Object orientation detection method and device, computer equipment and storage medium

Abstract

The invention relates to an object orientation detection method and device, computer equipment and a storage medium, which are used for solving the problem of poor accuracy and reliability of object pose adjustment in an existing industrial assembly line. The method comprises the following steps: carrying out image acquisition on a target object through a super-resolution full-spectrum imaging device to obtain a plurality of original image sets; carrying out image noise removal, enhancement and edge detection on the image set to obtain a preprocessed image set; inputting the preprocessed image set into a sparse self-encoding convolutional neural network for feature extraction to obtain object orientation feature data; identifying key feature points of the target object by using improved generalized Hough transform; and estimating the orientation of the object based on a fuzzy logic orientation discrimination algorithm, and generating an operation strategy. Through deep learning feature extraction and fuzzy logic orientation discrimination, accurate detection of the object orientation in a complex industrial environment can be realized, and the automation level and stability of assembly line work are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method, device, computer equipment and storage medium for detecting the orientation of an object. Background Art

[0002] In the development of industrial automation and intelligent manufacturing, machine vision systems are being used more and more widely, among which accurate detection of object orientation plays a key role in object grasping and refined operations. However, existing visual inspection technologies still face many technical bottlenecks when dealing with complex environments, mainly in terms of noise interference, lighting changes, occlusion, and complex backgrounds, which greatly reduces the accuracy of traditional methods in feature extraction and orientation estimation. In actual industrial scenarios, objects vary in shape, material, and color. Especially for objects with blurred edges or weak surface textures, how to achieve accurate and fast orientation detection has become an urgent problem to be solved.

[0003] Traditional image processing and machine vision methods usually rely on simple edge detection and feature matching algorithms. These methods can achieve certain results in a single and ideal environment, but they are powerless in the face of the complexity and variability of industrial environments. In many cases, the target object may be affected by light from different angles or partially blocked by other objects. These factors will cause the loss of visual information and cannot guarantee the stability and reliability of feature points.

[0004] In summary: When adjusting the position and posture of objects on existing industrial assembly lines, it is often difficult to stably extract the key feature points of the target object, which leads to technical problems such as insufficient adjustment accuracy and reliability. There is currently no effective solution. Summary of the invention

[0005] The main purpose of the present invention is to provide a method, device, computer equipment and storage medium for detecting the orientation of an object, so as to solve the technical problem of poor accuracy and reliability when adjusting the position of objects in existing industrial assembly lines.

[0006] To achieve the above-mentioned purpose, the present invention provides a method for detecting the orientation of an object, comprising the following steps: acquiring images of a target object by means of a plurality of super-resolution full-spectrum imaging devices arranged in the path of the target object to obtain a plurality of original image sets; performing image preprocessing on the original image sets, including image noise removal, image enhancement based on anisotropic diffusion, and edge detection to obtain a preprocessed image set; inputting the preprocessed image set into a preset sparse autoencoder convolutional neural network model for feature extraction to obtain object orientation feature data; based on the object orientation feature data, identifying key feature points of the target object using an improved generalized Hough transform to obtain feature point position information; estimating the orientation of the target object using the feature point position information and an orientation discrimination algorithm based on fuzzy logic to obtain object orientation estimation data; inputting the object orientation estimation data into an object operation module to generate an operation strategy for object grasping or processing according to the object orientation estimation data.

[0007] The present invention also provides a device for detecting the orientation of an object, comprising: an acquisition unit, used to acquire images of a target object through a plurality of super-resolution full-spectrum imaging devices arranged in a path of the target object, so as to obtain a plurality of original image sets; a preprocessing unit, used to perform image preprocessing on the original image set, including image noise removal, image enhancement based on anisotropic diffusion, and edge detection, so as to obtain a preprocessed image set; a feature extraction unit, used to input the preprocessed image set into a preset sparse autoencoder convolutional neural network model for feature extraction, so as to obtain object orientation feature data; an identification unit, used to identify key feature points of the target object based on the object orientation feature data by using an improved generalized Hough transform, so as to obtain feature point position information; an orientation discrimination unit, used to estimate the orientation of the target object by using the feature point position information and an orientation discrimination algorithm based on fuzzy logic, so as to obtain object orientation estimation data; a generation unit, used to input the object orientation estimation data into an object operation module, and generate an operation strategy for object grasping or processing according to the object orientation estimation data.

[0008] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0009] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0010] The object orientation detection method, device, computer equipment and storage medium provided by the present invention extract features from images by using a sparse autoencoder convolutional neural network model, making full use of sparsity constraints and deep learning capabilities to extract effective orientation features of objects; and combining an improved generalized Hough transform with a fuzzy logic orientation discrimination algorithm to identify key feature points of an object and estimate its orientation, thereby achieving high-precision orientation estimation and operation strategy generation. In short, the present invention effectively solves the accuracy and reliability problems of object posture adjustment in industrial assembly lines, and significantly improves the stability and efficiency of the automation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 1 is a schematic diagram of the steps of a method for detecting the orientation of an object in one embodiment of the present invention; Figure 2 is a structural block diagram of a device for detecting the orientation of an object in one embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0012] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0014] Reference Figure 1 , an embodiment of the present invention provides a method for detecting the orientation of an object, comprising: S1, collecting images of a target object by using a plurality of super-resolution full-spectrum imaging devices arranged in a path of the target object to obtain a plurality of original image sets.

[0015] That is, the target object is imaged by multiple super-resolution full-spectrum imaging devices arranged in the path of the target object to obtain multiple sets of original images. During the implementation process, each super-resolution full-spectrum imaging device has the multi-band acquisition capability of visible light and infrared spectrum, and can obtain rich detail information of the target object under different lighting conditions. These imaging devices are arranged at different positions of the target object's path to perform multi-view imaging of the target object at different angles. For example, several imaging devices are installed at different positions of the assembly line to ensure that the front, side and back of the object are fully captured. Compared with existing ordinary visible light cameras, super-resolution full-spectrum imaging can improve the spatial resolution and spectral detail expression of the image, provide high-quality data input for subsequent image processing and feature extraction, and effectively solve the problem of poor image quality caused by uneven lighting, occlusion or changes in object material by acquiring multi-view and multi-band images.

[0016] S2, performing image preprocessing on the original image set, including image noise removal, image enhancement based on anisotropic diffusion and edge detection, to obtain a preprocessed image set.

