1207 results about "Characteristic point" patented technology

The invention relates to the technical field of digital twinning control, and discloses a composite structure prefabricated segment digital twinning-adaptive assembly correction method and system, and the method comprises the steps: reading BIM geometric model data, and constructing an assembly reference coordinate system and a digital twinning geometry of a prefabricated segment; composite data are collected in real time, and end tooth groove boundary feature point cloud is extracted; matching the actually measured point cloud with the digital twinborn geometry through a dynamic point cloud registration technology, calculating a six-degree-of-freedom pose error vector, and generating a predicted total pose error vector in combination with a pre-constructed pose drift prediction model; based on the error vector, a mechanical fine adjustment jack is driven to execute position and posture adjustment until the position and posture are converged, and then tooth groove precise meshing and mechanical locking are completed; according to the method, through the synergistic effect of the dynamic mapping of the digital twinborn model and the self-adaptive correction algorithm, high-precision dynamic correction of the multi-combination structure segment assembly process is achieved on the premise that an original mechanical connection structure is not changed.

The invention provides an interaction method and system based on deep learning, and relates to the technical field of computer vision, and the method comprises the steps: collecting a continuous image sequence, extracting feature points, calculating the position variation and gradient variation of the feature points, constructing a two-dimensional feature vector, and carrying out the clustering through a Gaussian mixture model, thereby obtaining a stable feature class. And weighting the feature points based on a feature weight distribution function, calculating feature similarity and constructing a matching relationship. And calculating a dynamic parallax threshold in combination with the vehicle movement speed, screening reliable matching feature pairs, and solving pose transformation parameters. A feature change sequence is extracted through a feature detection network, a motion feature mapping matrix is constructed, and a compensation depth map is generated in combination with depth information. And scene reconstruction is carried out based on the compensation depth map and the pose transformation parameters, the scene confidence score is calculated, the pose parameters are calibrated, a high-precision pose parameter sequence is obtained, and the robustness of feature matching and the accuracy of pose estimation can be improved.

The invention discloses a visual inertial positioning method based on dynamic target detection and semantic information constraint, and belongs to the field of motion estimation and dynamic environment processing. According to the method, a dynamic target detection mechanism is introduced, the dynamic target is effectively detected based on the target detection network, inertial navigation information and geometric constraints, the dynamic target is effectively recognized in the image processing process, the corresponding dynamic feature points are screened out, the mismatching rate of the dynamic features is remarkably reduced, and high-quality observation input is provided for back-end optimization. In the back-end sliding window optimization stage, a semantic information consistency constraint method is constructed, and the estimation stability of the system in a weak texture area or a repeated texture area is enhanced by utilizing the consistency of feature points in the same semantic area on a geometric structure. Visual inertia pose estimation is realized based on dynamic target detection and semantic information constraint, and high-robustness and high-precision pose estimation can still be realized in a complex environment with dynamic interference of pedestrians, vehicles and the like and severe scene change.

The invention discloses a three-dimensional temperature field reconstruction method based on 3D Gaussian splashing, and the method comprises the steps: firstly carrying out the feature point extraction of an input color image, obtaining a descriptor, and carrying out the feature matching of the extracted features; and then generating a camera pose and realizing reconstruction of sparse point clouds, and finally selecting corresponding strategies to realize three-dimensional temperature field reconstruction according to different requirements. According to the method, three different strategies can be formulated according to different application requirements of three-dimensional reconstruction on storage space, reconstruction speed, reconstruction quality and the like in the actual application process so as to realize reconstruction of the three-dimensional temperature field, and a user can select corresponding strategies to realize reconstruction of the three-dimensional temperature field according to different requirements. The on-demand selection strategy can flexibly adapt to various scene requirements, and the applicability and reconstruction efficiency of the model are improved, so that calculation resources are effectively optimized, the reconstruction precision is ensured, unnecessary calculation consumption is reduced, and the reconstruction effects of the color image and the temperature image are improved at the same time.

The invention discloses a self-adaptive dynamic SLAM method based on Gaussian distribution, and relates to the technical field of dynamic environments. According to the method, the dynamic environment positioning precision is improved, the system can accurately recognize and eliminate interference of dynamic objects through a dynamic target elimination strategy based on CRF in combination with YOLO detection and conditional random field optimization, and a traditional SLAM system is likely to be affected by moving objects in the dynamic environment to generate pose offset, so that the system is not easy to operate. Reprojection error evaluation and feature point weighting processing are introduced, differential modeling is carried out on static points and dynamic points, the dependence of pose estimation on static features is strengthened, estimation errors caused by dynamic interference are effectively reduced, the positioning precision and stability of the system in a complex dynamic scene are improved, and the positioning accuracy of the system in the complex dynamic scene is improved. Moreover, the consistency and integrity of the map are enhanced, and the mapping efficiency and scene adaptability are improved.

