.

[0047] Preferably, obtaining the three-dimensional point cloud and the image features of the workpiece, estimating the pose vector of the workpiece by using a deep convolutional neural network, and dynamically positioning and tracking the workpiece by combining the pose vector of the workpiece with the three-dimensional point cloud of the workpiece comprises the following steps: S41. Extract features from workpiece images using a deep residual network and design a pose estimation network. S42, transforming the output of the feature extraction part through a fully connected layer to regress a pose vector; S43, dividing the output of the posture regression part into a position vector and a posture vector, as outputs of the estimated postures of the multiple classes of objects; S44, collecting the three-dimensional point cloud and image data of the workpiece in real time in a dynamic environment, and using the trained deep residual network to continuously update the pose estimation of the workpiece; S45. Combine the real-time updated pose vector and 3D point cloud to implement accurate dynamic positioning and tracking of workpieces.

[0048] It should be explained that the image of the workpiece is collected in real time by the camera and passed to the deep residual network for feature extraction. The deep residual network is an improved convolutional neural network. It solves the gradient vanishing problem in the deep network by introducing residual connections, so that it can extract more refined and deeper image features. Based on these features, a pose estimation network is designed to predict the pose of the workpiece. The workpiece image features extracted from the deep residual network are transformed through several layers of fully connected layers, and finally the pose vector of the workpiece is regressed. The pose vector usually includes six degrees of freedom of position and attitude (rotation angle). This regression task requires optimizing the loss function so that the error between the predicted pose vector and the actual pose of the workpiece is minimized, and the output of the regression network is divided into two parts: position vector and attitude vector. The position vector describes the coordinates of the workpiece in three-dimensional space, while the attitude vector describes the rotation of the workpiece (represented by Euler angles or quaternions). For various types of workpieces, the network will output the pose estimates of these workpieces according to the input workpiece image features, and obtain the three-dimensional point cloud data of the workpiece in real time through a laser scanner or a structured light camera, and collect the two-dimensional image data of the workpiece at the same time. The trained deep residual network continuously receives new image data and continuously updates the pose estimation of the workpiece in a dynamic environment. By fusing 3D point cloud information and image features, the accuracy of pose estimation can be improved, especially in complex scenes, by combining the real-time updated pose vector with 3D point cloud data to implement dynamic positioning and tracking. Based on the current pose estimation, the position and posture of the workpiece at the next moment can be predicted, and the robot's operation on the workpiece (such as grasping or placing) can be adjusted. In complex production environments, it can cope with problems such as movement, rotation or occlusion of the workpiece to achieve high-precision positioning and tracking.

[0049] According to another embodiment of the present invention, Figure 2 As shown, a variable-scale moving target positioning and tracking device integrating multiple features is also provided, and the variable-scale moving target positioning and tracking device integrating multiple features includes: Image processing module 1, used to obtain the image of the workpiece to be marked, extract features from the workpiece image using a multi-scale feature fusion algorithm, obtain feature points of the workpiece image, and establish a feature point matching relationship between adjacent workpiece image frames; Workpiece detection module 2, used for performing preliminary detection on the contour of the workpiece image by using Hough transform, determining the preliminary workpiece contour area, and classifying the workpiece categories based on the preliminary workpiece contour area to obtain the final workpiece contour area; Data analysis and regression model module 3, used to regress the feature points in the final workpiece contour area using the double point removal least squares method to determine the contour parameters; The target positioning and tracking module 4 is used to obtain the three-dimensional point cloud and image features of the workpiece, estimate the pose vector of the workpiece by using a deep convolutional neural network, and dynamically position and track the workpiece by combining the pose vector of the workpiece with the three-dimensional point cloud of the workpiece; The image processing module 1 is connected via the workpiece detection module 2 and the data analysis and regression model module 3 , and the data analysis and regression model module 3 is connected with the target positioning and tracking module 4 .