[0017] That is, by performing image preprocessing on the original image set, including image noise removal, image enhancement based on anisotropic diffusion and edge detection, a preprocessed image set is obtained. In the image preprocessing, noise removal is first performed by a composite bilateral filtering method to reduce the influence of ambient light interference and noise caused by the image sensor. Composite bilateral filtering combines spatial similarity and color similarity, and can effectively remove noise while maintaining the edge clarity of the object. Next, anisotropic diffusion is used for image enhancement, and by calculating the image gradient and diffusion coefficient, the diffusion rate of the edge area is low and the diffusion rate of the background area is high, thereby enhancing the edge features of the object and making the edge information more significant. Finally, an edge detection method based on gradient field divergence analysis is used to process the image, and the fine edge of the target object is extracted by detecting the gradient change rate. Through the above multiple preprocessing steps, the present invention effectively improves the image quality in the feature extraction stage, and compared with the existing simple filtering and edge detection methods, it significantly improves the integrity and accuracy of the image edge information, providing high-quality input for subsequent feature extraction.

[0018] S3, inputting the preprocessed image set into a preset sparse autoencoder convolutional neural network model for feature extraction to obtain object orientation feature data.

[0019] That is, the preprocessed image set is input into the preset sparse autoencoder convolutional neural network model for feature extraction to obtain the object orientation feature data. The sparse autoencoder convolutional neural network model adaptively extracts image features through an encoding-decoding structure, in which sparsity constraints are used to suppress redundant information and retain only effective features that have the ability to discriminate the object orientation. In the specific implementation, the image is first efficiently represented by a sparse coding layer, and then input into a convolutional layer for spatial feature extraction, and finally the feature dimension is compressed through a pooling layer to obtain a low-dimensional but discriminative feature vector. The sparse autoencoder network can effectively extract features such as the edge and surface texture of the object to generate a data set for orientation discrimination. Compared with traditional convolutional neural networks, the sparse autoencoder model reduces the learning of a large number of irrelevant features through sparse constraints, thereby improving the efficiency and accuracy of feature extraction.

[0020] S4, based on the object orientation feature data, using improved generalized Hough transform to identify key feature points of the target object to obtain feature point position information.

[0021] That is, based on the object orientation feature data, the improved generalized Hough transform is used to identify the key feature points of the target object and obtain the location information of the feature points. The improved generalized Hough transform combines the quantization of the gradient direction and distance parameters to make the voting process of the feature points more accurate, especially in the case of complex backgrounds or partial occlusion of the target object, which can effectively improve the recognition ability of key feature points. Specifically, the improved Hough transform analyzes the gradient information of the feature points and combines it with a weighted voting mechanism to obtain a voting accumulation array, and then finds the key feature points of the target object through local extreme value detection. Compared with the existing feature point detection method based on edge matching, this method shows higher robustness and recognition rate when dealing with multiple interference factors in complex scenes.

[0022] S5, using the feature point position information and an orientation discrimination algorithm based on fuzzy logic to estimate the orientation of the target object to obtain object orientation estimation data.

[0023] That is, the orientation of the target object is estimated using the feature point position information and the orientation discrimination algorithm based on fuzzy logic to obtain the object orientation estimation data. Specifically, the present invention converts the position data of the feature points into fuzzy input variables through a fuzzy logic system, and constructs a series of fuzzy rules for orientation discrimination. In actual operation, the feature point position information is fuzzified by a fuzzy membership function, and then the object orientation is inferred by a fuzzy inference engine combined with a preset fuzzy rule library to obtain the fuzzy orientation data of the target object. The fuzzy inference result is then defuzzified by the centroid method to obtain the specific orientation estimation data of the object. Compared with traditional geometric calculation methods, fuzzy logic shows better robustness and flexibility when dealing with uncertainty factors such as measurement errors and illumination changes.

[0024] S6, inputting the object position estimation data into an object operation module, and generating an operation strategy for object grasping or processing according to the object position estimation data.

[0025] Finally, the object position estimation data is input into the motion trajectory planning unit to generate a preliminary grasping trajectory; the path is then refined through cubic spline interpolation to ensure the smoothness of the trajectory and the accuracy of the operation. Finally, an adaptive operation strategy is generated based on the optimized grasping trajectory and object position estimation data, and the operation process is decomposed into multiple independently executed action steps to construct a complete operation instruction sequence, so that the robot or robotic arm can accurately and efficiently perform the task of grasping or processing the target object.

[0026] In summary, the present invention achieves high-precision position detection of target objects in complex industrial environments and generates stable operation strategies through the above-mentioned technical means, thus solving the shortcomings of the prior art in position detection and grasping strategy generation in complex environments.

[0027] In one example, the original image set is subjected to image preprocessing, including image noise removal, image enhancement based on anisotropic diffusion, and edge detection, to obtain a preprocessed image set, including: calculating the spatial distance and color similarity between the pixel point in each image in the original image set and the neighboring pixel points, respectively calculating the weighting coefficient of each neighboring pixel by means of a spatial Gaussian kernel function and a color Gaussian kernel function, obtaining a denoised pixel value according to the weighting coefficient of each pixel, and obtaining a preliminary noise removal image set; calculating the Laplace gradient value of each image in the preliminary noise removal image set, constructing an anisotropic diffusion coefficient matrix, and iterating and calculating each pixel one by one according to the diffusion coefficient matrix to obtain an image gradient enhanced image set; calculating the gradient direction of each pixel in each image in the image gradient enhanced image set to obtain a gradient direction distribution matrix, and Perform local extreme value detection on the gradient direction distribution matrix to extract the edge direction, and obtain an edge direction image set; calculate the second-order structure tensor matrix of each image in the edge direction image set, and perform segmentation processing on the edge points in each image according to the structure tensor matrix to obtain the edge feature distribution of the image, and obtain an edge feature image set; calculate the gradient divergence of the edge feature points of each image in the edge feature image set, construct a divergence filter to perform filtering operations on the edge feature points, and connect the edge feature points processed by the divergence filter to obtain a refined edge image; mark the edge feature points and non-edge areas in each image in the refined edge image, expand and shrink the edge areas through morphological dilation and erosion operations, and color the edge areas and non-edge areas in the image respectively to obtain a final preprocessed image set.