The invention relates to a defect positioning method based on fusion of weld defect features and trajectory tracking data, and belongs to the technical field of weld defect detection and positioning. The method comprises the following steps: capturing welding seam track dynamic data and defect feature data, constructing a dynamic coordinate system based on a welding seam initial feature point, and establishing double-data-set reference mapping; performing multi-physics field interference decoupling correction on the trajectory data, and performing cross-modal feature purification and core feature consistency verification on the defect data; converting the preprocessed data into a feature form adaptive to fusion, and constructing a welding process-defect formation mechanism association network to regulate and control fusion weight; and finally, reconstructing a three-dimensional dynamic contour of the welding seam, calling dynamic positioning logic to position the defect, and outputting a result carrying the process-defect causal confidence coefficient. The positioning precision is improved through multi-dimensional data fusion and mechanism association, and technical support is provided for welding quality management and control.

The invention relates to the field of data processing, in particular to a part machining detection method and system, and the method comprises the steps: obtaining images of a to-be-detected part and a template part on a production line, and carrying out the preprocessing of the images, and obtaining corresponding gray images; using an edge detection algorithm to obtain a gradient magnitude and a gradient direction of each pixel point in the grayscale image, performing non-maximum suppression in a neighborhood of each pixel point, screening candidate feature points, and based on the candidate feature points and a plurality of attributes in the neighborhood, respectively obtaining significance feature values of the candidate feature points of the to-be-detected part and the template part; matching is carried out according to the saliency characteristic values of the to-be-detected part and the template part, the comprehensive matching degree is calculated, whether the image of the to-be-detected part is abnormal or not is judged based on the comprehensive matching degree, and part machining detection is completed; according to the method, the real feature points and the false feature points are effectively distinguished by introducing multi-dimensional features such as significance feature values and shape complexity, and the mismatching rate caused by similar texture structures is reduced.

The invention relates to a point cloud registration method based on improved ISS-TOLDI features in combination with ICP, and belongs to the technical field of laser point cloud application. The method comprises the steps of obtaining a source point cloud and a target point cloud, and performing preprocessing; extracting feature points from the point cloud data by using an internal shape descriptor ISS algorithm; performing feature description on the extracted feature point set by using an improved TOLDI algorithm; performing coarse registration on the point cloud by using a sampling consistency initial registration algorithm; and performing fine registration on the coarsely registered point cloud data by using an iterative closest point ICP algorithm to complete point cloud registration. According to the method, the TOLDI algorithm is improved, so that the calculation complexity is reduced, the feature integrity is ensured, and the point cloud registration precision is higher.

The invention discloses a stereo matching method based on context geometric features and distortion parallax optimization, which comprises the following steps of: firstly, shooting a calibrated left picture and a calibrated right picture through a binocular camera, extracting multi-scale context features of the two pictures, and calculating the similarity between feature points according to the feature points of the left feature graph and the right feature graph under different parallaxes; the matching cost is the matching cost, so that the constructed geometric cube contains rich geometric matching information. Meanwhile, context features are copied and expanded into context feature cubes with the same size according to parallax, and rich context information is contained. The geometric cube contains matching conditions of feature points of left and right views with different parallax errors, cost aggregation is guided through context features of the left view, and context feature information, feature point cost matching information under different parallax errors and geometric information of adjacent feature points are preliminarily fused to improve cost matching reliability. And generating a geometric attention weight cube. And filtering the contextual feature cube.

The invention provides a deep learning-based die process die surface automatic supplement generation method and system, and relates to the technical field of deep learning, and the method comprises the steps: extracting the contour line and feature point information of a die surface as a training sample; feature learning is carried out based on a multi-scale image convolutional network and a double-flow network, and optimized local geometric features and global topological features are extracted; generating a supplementary modular surface through the generative adversarial network; and carrying out stress field constraint verification and then fusing with an original model. According to the method, intelligent supplementary generation of the mold surface is realized, and the mold design efficiency and quality are improved.