[0050] In summary, with the help of the above technical solution of the present invention, the present invention can capture the details and structure of the image at different scale levels through a multi-scale feature fusion algorithm. Such a method is extremely effective for detecting feature points that remain stable under scale transformation. This capability is particularly suitable for scenarios where the size, shape or orientation of the workpiece changes rapidly. The descriptor constructed by calculating the gradient and direction of the pixels around each feature point not only provides the location information of the point, but also accurately describes its local image features. The distance and similarity between the descriptors are used to match the feature points, making the matching process more robust, especially when the image is slightly distorted or occluded due to the movement of the workpiece. It can then effectively connect the same workpiece in consecutive image frames to provide a stable input for dynamic tracking. The present invention enhances the accurate recognition of the workpiece contour by combining multi-step edge detection and shape parameter extraction using the Hough transform, especially when the image quality is disturbed by noise. Grayscale conversion and Gaussian filtering help reduce processing errors and ensure the accuracy of edge detection. The partitioning algorithm used in the classification process further optimizes the category recognition of workpieces, thereby ensuring that each type of workpiece can be correctly classified. The chaotic search technology is used to optimize the workpiece contour after the classification is determined, which not only improves the accuracy of tracking, but also increases the adaptability to dynamic changes. By integrating different solutions and adjusting algorithm parameters such as scaling factors and cross factors, it can flexibly respond to the diversity and complexity of the visual presentation of workpieces, thereby maintaining efficient performance in a changing industrial environment. The present invention uses the least squares method for preliminary regression to provide a robust baseline, and by removing potential abnormal feature points one by one and recalculating the regression, the influence of noise and outliers can be significantly reduced, and the accuracy and robustness of parameter estimation can be improved. By continuously analyzing the difference value and selecting the regression result with the smallest difference value, the model is guaranteed to perform optimally when facing variable and complex data. This method is particularly suitable for dynamically changing industrial environments, ensuring that the estimation of contour parameters can adapt to actual physical changes. By finely managing each data point and iterative process, it can more effectively process large amounts of data, reduce the waste of computing resources, and thus optimize operating efficiency and output quality. The present invention extracts high-quality features from the workpiece image through a deep residual network, and accurately calculates the pose vector through a pose estimation network, thereby achieving efficient dynamic tracking. In a dynamic environment, it can collect and process three-dimensional point clouds and image data in real time, and continuously update the pose estimate of the workpiece. This not only improves the system's ability to adapt to environmental changes, but also ensures that the continuity and accuracy of tracking can be maintained under various working conditions.

[0051] 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 principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A variable-scale moving target positioning and tracking method integrating multiple features, characterized in that: The variable-scale moving target positioning and tracking method comprises the following steps: S1. Acquire an image of the workpiece to be identified, extract features from the workpiece image using a multi-scale feature fusion algorithm, obtain feature points of the workpiece image, and establish a feature point matching relationship between adjacent workpiece image frames; S2. Performing preliminary detection on the contour of the workpiece image using Hough transform, determining a preliminary workpiece contour area, and classifying the workpiece categories based on the preliminary workpiece contour area to obtain a final workpiece contour area; S3, using the double point removal least squares method to regress the feature points in the final workpiece contour area to determine the contour parameters; S4, obtaining the three-dimensional point cloud and image features of the workpiece, estimating the pose vector of the workpiece using a deep convolutional neural network, and dynamically positioning and tracking the workpiece by combining the pose vector of the workpiece with the three-dimensional point cloud of the workpiece; The method of acquiring an image of the workpiece to be identified, extracting features from the workpiece image using a multi-scale feature fusion algorithm, obtaining feature points of the workpiece image, and establishing a feature point matching relationship between adjacent workpiece image frames includes the following steps: S11, extracting feature points from the acquired workpiece image using a multi-scale feature fusion algorithm; S12, extracting stable feature points from the workpiece image, and calculating a descriptor of each feature point to characterize image information around the feature point; S13, matching feature points in two adjacent workpiece images, and realizing feature point matching relationship by calculating the distance and similarity between two descriptors.