[0028] That is, the first step of image preprocessing is to remove noise from each image in the original image set. Specifically, by calculating the spatial distance and color similarity between each pixel and its neighboring pixels, the weighting coefficients of each neighboring pixel are calculated using the spatial Gaussian kernel function and the color Gaussian kernel function, respectively, and the denoised pixel value is calculated based on these weighting coefficients. This method not only considers the spatial distance from the central pixel, but also measures the color similarity of neighboring pixels through the color Gaussian kernel function to avoid excessive blurring in the edge area. For example, if the target object is a part with complex colors, the spatial similarity of the neighboring pixels can effectively maintain the edge details of the part, while the color similarity can help highlight the color features and ensure that key edge information is not lost during the denoising process. In this way, the generation of the preliminary noise removal image set can significantly reduce the impact of environmental noise, reduce the noise interference caused by the image sensor, and significantly improve the image quality.

[0029] Next, based on the denoised preliminary image, the image is gradient enhanced using anisotropic diffusion technology. Specifically, the anisotropic diffusion coefficient matrix is ​​constructed by calculating the Laplace gradient value of each image, and each pixel is iteratively calculated according to the diffusion coefficient matrix to achieve image enhancement. The core of anisotropic diffusion is that it can effectively distinguish edge areas from non-edge areas: the diffusion rate in the edge area is low, which can maintain the clarity of the edge, while the diffusion rate in the flat area is high, which can effectively reduce noise. For example, for a gear with a complex contour, the Laplace gradient value can help locate the edge of the gear, while anisotropic diffusion makes the edge of the entire image more prominent by retaining the edge and suppressing background noise. The image gradient enhanced image set obtained in this way not only improves the saliency of the edge of the target object, but also ensures further suppression of noise, providing a good foundation for edge detection.

[0030] After image enhancement, the next step is to calculate the gradient direction of each pixel in each image in the image gradient enhancement image set, obtain the gradient direction distribution matrix, and perform local extremum detection on the gradient direction distribution matrix to extract the edge direction, thereby obtaining an edge direction image set. The purpose of this step is to further identify and accurately extract the edge direction information in the image. By calculating the gradient direction distribution, the change trend of each pixel in the image can be obtained, so as to locate and extract the edge. Local extremum detection finds those significant edge directions by judging the gradient direction, which is crucial for extracting the contour of the object and determining the orientation of the object. For example, when detecting industrial parts, the change in gradient direction can reflect the edge direction of the part, and the detection of local extremum can help find the sharp turning point of the contour, which is very important for accurately locating the position and shape of the object.

[0031] In order to further extract more detailed edge information, the second-order structure tensor matrix is ​​calculated for the edge direction image set, and the edge points in the image are segmented according to the structure tensor matrix to obtain the edge feature image set. The second-order structure tensor is used to describe the local gradient characteristics in the image, and can accurately capture the direction and intensity characteristics of the edge. In this step, by analyzing the neighborhood of each pixel point, a structure tensor matrix can be constructed, and the edge points of the image can be segmented using this matrix. For example, in a scene composed of multiple intersecting lines, the second-order structure tensor can effectively distinguish the edges of different lines, thereby ensuring that the features of each line can be correctly segmented and marked.

[0032] Then, the gradient divergence of the edge feature points of each image in the edge feature image set is calculated, a divergence filter is constructed to filter the edge feature points, and the edge feature points processed by the divergence filter are connected to obtain a refined edge image. The calculation of gradient divergence is mainly used to further refine the features of edge points. By constructing a divergence filter, edge points with pseudo edges or noise interference can be effectively removed. The divergence filter analyzes the gradient changes of edge points to ensure that only those edge points with significant gradient changes are retained, and then connects these points to form a complete edge contour. For example, when the target object is a plate with many complex textures on the surface, gradient divergence analysis can help remove unnecessary details and retain only the main edge information of the plate, which makes the boundary of the object clearer.

[0033] Finally, the edge feature points and non-edge areas of each image in the refined edge image are marked, the edge area is expanded and contracted by morphological dilation and erosion operations, and the edge area and non-edge area in the image are colored respectively to obtain the final preprocessed image set. Morphological dilation and erosion operations are key steps for further processing of edges. The dilation operation can fill the small breaks in the edge to make the edge more coherent, while the erosion operation can remove the redundant parts introduced during dilation, so that the edge refinement processing is more accurate. When marking and coloring the edge and non-edge area, the outline of the object can be intuitively distinguished from the background. For example, for objects on the assembly line, by coloring the edge of the object, it can be clearly distinguished from other areas of the background, which is helpful for subsequent feature extraction and orientation detection.

[0034] Through the above steps, this technical solution realizes an effective image preprocessing process before object orientation detection. Compared with the prior art, this solution adopts a variety of complex image processing methods such as composite bilateral filtering, anisotropic diffusion, structural tensor calculation and gradient divergence, which significantly improves the quality of edge feature extraction of target objects, making the final edge information more accurate and complete, and providing reliable data support for subsequent orientation estimation and operation strategy generation. Through this multi-step image preprocessing, not only the signal-to-noise ratio of the image is improved, but also the significance of the object edge is enhanced, so that high-precision detection and operation of target objects can be achieved in complex industrial environments.

[0035] In one example, the preprocessed image set is input into a preset sparse autoencoder convolutional neural network model for feature extraction to obtain object orientation feature data, including: performing local block division on each image in the preprocessed image set, constructing a sparse representation matrix for each image block according to a sparsity constraint, and calculating the sparse coding result of each image by an L1 regularization minimization method; inputting the sparse coding result of each image into the encoding layer of the sparse autoencoder of the sparse autoencoder convolutional neural network model, and performing a dimensionality reduction transformation on the coding result using an activation function of a hidden layer to obtain a low-dimensional feature representation of each image; constructing a convolution kernel of the sparse autoencoder convolutional neural network model, and using the convolution kernel to perform a sparse representation on each image. The low-dimensional feature representation of the image is locally weighted summed, and the convolution result is processed with nonlinear activation to obtain a convolution feature image containing the local spatial features of the image; and the convolution features of each image are locally downsampled using the maximum pooling method to retain the most significant part of the features to obtain the pooled features of each image after downsampling; the pooled features of each image are inversely encoded using a sparse autoencoder again, and the pooled features of each image are sparsely reconstructed through sparsity constraints and reconstruction error minimization conditions to obtain the reconstructed image features of each image; the reconstructed image features of each image are fully connected to map the image features to the feature space, and the object orientation feature data is calculated using an activation function.