The invention provides a material data curve identification and fitting method based on deep learning, and relates to the technical field of material data extraction, and the method comprises the steps: obtaining an image file containing a plurality of material data curves; separating the material data curve from the image background by adopting an image segmentation technology, and extracting feature points of the material data curve through a convolutional neural network; and taking the feature points as control points, performing parametric fitting on the curve through polynomial or spline interpolation, and generating a smooth continuous curve. The method further comprises the following steps: detecting the position and boundary of a coordinate axis in the image file through an image processing technology; the OCR technology is used to identify the scale label on the coordinate axis, and the deep learning model is combined to analyze the numerical value and unit of the scale. According to the method, through innovative technologies such as multi-modal feature fusion, self-adaptive preprocessing and intelligent data labeling, precise recognition of complex curves, multi-curve separation and association, intelligent recognition of non-standard coordinate axes and efficient real-time processing are achieved.

The invention relates to the technical field of weld joint detection, in particular to a CO2 welding automatic trolley weld joint recognition method. According to the method, by measuring current signals and magnetic conductivity parameters in real time and combining comparison of dynamic change amplitudes, the initial position of a weld joint is accurately recognized, the initial accuracy of weld joint detection is improved, meanwhile, a displacement characteristic curve is extracted, weld joint form data is dynamically adjusted, the continuity and accuracy of weld joint path recognition are ensured, and the path tracking effect is improved; the welding seam boundary and width range are accurately judged in combination with electromagnetic parameter differences, the reliability of welding seam boundary detection is facilitated, irrelevant parameters are removed through frequency selection, edge contours are captured, welding seam surface features are marked, feature point layered clustering and normalization processing are conducted through a Gaussian mixture model, welding seam feature point data are comprehensively integrated, and the accuracy of welding seam boundary detection is improved. The generated feature point model provides an important basis for high efficiency and accuracy of weld joint recognition, and the performance of weld joint detection in the aspects of paths, forms and boundaries is integrally optimized.

The invention relates to the technical field of camera measurement, and discloses a target positioning method and device, computer equipment and a storage medium, and the method comprises the steps: determining a model point and candidate feature points corresponding to a target point on a current frame image; determining a reference feature point according to the distance between the model point and the candidate feature point; determining matching point pairs according to the reference feature points; constructing a similar transformation matrix according to the matching point pairs, and determining a registration error corresponding to the similar transformation matrix; when the registration error is smaller than a preset threshold value, determining the similar transformation matrix as a target transformation matrix; determining a search area on the next frame of image through the covariance matrix; fitting a plurality of prediction points in the search area by using a least square method to obtain a fitting function, and solving an extreme point of the fitting function; and correcting the extreme point according to the target transformation matrix to obtain the position and speed of the target point on the next frame of image. Based on the technical scheme of the invention, the precision and efficiency of target point positioning can be improved.

The invention discloses a roof photovoltaic identification and evaluation method and system based on key point detection enhanced semantic segmentation, and a storage medium. The method comprises the following steps: extracting significant feature points of a roof from a satellite image as key points; for each detected key point, generating a descriptor vector; combining the generated key point information with a pre-trained semantic segmentation network, and optimizing a preliminary segmentation result of the semantic segmentation network by using the key point information; and carrying out key point matching and similarity comparison by using the descriptor vector to identify the contour and important structural features of the roof, and classifying and identifying the roof according to a matching result to obtain a final roof segmentation result. According to the method, through combination of key point detection and semantic segmentation, high-precision identification and description of a roof in a complex environment are realized, and the method has good robustness and adaptability.

The invention relates to an intelligent solid state disk quality detection method, system and device and a medium, and the method comprises the steps: obtaining test demand data of a to-be-detected solid state disk, extracting feature point data in the test demand data from the test demand data through a feature extraction algorithm, and obtaining demand feature point data; acquiring environment data, and transmitting the demand feature point data and the environment data to a pre-trained detection model for comparative analysis to obtain a test adjustment scheme; performing a virtual environment test according to the test adjustment scheme to obtain a virtual environment test result; and when the virtual environment test result does not conform to the expectation, obtaining difference information from the virtual environment test result, correspondingly modifying the test adjustment scheme according to the difference information to obtain an optimized adjustment scheme, and adjusting the test process and parameter configuration of the to-be-detected solid state disk according to the optimized adjustment scheme. The method has the effect of improving the accuracy of solid state disk quality detection.