2. The variable-scale moving target positioning and tracking method integrating multiple features according to claim 1 is characterized in that: The method of extracting feature points from the acquired workpiece image using a multi-scale feature fusion algorithm comprises the following steps: S111, performing multi-scale transformation on the workpiece image by using a Gaussian difference pyramid, and generating workpiece images in different scale spaces; S112, detecting local extreme points as candidate feature points by comparing adjacent pixels in each scale space workpiece image; S113, calculating the gradient and direction of the surrounding pixels for each candidate feature point, generating a gradient direction histogram, and determining the main direction of the feature point with the peak value of the histogram; S114, constructing a local area with each feature point as the center, and calculating the gradient direction and amplitude in the local area, and combining them into a descriptor vector as a descriptor of the feature point to perform feature matching.

3. The variable-scale moving target positioning and tracking method integrating multiple features according to claim 1 is characterized in that: The method of using Hough transform to preliminarily detect the contour of the workpiece image, determine the preliminary workpiece contour area, and classify the workpiece categories based on the preliminary workpiece contour area to obtain the final workpiece contour area includes the following steps: S21, converting the workpiece color image into a grayscale image, and applying Gaussian filtering to remove image noise; S22, extracting edges in the grayscale image using an edge detection algorithm; S23, setting the parameter space according to the shape of the workpiece to be detected; S24, increasing the count of all parameters corresponding to each edge point on the grayscale image in the parameter space to form a cumulative array, and searching for a local maximum value in the cumulative array; S25, obtaining a peak value from the accumulation array as a parameter of the detected workpiece shape, and using the parameter of the workpiece shape to draw a contour on the original workpiece image as a preliminary workpiece contour area; S26. Based on the preliminary workpiece contour area, the workpiece categories are divided using a division algorithm to obtain a final workpiece contour area.

4. The variable-scale moving target positioning and tracking method integrating multiple features according to claim 3 is characterized in that: The method of dividing the workpiece categories by using a division algorithm based on the preliminary workpiece contour area to obtain the final workpiece contour area includes the following steps: S261, setting the scaling factor and cross factor of the partitioning algorithm, and determining the maximum number of iterations and the number of chaotic searches; S262, classifying the workpiece according to the characteristics of the workpiece image and initializing candidate solutions for the workpiece contour; S263, randomly selecting a workpiece contour candidate solution from the workpiece contour candidate solutions according to each workpiece category, and generating a new variant workpiece contour candidate solution; S264, generating a new candidate solution for the workpiece contour by fusing different candidate solutions for the workpiece contour through cross factors; S265, comparing the newly generated candidate workpiece contour solution with the current optimal workpiece contour solution, and retaining the optimal workpiece contour solution; S266, repeating steps S263 to S265 until a predetermined number of iterations is reached; S267, performing a chaotic search around the best workpiece contour solution of each category, and selecting the workpiece contour with the highest fitness from the chaotic search as the final workpiece contour region.

5. The variable-scale moving target positioning and tracking method integrating multiple features according to claim 4 is characterized in that: The formula for generating a new candidate solution for the variant workpiece contour is: ； In the formula, A i ( m +1) indicates the next iteration m +1 at the time i Candidate solutions for variant workpiece contours; G Expressed as a scaling factor; C r1 ( m )、 C r2 ( m )and C r2 ( m ) are respectively represented as m A candidate solution of the workpiece contour randomly selected from the candidate solutions of the workpiece contour at the time; r 1. r 2 and r 3 are random variables.

6. The variable-scale moving target positioning and tracking method integrating multiple features according to claim 1 is characterized in that: The method of regressing the feature points in the final workpiece contour area using the double point removal least squares method to determine the contour parameters comprises the following steps: S31, obtaining a feature point data set within the final workpiece contour area; S32, using the least squares method to perform a preliminary linear regression analysis on all feature points to obtain an initial contour parameter estimate; S33, removing two feature points in the feature point data set one by one, and re-performing least squares regression using the remaining feature points; S34, calculating the difference between the contour parameters obtained by least square regression after removing two feature points each time and the initial contour parameter estimation, and recording the difference value; S35, repeating step S33 and step S34 until a preset number of iterations is reached; S36. Analyze the difference values ​​of all records, and select a set of regression results with the smallest difference values ​​as the estimation of the final profile parameters.