[0036] That is, each image in the preprocessed image set is divided into local blocks, a sparse representation matrix is ​​constructed for each image block according to the sparsity constraint, and the sparse coding result of each image is calculated by the L1 regularization minimization method. Local block partitioning divides the image into multiple small blocks, so that each block can be sparsely represented separately. This processing method effectively reduces the computational complexity of the image while retaining the significance of local features. Sparse representation makes feature representation more concise and efficient by applying sparsity constraints to each image block, that is, only selecting the most significant parts among many possible features. For example, for an industrial part containing multiple threads, the sparse representation of each local block will focus on the significant features that describe the thread profile and ignore redundant flat areas, thereby effectively reducing the computational burden in subsequent processing.

[0037] Then, the sparse coding result of each image is input into the coding layer of the sparse autoencoder convolutional neural network model, and the coding result is transformed into a low-dimensional feature representation using the activation function of the hidden layer to obtain a low-dimensional feature representation of each image. The autoencoder performs adaptive dimensionality reduction on the input image through the encoding-decoding structure. The encoding layer is responsible for compressing the high-dimensional sparse coding result into a low-dimensional representation, while the activation function of the hidden layer ensures the retention of nonlinear features in this process. Through dimensionality reduction, the feature representation of each image extracts the most discriminative information from the high-dimensional feature space to form a more compact feature description. For example, for an image containing complex geometric shapes, the autoencoder can remove duplicate or unimportant information through dimensionality reduction, retaining features such as edges and corners that are important for the orientation of the object, making orientation discrimination easier.

[0038] Then, the convolution kernel of the sparse autoencoder convolutional neural network model is constructed, and the convolution kernel is used to perform local weighted summation on the low-dimensional feature representation of each image, and the convolution result is processed by nonlinear activation to obtain a convolution feature image containing the local spatial features of the image. The convolution operation constructs a convolution kernel to slide on the image, thereby extracting features with local similarity in the image. This processing method enables the network to learn local patterns in the image, such as local spatial features such as edges and corners. Nonlinear activation functions (such as ReLU) are used to activate the convolution results to make the features more diverse. For example, when detecting a gear composed of multiple teeth, the convolution operation can identify the edge of each tooth, and the nonlinear activation further enhances the saliency of the teeth, thereby providing richer input for subsequent pooling and feature extraction.

[0039] After the convolution operation, the maximum pooling method is used to locally downsample the convolution features of each image, thereby retaining the most significant part of the feature and obtaining the downsampled pooled feature. The pooling operation further reduces the dimension of the feature representation by taking the maximum value in the local area, while retaining the most distinctive part of the feature. Maximum pooling can effectively reduce the redundant information of image features and improve the robustness of the network. For example, for an image containing multiple complex edges, the most obvious part of these edges can be extracted through the pooling operation, thereby reducing the interference of detail noise on subsequent processing, so that the model focuses more on the global important features of the image.

[0040] Then, the sparse autoencoder is used again to inversely encode the pooled features of each image, and the pooled features are sparsely reconstructed through sparsity constraints and reconstruction error minimization conditions to obtain the reconstructed image features. Sparse reconstruction restores the pooled features through the inverse encoder so that the features can be restored to a form close to the original image, and combines sparsity constraints and reconstruction error minimization to ensure that the restored features remain concise and effective. For example, when sparsely reconstructing a pooled image of a mechanical part, the inverse encoder will try to restore key features (such as edges and corners) to their complete form, while the sparsity constraint ensures that only those features that are of practical significance for orientation discrimination are restored, thereby reducing unnecessary complexity.

[0041] Finally, the reconstructed image features of each image are fully connected, mapped to the feature space, and the object orientation feature data is calculated using the activation function. The fully connected operation integrates the reconstructed features and combines different features together to form a complete feature representation. The activation function performs nonlinear processing on the features, making the data in the feature space more discriminative. For example, for a workpiece with complex geometric features, the edges, corners, surface textures and other features are integrated together through the fully connected operation, and finally a comprehensive feature orientation description is formed, which can more accurately determine the orientation of the workpiece.

[0042] The entire feature extraction process mentioned above ensures efficient extraction and accurate expression of image features. Sparse coding reduces unnecessary redundant information and improves the sparsity of features; convolution and pooling extract local spatial features in the image and enhance the network's ability to learn local patterns; sparse autoencoder reconstruction ensures the integrity and simplicity of features; fully connected operations integrate different features to form complete object orientation feature data. Compared with traditional feature extraction methods, this scheme significantly improves the accuracy and effectiveness of feature extraction by introducing multi-level deep learning technology, especially in complex backgrounds and with more noise interference. It can better extract the salient features of the target object, providing strong data support for subsequent orientation estimation and automated operations.

[0043] Through these processing steps, the final object orientation feature data not only retains the most significant features in the image, but also reduces unnecessary information through dimensionality reduction, sparsity constraints, etc., making the features easier to identify and analyze. This feature extraction method can ensure that in complex industrial environments, even in the presence of unfavorable factors such as noise and occlusion, the orientation features of objects can be obtained efficiently and accurately, thus providing a reliable basis for subsequent operations.

[0044] In one example, based on the object orientation feature data, an improved generalized Hough transform is used to identify key feature points of the target object to obtain feature point position information, including: obtaining prior geometric shape information of the target object, and constructing a shape template model based on the prior geometric shape information of the target object, matching the object orientation feature data with the shape template model pixel by pixel to extract the initial feature point position that meets the template; calculating the relative position relationship between each initial feature point and the surrounding pixels, and using the gradient direction consistency and gradient amplitude difference metrics to eliminate feature points that do not meet the threshold conditions and retain candidate feature points that meet the conditions; performing gradient direction matching on each candidate feature point; The direction and distance parameters are quantified, the quantization results are mapped to the voting space, and the weighted voting mechanism is used for cumulative calculation to obtain the voting accumulation array for feature point detection; a neighborhood comparison is performed on each voting position in the voting accumulation array to find the point with the largest voting value, and it is used as the candidate point of the key feature point to obtain a preliminary set of key feature points; the geometric consistency calculation is performed on the relative position and distance of each candidate feature point, and the candidate points that do not meet the preset geometric constraints are eliminated, and the key feature points that meet the geometric consistency are retained; the coordinates of each key feature point are accurately located, and the sub-pixel interpolation method is used to adjust the feature point position with sub-pixel accuracy to obtain the final feature point position information.

[0045] That is, by obtaining the prior geometric shape information of the target object and building a shape template model based on this information. Prior geometric shape information can include the basic outline of the object, feature edges, corner point positions, etc., which is the basis for effective detection of the target object. For example, when detecting an industrial part, if it is known that the outer contour of the part is a circle or a specific polygonal structure, a shape template model can be built based on this geometric information. This template model can be used as a standard to identify the feature points of the object. In practical applications, the construction of the shape template model enables the subsequent feature point extraction to have a clearer directionality, thereby improving the accuracy of matching.