The invention discloses an IMU (Inertial Measurement Unit)-assisted moving target detection method under a dynamic condition, which comprises the following steps of: firstly, extracting feature points of a target image for matching to obtain feature point pairs; then, relative coordinate conversion is carried out on IMU sensor data, a corresponding basic matrix or homography matrix is solved, and then corresponding polar lines or mapping projection feature points are obtained; classifying the extracted feature points into two parts, namely static background feature points and dynamic target feature points; an affine transformation matrix is solved according to the corresponding relation of the static background feature points of the two adjacent frames of images and the motion parameter model, an inter-frame difference method is used for carrying out difference on the two frames of images and carrying out image binarization and morphological processing, and a corresponding motion target frame is obtained; and finally, traversing the moving target frame to carry out overlap ratio matching, and screening out a detection object of which the overlap ratio is greater than a set threshold value, namely an actual moving target in the current image. The method provided by the invention can effectively improve the precision and accuracy of moving target identification in a dynamic background.

The invention discloses an image registration method based on improved ORB feature extraction and grid feature matching. The method comprises the following steps: inputting an image to be extracted; using an improved FAST algorithm to extract feature points of the to-be-spliced image; establishing a BRIEF descriptor, and obtaining a BRIEF descriptor with a direction; according to the feature point set of the to-be-spliced image, establishing a Delaunay triangular grid taking the feature points as vertexes; judging the similarity of triangles, and obtaining similar triangles formed by the feature points; matching the similar triangle vertexes by using a Hamming distance to obtain an optimal matching feature point pair; obtaining affine matrix parameters by using an SVD method; according to the image registration method provided by the invention, the image quality is ensured, the feature extraction efficiency and the extraction precision can be improved, wrong matching is reduced, the stability is improved, the real-time performance of video stitching is improved, the waste of memory resources is reduced, the equipment cost is reduced, the working efficiency is greatly improved, and the method is suitable for popularization and application. And a better method is provided for image splicing.

The invention relates to the technical field of point cloud data processing, in particular to an indoor scene point cloud semantic segmentation method and system for a wearable device, which can better capture tiny details in a scene through curvature-based feature point selection and local geometric feature extraction, improve the capture capability of fine object edges, and improve the accuracy of point cloud data processing. Therefore, the segmentation precision is obviously improved; meanwhile, by introducing multi-scale local geometric features and self-adaptive neighborhood definition, the neighborhood radius is adjusted in a self-adaptive mode through curvature change, and therefore the method is excellent in performance in a non-uniform point cloud scene; according to the method, the feature extraction process is simplified, and the lightweight multilayer perceptron is used as a classifier, so that the calculation complexity is remarkably reduced. In combination with optimization strategies such as farthest point sampling (FPS) for reducing point cloud processing points, KD tree for accelerating neighborhood search, parallel processing feature extraction and classification and the like, the method can keep real-time performance under the condition of limited computing resources, and the requirements of low-power-consumption equipment are met.

The invention provides a self-adaptive workpiece grabbing pose correction method and system based on feature point matching. The method comprises the steps that a standard image of a workpiece which is successfully grabbed to a jig and an auxiliary image of the workpiece which is not successfully grabbed to the jig are shot; the template image and the auxiliary image are processed, feature points in the images are recognized and compared, matched feature points between the template image and the similar image are found out, and three-dimensional coordinates matched with the feature points are obtained; and based on the matched feature points and the three-dimensional coordinates of the feature points, the pose of the workpiece is solved through an RANSAC PNP algorithm, the pose of the current unsuccessfully grabbed workpiece relative to the position of the template drawing is recovered, and it is ensured that the workpiece can be accurately placed on the jig by adjusting the pose of the mechanical arm. And through feature extraction and matching of the SIFT algorithm and the LightGlue algorithm, accurate and effective correction of the workpiece placement pose is achieved, important industrial application value and market potential are achieved, and powerful technical support is provided for improving the placement precision and efficiency of the mechanical arm.

The invention relates to the technical field of bolt length detection data processing, in particular to a bolt length detection system, which is characterized in that a three-dimensional data acquisition module generates a high-density point cloud and extracts a thread curvature extreme point, and a dynamic model matching module screens an optimal lead angle parameter based on a Hausdorff distance; and constructing an axial expansion model carrying phase offset gradient features. The projection mapping correction module adopts a genetic algorithm to optimize a compensation coefficient matrix, and generates a dynamic weight correction parameter to suppress geometric distortion; and the phase error correction module combines Fourier spectrum decomposition and an exponential decay weight coefficient to construct a frequency domain transfer function to eliminate an axial accumulative error. According to the method, the problems of projection distortion and phase deviation caused by the lead angle are effectively solved through geometric-phase double-domain correction, dynamic parameter iteration and an error collaborative suppression mechanism, and the spatial positioning precision of the feature points at the tail end of the thread is improved.