7. The variable-scale moving target positioning and tracking method integrating multiple features according to claim 6 is characterized in that: The step of analyzing the difference values ​​of all records and selecting a set of regression results with the smallest difference values ​​as the estimation of the final profile parameters comprises the following steps: S361, obtain all historical regression result data sets, and initialize the regression result data sets and related constraints; S362, accepting a new regression result data update instruction and selecting a data processing mode according to analysis requirements; S363, if the priority of minimizing the difference value is selected, the constraint boundary of the data set is calculated by a specific formula, and if other priorities are selected, the boundary of the original data set is kept unchanged; S364, calculating the deviation of each set of regression result data and the deviation of the overall regression result data set, and selecting data points that need to be adjusted first; S365, performing regression analysis optimization on the selected data points using an optimization algorithm, and recording the difference value of each optimization; S366. Analyze the difference values ​​of all records, and select a set of regression results with the smallest difference value as the final profile parameters by comparing the calculated difference values.

8. The variable-scale moving target positioning and tracking method integrating multiple features according to claim 7 is characterized in that: The method of performing regression analysis optimization on the selected data points using an optimization algorithm and recording the difference value of each optimization includes the following steps: S3651. Randomly initialize several groups of parameters of the regression model, each group of parameters represents a potential regression model solution; S3652, calculating the fitness of each set of regression model parameters, and selecting the regression model with the minimum difference value as the core regression model; S3653, adjusting the non-core regression model parameter set, using different adjustment strategies for the core regression model parameter set, and determining the number of new parameters to be generated according to their respective difference values; S3654. Generate a new parameter set based on the current regression model parameters and perform boundary detection; S3655, applying Gaussian mutation to selected regression model parameters to generate additional parameter sets; S3656. From all current regression model parameters, select a set of parameters with the smallest difference value to continue iteration, and record the best difference value; S3657. Check whether the predetermined number of iterations has been reached. If reached, stop the iteration and record the minimum difference value as the regression result.

9. The variable-scale moving target positioning and tracking method integrating multiple features according to claim 1 is characterized in that: The method of acquiring the three-dimensional point cloud and the image features of the workpiece, estimating the pose vector of the workpiece by using a deep convolutional neural network, and dynamically positioning and tracking the workpiece by combining the pose vector of the workpiece with the three-dimensional point cloud of the workpiece includes the following steps: S41. Extract features from workpiece images using a deep residual network and design a pose estimation network. S42, transforming the output of the feature extraction part through a fully connected layer to regress a pose vector; S43, dividing the output of the posture regression part into a position vector and a posture vector, as outputs of the estimated postures of the multiple classes of objects; S44, collecting the three-dimensional point cloud and image data of the workpiece in real time in a dynamic environment, and using the trained deep residual network to continuously update the pose estimation of the workpiece; S45. Combine the real-time updated pose vector and 3D point cloud to implement accurate dynamic positioning and tracking of workpieces.

10. A multi-feature integrated variable-scale moving target positioning and tracking device, used to implement the multi-feature integrated variable-scale moving target positioning and tracking method according to any one of claims 1 to 9, characterized in that: The variable-scale moving target positioning and tracking equipment integrating multiple features includes: An image processing module is used to obtain an image of the workpiece to be marked, extract features from the workpiece image using a multi-scale feature fusion algorithm, obtain feature points of the workpiece image, and establish a feature point matching relationship between adjacent workpiece image frames; The workpiece detection module is used to perform preliminary detection on the contour of the workpiece image using Hough transform, determine the preliminary workpiece contour area, and classify the workpiece categories based on the preliminary workpiece contour area to obtain the final workpiece contour area; The data analysis and regression model module is used to regress the feature points in the final workpiece contour area using the double point removal least squares method to determine the contour parameters; The target positioning and tracking module is used to obtain the three-dimensional point cloud and image features of the workpiece, estimate the workpiece's pose vector using a deep convolutional neural network, and dynamically position and track the workpiece by combining the workpiece's pose vector with the workpiece's three-dimensional point cloud; Among them, the image processing module is connected through the workpiece detection module and the data analysis and regression model module, and the data analysis and regression model module is connected with the target positioning and tracking module.

Images