[0046] Next, the object orientation feature data is matched pixel by pixel with the constructed shape template model to extract the initial feature point positions that match the template. This step ensures that the feature points extracted from the object orientation feature data are consistent with the geometric shape of the target object. This pixel-by-pixel matching method can fully capture the distribution of the feature points of the target object. For example, when detecting a set of complex gears, by comparing the image features with the template model pixel by pixel, feature points that match the geometric shape can be extracted. These points are usually located at the edge or junction of the gear, reflecting the important geometric features of the target object.

[0047] In order to improve the reliability of feature points, after the initial feature point extraction, it is necessary to calculate the relative position relationship between each feature point and the surrounding pixels, and use the gradient direction consistency and gradient amplitude difference metrics to screen feature points. Gradient direction consistency is used to determine whether the edge features of the feature point and its surrounding neighborhood are consistent, and the gradient amplitude difference measures the significance of the feature point. Through these calculations, feature points that do not meet the threshold conditions are eliminated, and only those candidate feature points that meet the requirements are retained. For example, in a complex background, some points may meet the template matching requirements, but because the gradient direction or amplitude of the surrounding pixels is inconsistent with the feature points, they may be part of the noise or background. At this time, they can be eliminated through this process, thereby improving the accuracy of feature point detection.

[0048] For the retained candidate feature points, the gradient direction and distance parameters of each candidate point are further quantified, and the quantization results are mapped to the voting space. This quantization process allows each candidate feature point to be represented by fixed parameters, which is convenient for subsequent voting calculations. In the voting stage, the candidate feature points are accumulated and calculated using a weighted voting mechanism to generate a voting accumulation array. The weighted voting mechanism ensures that the voting for the key feature points of the target object is more accurate by weighting the importance of the candidate feature points. For example, if the gradient direction of a candidate feature point is consistent with multiple feature points of the template, it can be given a higher weight, so that the point will contribute more in the voting accumulation array. This weighted method ensures the accuracy of the voting process, especially when there is interference around the feature points, and can effectively highlight the truly key feature points.

[0049] After generating the voting accumulation array, a neighborhood comparison is performed on each voting position to find the point with the largest voting value, and it is used as a candidate point for the key feature point to obtain a preliminary set of key feature points. This process ensures that the selected feature point is the most significant in the local area by comparing the voting situation of each candidate point in its neighborhood. For example, during the detection of a polygonal part, there may be multiple similar positions that have received high voting values. Through neighborhood comparison, it can be ensured that the selected feature point is the point with the highest importance in the local area, which further improves the robustness and stability of feature point selection.

[0050] In order to ensure that the selected preliminary key feature points meet the geometric characteristics of the target object, the geometric consistency calculation is then performed on the relative position and distance of each candidate feature point. By comparing with the preset geometric constraint relationship, the candidate points that do not meet the geometric constraints are eliminated, and the key feature points that meet the geometric consistency are retained. This process analyzes the spatial relationship between the feature points to ensure that these points are consistent with the geometric structure of the target object as a whole. For example, when detecting a set of symmetrical mechanical parts, the geometric consistency calculation can ensure that the relative positions of the feature points on both sides of the symmetry axis meet a specific ratio or distance, thereby eliminating points that do not meet the symmetry characteristics.

[0051] Finally, the coordinates of each key feature point are precisely located, and the sub-pixel interpolation method is used to adjust the feature point position with sub-pixel precision to obtain the final feature point position information. Sub-pixel interpolation can obtain coordinates that are more accurate than the original pixel resolution by interpolating the grayscale values ​​of the feature points in the neighborhood. This is very important for application scenarios that require high-precision orientation estimation. For example, when an industrial robot grasps a small precision part, sub-pixel accuracy can ensure that the position accuracy of the feature point is high enough, thereby providing more accurate data support for the robot's grasping and positioning.

[0052] The above technical solution ensures that the feature point extraction process is highly robust and accurate. First, by constructing a shape template and matching pixel by pixel, the initially extracted feature points are ensured to conform to the geometric shape of the object; then, through mechanisms such as gradient consistency and weighted voting, the feature points are screened and optimized to remove background noise and interference points; then, through geometric consistency verification, the feature points are ensured to conform to the overall geometric features of the target object, and finally, sub-pixel interpolation is used to further improve the positioning accuracy. Compared with traditional feature point detection methods, this solution shows higher robustness and recognition rate when dealing with various interference factors in complex environments, so that the significant feature points of the target object can be better extracted, ensuring that reliable reference data is provided in subsequent orientation estimation and automated operations.

[0053] In one example, the orientation of the target object is estimated using the feature point position information and an orientation discrimination algorithm based on fuzzy logic to obtain object orientation estimation data, including: vectorizing the coordinates of each feature point in the feature point position information, calculating the angular offset value of the feature point relative to the image center, and obtaining preliminary azimuth angle data of each feature point; fuzzifying the angular difference between every two feature points, constructing a fuzzy membership function, mapping the angular difference to a membership value, and forming a fuzzy membership matrix, wherein the fuzzy membership matrix is ​​used to describe the spatial relationship between the feature points; constructing a set of fuzzy rule bases based on the object orientation, and using a fuzzy inference engine to map the model. The fuzzy membership matrix is ​​used for reasoning operations, and the orientation of the target object is judged in combination with the fuzzy rule base to obtain the fuzzy reasoning result of each feature point; the fuzzy set operation is performed on the fuzzy reasoning result of each feature point, and the membership of the reasoning results of different feature points is weighted summed to obtain the aggregated orientation data set of the target object as a whole; the fuzzy membership in the aggregated orientation data set is defuzzified based on the centroid method, and the fuzzy membership is mapped to a specific orientation value to obtain the orientation estimation result of the target object; the defuzzified orientation estimation result is smoothed, and the estimated values ​​of different feature points are combined using the weighted average method to obtain the final orientation estimation data of the object.

[0054] That is, by vectorizing the coordinates of each feature point in the feature point position information, the angular offset value of the feature point relative to the image center is calculated to obtain the preliminary azimuth data of each feature point. The core of vectorization processing is to convert the position information of each feature point into a relationship with the image center. This conversion method can express the position of the object as a relative orientation description, which is convenient for subsequent orientation analysis. For example, in an image, if the feature point of the target object is located to the left of the image center, the offset angle relative to the image center can be obtained through vectorization calculation. This offset angle describes the relative orientation of the feature point and is the basic data for subsequent analysis.