The present application provides a method for extracting feature points, a method for reconstructing an image, and a device. The method for extracting feature points includes: determining the actual grid size based on the size of an input image and the expected number of feature points n; dividing the input image into grids based on the actual grid size to obtain n grids, where n is an integer greater than 1; extracting feature points from each of the n grids based on an initial threshold to obtain g feature points for the grid; when g is greater than K, determining K feature points with the highest response values ​​among the g feature points as feature points of the grid, where K is the number of feature points to be extracted for each grid. The technical solution of the present application can obtain uniformly distributed high-quality feature points.

The invention provides a seamless rapid fusion method and system for multivariate three-dimensional data, a product and a medium, and relates to the technical field of electric digital data processing. By acquiring multi-source heterogeneous data and unifying the multi-source heterogeneous data to a coordinate system of a geographic base map layer, a datum reference for data fusion is established. And the geographic base map layer is used as a benchmark to provide a stable spatial reference for subsequent data fusion. The surface smoothing algorithm is adopted to preprocess the live-action three-dimensional model and the monomer model data, so that the influence of data noise on the fusion precision is reduced. Through cooperative use of feature point extraction and an iterative nearest point algorithm, accurate alignment of different types of three-dimensional data is realized. According to the feature-based iterative optimization mode, the error accumulation problem caused by direct overall fusion is avoided, the fusion precision is improved, and the geometric consistency of the fused scene is ensured.

The invention discloses a pose estimation method, computer equipment and a storage medium. The method comprises the following steps: acquiring point cloud data of a current frame, and performing feature point extraction on the point cloud data to obtain at least two types of feature points; performing region division on the feature points of each category to obtain at least two categories of feature points corresponding to each region; clustering processing is carried out on the feature points, corresponding to the areas, of the categories to obtain centroid point sets corresponding to at least two categories; and performing pose estimation by using the centroid point sets corresponding to the at least two categories to obtain the current pose of the target equipment. According to the scheme, the accuracy of pose estimation can be improved.

The invention discloses a target object pose estimation method, system and device and a medium, and relates to the technical field of pose estimation, and the method comprises the steps of offline data set establishment and online real-time multi-target pose estimation. In the off-line process, continuous RGBD streams surrounding a single object are collected, the relative poses of front and back frames are calculated frame by frame, and meanwhile, the three-dimensional surface of a target object is reconstructed to obtain a key frame off-line data set; in the online process, a visible target set in a data stream is firstly detected, VOS segmentation is performed on each target to obtain real-time effective texture and point cloud, offline key frame images of to-be-detected targets are arranged to realize single matching of all feature points and find out a closest reference frame at the same time, and RANSAC is adopted to calculate relative poses of a current frame target and the reference frame; according to the method, an expensive data labeling process is avoided, rapid integrated deployment of detection of an object without a new target is realized, the matching speed is improved, and meanwhile, online real-time multi-target pose estimation can be realized.

The invention discloses a multi-scale machine vision matching method, and belongs to the technical field of vision matching, and the method comprises the steps: determining an original scale template image I1 based on a search image I0 inputted by a user, and obtaining a template feature point set under each layer of pyramid of the original scale; calculating a template scale step s; calculating upper and lower limits of a template scale; calculating a scale factor set S; traversing the scale factor set S to obtain each layer of pyramid feature point set corresponding to different scale factors; performing feature correction on each layer of pyramid feature point set of which the scale factor is less than 1; calculating a pyramid image sequence set of a search image I0 input by a user; performing top layer template similarity calculation to obtain a top layer matching result; performing neighborhood suppression on all the top-layer matching results to obtain a top-layer matching result set; and traversing all pyramid layers by adopting a layer-by-layer approximation mode to obtain a final matching result. According to the invention, targets of multiple scales can be effectively positioned.