[0055] Next, the angle difference between each two feature points is fuzzified, a fuzzy membership function is constructed, and the angle difference is mapped to a membership value to form a fuzzy membership matrix. The purpose of fuzzification is to deal with the uncertainty between feature points. By constructing a fuzzy membership function, the angle difference between feature points is expressed as a membership in a fuzzy set. For example, for two feature points, if their angle difference is small, they can be mapped to a higher membership value through a fuzzy membership function, which means that the two points are more likely to be in the same orientation. If the angle difference is large, the membership will be lower, indicating that there is a large difference in the orientation between the two points. The fuzzy membership matrix can effectively describe the spatial relationship between feature points, especially when there is a certain measurement error or a complex background. The uncertainty between these points can be better characterized by fuzzification.

[0056] Then, a set of fuzzy rule bases based on the object orientation is constructed, and the fuzzy inference engine is used to perform inference operations on the fuzzy membership matrix. The fuzzy rule base consists of a set of rules that describe the orientation relationship of objects. These rules can be used to define the orientation relationship between feature points under different membership conditions. The fuzzy inference engine combines the fuzzy rule base to infer the input fuzzy membership matrix to determine the orientation of the target object. For example, if the membership of multiple feature points shows that most of them are located on the same side of the image center, then according to the fuzzy rule base, the inference engine will determine the orientation of the object as being offset to that side. Through this fuzzy reasoning, the fuzzy reasoning results of each feature point can be obtained, which represent the possible values ​​of the object orientation in the fuzzy set.

[0057] In order to further integrate the reasoning results of each feature point, fuzzy set operations are performed on the fuzzy reasoning results of each feature point. By weighted summing the membership of the reasoning results of different feature points, the aggregated orientation data set of the target object as a whole is obtained. The purpose of fuzzy set operations is to merge the reasoning results of each feature point to form an aggregated representation that describes the orientation of the entire object. In this process, the membership of different feature points is weighted, and the weights can be set according to the significance of the feature points or their importance in orientation discrimination. For example, if some feature points are located at the edge or corner of the object, these locations are highly representative of the orientation estimation, so they can be given higher weights when weighting to ensure that these points contribute more to the final aggregated orientation data set.

[0058] Next, the fuzzy membership in the aggregated orientation data set is defuzzified based on the centroid method, and the fuzzy membership is mapped to a specific orientation value to obtain the orientation estimation result of the target object. The centroid method is a common defuzzification method. By calculating the centroid position of the fuzzy set, the fuzzy description is converted into a specific numerical value. For example, if the aggregated orientation data set contains multiple fuzzy membership values, a numerical orientation angle can be calculated by the centroid method, which can be used to describe the specific position of the target object. The centroid method can well balance the differences between different fuzzy memberships, thereby obtaining a robust orientation estimate.

[0059] Finally, the defuzzified orientation estimation results are smoothed, and the estimated values ​​of different feature points are combined by weighted averaging to obtain the final orientation estimation data of the object. The purpose of smoothing is to reduce the fluctuation of orientation estimation caused by noise or local anomalies. By combining the orientation estimation results of multiple feature points by weighted averaging, a smoother and more stable orientation representation can be obtained. For example, if the orientation estimation of a feature point fluctuates greatly, while the orientations of other feature points are relatively consistent, the fluctuation of the feature point can be effectively suppressed by weighted averaging, ensuring that the final orientation estimation is more reliable.

[0060] The entire technical solution gradually processes the feature point position information to achieve accurate estimation of the target object's orientation. Compared with traditional geometric calculation methods, the introduction of fuzzy logic makes the entire process more robust and adaptable in the face of environmental uncertainties. For example, in complex industrial environments, the surface of the target object may have reflections, occlusions or other interferences, resulting in inaccurate feature point detection. The fuzzy logic method can effectively deal with these interference factors through fuzzy membership, fuzzy reasoning, etc., to ensure stable judgment of the object's orientation. In addition, the accuracy of orientation estimation is further improved through sub-pixel interpolation and smoothing, making this method particularly suitable for industrial application scenarios that require high-precision positioning, such as robot grasping, precision assembly, etc.

[0061] In one example, the object orientation estimation data is input into an object operation module, and an operation strategy for object grasping or processing is generated based on the object orientation estimation data, including: determining the position and orientation of the object in space based on the object orientation estimation data, and generating a preliminary operation path plan in combination with the degree of freedom parameters of the operation robot or robotic arm; performing detailed calculations on each operation point in the preliminary operation path plan, and using cubic spline interpolation to interpolate key points in the path to obtain a final operation path plan; decomposing the final operation path plan into operation steps to obtain multiple independent execution actions, and constructing an operation command sequence, and determining the optimal execution order and execution time through a path optimization algorithm to generate an operation strategy for controlling the object operation module.

[0062] Determine the position and orientation of the object in space based on the object orientation estimation data. The core of this process is to use the object orientation information obtained in the previous step to determine the specific coordinates and posture of the object in the workspace, including its position, rotation angle, orientation and other information in three-dimensional space. These data are crucial for planning the operating path of the robot arm because they determine how the robot arm reaches the object and effectively grasps it. For example, in an assembly line application, the target object may pass through the working area of ​​the robot arm at different angles and positions. Therefore, determining the position and orientation of the object based on the orientation estimation data can provide an accurate basis for path planning and ensure that the end effector of the robot arm can align with the object.

[0063] Next, the preliminary operation path planning is generated by combining the degree of freedom parameters of the operating robot or manipulator. Degrees of freedom are important parameters that describe the flexibility of the manipulator's operation, including different dimensions such as rotation, extension, and joint movement of the manipulator. Based on these degree of freedom parameters, the purpose of preliminary path planning is to design a reasonable path for the manipulator to reach the target object. This path must be feasible within all degrees of freedom to avoid restricted movement or collision of the manipulator. For example, for a manipulator with six degrees of freedom, its preliminary path planning will ensure that the joints of the manipulator can move freely in the space around the object while avoiding collisions with other objects, so as to reach the best grasping position.

[0064] After the preliminary path planning, each operation point is refined and calculated, and the key points in the path are interpolated using cubic spline interpolation to obtain the final operation path planning. Cubic spline interpolation is a mathematical method for smoothing the path, which can generate continuous and smooth curves in space to ensure that the movement of the robot arm is more stable and natural. In the preliminary path planning, there may be only a few key points, and the motion trajectory between these points may be discontinuous or not smooth enough. Using cubic spline interpolation, smooth interpolation can be performed between these key points, making the entire operation path smoother. For example, assuming that the preliminary path planning of the robot arm contains several position key points, if these points are directly connected linearly, it may cause the robot arm to suddenly accelerate or decelerate during operation, increasing the motion burden of the robot arm and the wear of the system. Through cubic spline interpolation, the path becomes smooth, the continuity of the motion is guaranteed, and the robot arm can maintain a stable speed and posture during execution, reducing the mechanical load.

[0065] After completing the path refinement, the final operation path planning is decomposed into operation steps to obtain multiple independent execution actions and construct an operation command sequence. The decomposition of operation steps is to decompose the complex overall path planning into actions that can be executed by the robot or robotic arm in sequence, for example: first extend the robotic arm to a certain position, then rotate a certain joint, and then move the end effector to grasp the object. These independent execution actions are a key step from path planning to actual operation, converting abstract path planning into specific mechanical instructions. For example, in the process of grasping a workpiece, the step decomposition may include: the robotic arm moves to the top of the target object, lowers the height, aligns the grasping position, grasps the object, and lifts to a safe height. Each action must be accurately executed based on the previous orientation estimation and path planning to ensure the success rate of the overall operation.

[0066] After constructing the sequence of operation commands, the path optimization algorithm is used to determine the optimal execution order and execution time, and generate an operation strategy for controlling the object operation module. The role of the path optimization algorithm is to find an execution plan that can minimize the operation time and maximize the operation efficiency by optimizing the order and time of executing actions. Especially in industrial automation, production efficiency is very important, so path optimization can reduce the idle time and unnecessary movement of the robot arm, thereby improving overall efficiency. For example, if the robot arm needs to grab workpieces in two different positions at the same time, the path optimization algorithm can help determine which position is more efficient to grab first, thereby reducing the total motion path and shortening the operation time.

[0067] The technical effect of the entire solution is that it starts with orientation estimation data, combines multiple means such as degree of freedom analysis, path refinement, operation decomposition and path optimization, and finally generates a stable and reliable operation strategy, enabling the robot or robotic arm to accurately handle objects in complex environments.

[0068] Specifically, for example, in the assembly process of automobile parts, the robot arm needs to grab multiple parts of different shapes in turn and assemble them. The position and orientation of each part may be different, and the assembly requires extremely high precision. Using the above technical solution, the orientation of each part is first determined, and then the operation path is generated according to the degree of freedom and spatial position. The smoothness of the path is ensured by cubic spline interpolation, and finally the operation efficiency is further improved by path optimization. In this way, the robot arm can accurately complete each assembly action in the shortest time and with the smallest movement amplitude to ensure the correct installation of the parts.

[0069] Therefore, this solution has achieved significant improvements in accuracy, flexibility and operational efficiency. It is particularly suitable for scenarios that require high-precision and high-stability operations, such as industrial automation, robot assembly, and precision grasping. It can effectively solve the problems of inflexible path planning and low operational efficiency in existing technologies, and provide strong technical support for the realization of more intelligent industrial production. It should be noted that this technical solution is not only applicable to the field of industrial manufacturing, but also very suitable for the field of item sorting and packaging. For example, after the fruit on the conveyor belt is identified, the robotic arm can adjust the placement according to the orientation of the fruit for subsequent packaging or further processing. By accurately estimating the orientation of the fruit and combining the operation path planning and optimization algorithm of the robotic arm, the system can automatically adjust the position and orientation of the fruit to ensure that they meet specific placement requirements. This flexible and efficient operation strategy can significantly improve the level of automation in fruit processing, reduce manual intervention, and improve the overall efficiency and processing accuracy of the production line.

[0070] like Figure 2 As shown, an embodiment of the present invention further provides a device for detecting the orientation of an object, comprising: An acquisition unit 1 is used to acquire images of a target object through a plurality of super-resolution full-spectrum imaging devices arranged in a path of the target object to obtain a plurality of original image sets; A preprocessing unit 2 is used to perform image preprocessing on the original image set, including image noise removal, image enhancement based on anisotropic diffusion and edge detection, to obtain a preprocessed image set; A feature extraction unit 3 is used to input the pre-processed image set into a preset sparse autoencoder convolutional neural network model to extract features and obtain object orientation feature data; The identification unit 4 is used to identify key feature points of the target object based on the object orientation feature data by using an improved generalized Hough transform to obtain feature point position information; An orientation determination unit 5 is used to estimate the orientation of the target object by using the feature point position information and an orientation determination algorithm based on fuzzy logic to obtain object orientation estimation data; The generating unit 6 is used to input the object position estimation data into the object operation module, and generate an operation strategy for grasping or processing the object according to the object position estimation data.

[0071] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0072] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0073] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0074] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0075] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0076] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0077] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for detecting the orientation of an object, characterized in that: The following steps are involved: Capturing images of the target object by using a plurality of super-resolution full-spectrum imaging devices arranged in the path of the target object to obtain a plurality of original image sets; Performing image preprocessing on the original image set, including image noise removal, image enhancement based on anisotropic diffusion and edge detection, to obtain a preprocessed image set; Inputting the preprocessed image set into a preset sparse autoencoder convolutional neural network model for feature extraction to obtain object orientation feature data; Based on the object orientation feature data, an improved generalized Hough transform is used to identify key feature points of the target object to obtain feature point position information; Using the feature point position information and an orientation discrimination algorithm based on fuzzy logic, the orientation of the target object is estimated to obtain object orientation estimation data; The object position estimation data is input into an object operation module, and an operation strategy for object grasping or processing is generated according to the object position estimation data.

2. The detection method according to claim 1, characterized in that: The original image set is subjected to image preprocessing, including image noise removal, image enhancement based on anisotropic diffusion, and edge detection, to obtain a preprocessed image set, including: For each pixel in the original image set, the spatial distance and color similarity between the pixel and the neighboring pixel are calculated, and the weighting coefficient of each neighboring pixel is calculated by using the spatial Gaussian kernel function and the color Gaussian kernel function respectively, and the denoised pixel value is calculated according to the weighting coefficient of each pixel to obtain a preliminary noise removal image set; Calculating the Laplace gradient value of each image in the preliminary noise removal image set, constructing an anisotropic diffusion coefficient matrix, and iteratively calculating each pixel according to the diffusion coefficient matrix to obtain an image gradient enhanced image set; Calculating the gradient direction of each pixel in each image in the image gradient enhanced image set to obtain a gradient direction distribution matrix, and performing local extreme value detection on the gradient direction distribution matrix to extract edge directions to obtain an edge direction image set; Calculating the second-order structure tensor matrix of each image in the edge direction image set, performing segmentation processing on the edge points in each image according to the structure tensor matrix, obtaining the edge feature distribution of the image, and obtaining the edge feature image set; Calculating the gradient divergence of the edge feature points of each image in the edge feature image set, constructing a divergence filter to perform a filtering operation on the edge feature points, and connecting the edge feature points processed by the divergence filter to obtain a refined edge image; The edge feature points and non-edge areas in each image in the refined edge image are marked, the edge areas are expanded and contracted through morphological dilation and erosion operations, and the edge areas and non-edge areas in the image are colored respectively to obtain a final preprocessed image set.