The invention discloses a multi-domain feature level fusion expression recognition method based on a graph neural network, and relates to the technical field of computer vision, and the method comprises the steps: firstly, carrying out the data enhancement of an original image, and carrying out the region division of a face according to three courts and five eyes based on key point detection; secondly, constructing a graph structure of the face, endowing each edge with a corresponding weight according to an area where the feature points are located, obtaining weighted distance information, and storing the weighted distance information in an adjacent matrix; then spectral domain features and spatial domain features are obtained through a graph neural network, and local features and context features are obtained through a multistage convolutional neural network and ViT; and then, multi-domain feature fusion is carried out through a feature fusion network based on structural feature guidance, and then an expression recognition result is obtained. Therefore, by adopting the multi-domain feature level fusion expression recognition method based on the graph neural network, the model can understand the change of the expression more deeply, the recognition capability is enhanced, and the robustness and the stability of the model are improved.

To provide an abnormal behavior notification system capable of making an appropriate notification according to the attributes of a behavioral entity when abnormal behavior such as an abnormal behavior of a behavioral entity captured on a video is detected. [Solution] The abnormal behavior notification system 1 includes a determination unit 5 that determines whether abnormal behavior has occurred based on the displacement of stored feature points and the displacement of feature points detected from each time-series image Y, an estimation unit 6 that estimates the attribute or size of a target behavior entity Z based on the detected feature points or the appearance of the target behavior entity, and a notification unit 7 that issues a notification when it is determined that abnormal behavior has occurred. Even when it is determined that abnormal behavior has occurred, the notification unit 7 does not issue a notification if the attribute or size satisfies a first predetermined condition and the abnormal behavior satisfies a second predetermined condition set for the attribute or size.

At present, most researches on a visual synchronous localization and mapping (VSLAM) system are established on a static scene, and a dynamic target in a real scene can seriously influence the accuracy and robustness of the VSLAM system, so that the invention provides an RGB-D visual SLAM algorithm for an indoor dynamic scene. The method comprises the following steps: (1) fusing a lightweight YOLOv8 target detection network in an ORB-SLAM3 tracking thread, performing instance segmentation on an image frame, and generating semantic information; (2) processing the semantic information by using a sliding window mechanism and a template matching strategy, identifying a dynamic target and removing dynamic feature points; and (3) a PROSAC algorithm is adopted to replace an original RANSAC algorithm to remove mismatching of feature points, and the matching performance is improved. In order to evaluate the effectiveness of the method, an experiment is carried out on a disclosed TUM RGB-D data set, and compared with an original ORB-SLAM3 and a relatively advanced dynamic DynaSLAM algorithm, the method has the advantages that the algorithm positioning precision and robustness in a dynamic environment are greatly improved, the frame rate can reach 45 fps or above under the condition of no GPU acceleration, and real-time operation can be ensured.

The invention relates to a smart power grid deep learning model copyright protection method, device and system based on a model confidence domain, a computer readable storage medium and a computer program product. The method comprises the steps of obtaining a smart grid depth model set; aiming at each model in the intelligent power grid depth model set, searching a characteristic data sample of which the model prediction value is close to the model discrimination boundary in the special data set to obtain a confidence domain characteristic point set; carrying out dimensionality reduction on a gradient vector of the confidence domain feature point set on a model discrimination boundary according to a linear discriminant analysis method to obtain a disturbance vector of the confidence domain feature point set; generating a model feature identification set corresponding to the intelligent power grid deep model set according to the prediction label change before and after the confidence domain feature point set and the disturbance vector are combined by each model; and training a to-be-trained copyright detection model according to the model feature identification set to obtain a pre-trained copyright detection model. By adopting the method, the detection accuracy of the model copyright infringement attack can be improved.

The invention discloses an FP-XFat feature device and a typical rescue scene-oriented three-dimensional reconstruction method and system, and relates to the technical field of three-dimensional reconstruction. The FP-XFat feature device is internally provided with a pre-training model, and the pre-training model comprises an improved PIOU loss function; the implementation mode of the improved PIOU loss function is as follows: 1) constructing a dynamically adjusted weight coefficient, reinforcing geometric constraints when the IoU is low, and weakening the constraints when the IoU is high; 2) introducing a shape difference metric item based on an arc tangent function, converting the aspect ratio into an angle dimension, and eliminating the influence of scale difference on shape judgment; and 3) adopting a gradient separation technology to block back propagation of an IoU gradient to a weight coefficient. The FP-XFat feature device introduces an improved PIOU loss function, and through an adaptive weight mechanism and a shape perception penalty term, the accuracy of feature point positioning can be significantly improved.