3. The detection method according to claim 1, characterized in that: The preprocessed image set is input into a preset sparse autoencoder convolutional neural network model for feature extraction to obtain object orientation feature data, including: Performing local block division on each image in the preprocessed image set, constructing a sparse representation matrix for each image block according to a sparsity constraint, and calculating a sparse coding result of each image by an L1 regularized minimization method; Inputting the sparse coding result of each image into the coding layer of the sparse autoencoder of the sparse autoencoder convolutional neural network model, and performing a dimensionality reduction transformation on the coding result using the activation function of the hidden layer to obtain a low-dimensional feature representation of each image; Constructing a convolution kernel of the sparse autoencoder convolutional neural network model, and using the convolution kernel to perform local weighted summation on the low-dimensional feature representation of each image, and performing nonlinear activation processing on the convolution result to obtain a convolution feature image containing local spatial features of the image; and, The maximum pooling method is used to locally downsample the convolution features of each image, retaining the most significant part of the features, and obtaining the pooled features of each image after downsampling; The sparse autoencoder is used again to inversely encode the pooled features of each image, and the pooled features of each image are sparsely reconstructed through the sparsity constraint and the reconstruction error minimization condition to obtain the reconstructed image features of each image; A full connection operation is performed on the reconstructed image features of each image, the image features are mapped to the feature space, and the object orientation feature data is calculated using an activation function.

4. The detection method according to claim 1, characterized in that: Based on the object orientation feature data, the improved generalized Hough transform is used to identify the key feature points of the target object to obtain the feature point position information, including: Acquire the prior geometric shape information of the target object, and construct a shape template model according to the prior geometric shape information of the target object, and perform pixel-by-pixel matching between the object orientation feature data and the shape template model to extract the initial feature point position that meets the template; The relative position relationship between each initial feature point and the surrounding pixels is calculated, and the feature points that do not meet the threshold conditions are eliminated using the gradient direction consistency and gradient amplitude difference metrics, and the candidate feature points that meet the conditions are retained; Quantify the gradient direction and distance parameters of each candidate feature point, map the quantization results into the voting space, and use the weighted voting mechanism for cumulative calculation to obtain the voting accumulation array for feature point detection; Perform neighborhood comparison on each voting position in the voting accumulation array to find the point with the largest voting value and use it as a candidate point for the key feature point to obtain a preliminary set of key feature points; The geometric consistency of the relative position and distance of each candidate feature point is calculated, and the candidate points that do not meet the preset geometric constraints are eliminated, and the key feature points that meet the geometric consistency are retained; The coordinates of each key feature point are accurately located, and the sub-pixel interpolation method is used to adjust the feature point position with sub-pixel accuracy to obtain the final feature point position information.

5. The detection method according to claim 1, characterized in that: The position information of the feature points and the orientation discrimination algorithm based on fuzzy logic are used to estimate the orientation of the target object to obtain object orientation estimation data, including: Vectorize the coordinates of each feature point in the feature point position information, calculate the angular offset value of the feature point relative to the center of the image, and obtain preliminary azimuth data of each feature point; The angle difference between each two feature points is fuzzified, a fuzzy membership function is constructed, and the angle difference is mapped into a membership value to form a fuzzy membership matrix, wherein the fuzzy membership matrix is ​​used to describe the spatial relationship between the feature points; Construct a set of fuzzy rule bases based on the object orientation, use the fuzzy inference engine to perform reasoning operations on the fuzzy membership matrix, combine the fuzzy rule base to judge the orientation of the target object, and obtain the fuzzy inference results of each feature point; Perform fuzzy set operations on the fuzzy reasoning results of each feature point, and obtain the overall aggregated orientation data set of the target object by weighted summing up the membership of the reasoning results of different feature points; Based on the centroid method, the fuzzy membership in the aggregated orientation data set is defuzzified and mapped to a specific orientation value to obtain the orientation estimation result of the target object; The defuzzified orientation estimation result is smoothed, and the estimated values ​​of different feature points are integrated using the weighted average method to obtain the final orientation estimation data of the object.

6. The detection method according to claim 1, characterized in that: The object position estimation data is input into the object operation module, and an operation strategy for object grasping or processing is generated according to the object position estimation data, including: Determine the position and orientation of the object in space based on the object orientation estimation data, and generate a preliminary operation path plan based on the degree of freedom parameters of the operating robot or robotic arm; Performing detailed calculation on each operation point in the preliminary operation path planning, interpolating key points in the path using cubic spline interpolation, and obtaining a final operation path planning; The final operation path planning is decomposed into operation steps to obtain multiple independent execution actions, and an operation command sequence is constructed. The optimal execution order and execution time are determined through a path optimization algorithm to generate an operation strategy for controlling the object operation module.

7. A device for detecting the orientation of an object, characterized in that: include: A collection unit, used to collect images of the target object through a plurality of super-resolution full-spectrum imaging devices arranged in the path of the target object to obtain a plurality of original image sets; A preprocessing unit, used for performing image preprocessing on the original image set, including image noise removal, image enhancement based on anisotropic diffusion and edge detection, to obtain a preprocessed image set; A feature extraction unit, used for inputting the preprocessed image set into a preset sparse autoencoder convolutional neural network model for feature extraction to obtain object orientation feature data; An identification unit, used to identify key feature points of a target object based on the object orientation feature data by using an improved generalized Hough transform to obtain feature point position information; An orientation determination unit, used to estimate the orientation of the target object by using the feature point position information and an orientation determination algorithm based on fuzzy logic to obtain object orientation estimation data; The generating unit is used to input the object position estimation data into the object operation module, and generate an operation strategy for grasping or processing the object according to the object position estimation data.

8. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images