A fast conversion method for large space positioning based on image recognition

Abstract

The present invention discloses a rapid conversion method for large space positioning based on image recognition, and relates to the technical field of large space positioning. The present invention comprises the following steps: large space image acquisition and deployment, on-site measurement of large space size and obstacle positions, importing software to build a basic three-dimensional model, simulating the initial position of the camera, determining the final layout through evaluation and screening, and installing and debugging; the present invention significantly improves the accuracy and reliability of target positioning in a large space by comprehensively utilizing image recognition technology and a multi-camera collaborative positioning mechanism; through the combination of triangulation principle and Kalman filtering technology, the position of the target in three-dimensional space can be effectively inferred, and the positioning result can be corrected in real time, and the positioning jitter caused by image noise and temporary occlusion of the target can be smoothed; not only the continuity of positioning is improved, but also the stability and accuracy in complex environments are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of large space positioning, and in particular to a fast conversion method for large space positioning based on image recognition. Background Art

[0002] In traditional large-space positioning technologies, common challenges include insufficient accuracy, limited coverage, and cost-effectiveness issues. These technologies often rely on complex hardware facilities and high maintenance costs. In actual applications, environmental factors (such as occlusion, lighting changes, etc.) may cause positioning accuracy to decrease. In addition, as the size of the space increases, the viewing angle limitation of a single camera makes all-round monitoring difficult, and more cameras need to be deployed to achieve full coverage, which further increases costs. Therefore, existing technologies have obvious limitations in achieving fast and accurate positioning of targets in large spaces;

[0003] With the development of image recognition technology, more and more solutions are trying to use image data for spatial positioning. These methods identify and track targets by analyzing images captured by cameras, providing a relatively low-cost and easy-to-deploy solution. However, these methods face challenges when dealing with large spatial environments, including image preprocessing and enhancement, accurate positioning of targets, and collaborative processing of multi-camera data. Especially in dynamic environments, how to effectively extract useful information from continuous image streams and update the position of targets in real time is a key problem that image recognition technology needs to solve in large-space positioning. Summary of the invention

[0004] The purpose of the present invention is to solve the problems of insufficient accuracy, limited coverage and cost efficiency in the existing large-space positioning technology, and to propose a fast conversion method for large-space positioning based on image recognition.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] The fast conversion method for large space positioning based on image recognition includes the following steps:

[0007] B1: Large space image acquisition and deployment, on-site measurement of large space dimensions and obstacle locations, importing software to build a basic 3D model, simulating the initial position of the camera, determining the final layout plan after evaluation and screening, and installing and debugging the camera;

[0008] B2: Spatial positioning modeling: After receiving the image stream, it performs preprocessing and enhancement, locates the target using camera parameters and triangulation principles, and outputs position information through multi-camera collaboration and Kalman filter optimization;

[0009] B3: Target attribution judgment, obtain camera parameters to calculate the field of view boundary, convert the target coordinates to determine whether it is within the field of view, and determine the camera priority based on multiple factors when processing overlapping areas;

[0010] B4: Image retrieval and fusion annotation: retrieve and sort images according to target information, annotate targets and add metadata after registration and fusion, and finally select format for storage.

[0011] Furthermore, the process of B1 is as follows:

[0012] Technicians carry measuring equipment to conduct on-site measurements of large spaces, record length, width, and height data, draw spatial plans, and mark the locations of doors, windows, columns, partitions, and fixed obstacles;

[0013] Import the dimension data and obstacle location information obtained from field measurements into spatial topology modeling software, including AutoCAD Map 3D and SketchUp with topology analysis plug-ins; construct a basic three-dimensional model of the large space based on the data, restore the spatial structure, and provide a visual carrier for subsequent algorithm operations;

[0014] The initial camera installation points are simulated on the spatial topology model using the camera layout planning algorithm. The algorithm takes full coverage, no blind spots, and reasonable image overlap as optimization goals. It inputs large space dimensions, obstacle distribution, camera field of view, and focal length parameters. After multiple rounds of iterative calculations, it outputs several candidate installation point combinations. Specifically:

[0015] First, determine the camera layout planning algorithm, including genetic algorithm or simulated annealing algorithm; input the size information of the large space, the distribution data of marked doors, windows, columns, partition obstacles, and the field of view and focal length parameters of the camera into the algorithm; the algorithm randomly generates the initial camera installation site combination on the spatial topology model based on the requirements of full coverage and no blind spots, as the initial population or initial solution; then, calculate the camera's field of view coverage under each site combination, evaluate its coverage rate, whether there are monitoring blind spots, and whether the degree of image overlap meets the standards, and assign fitness values ​​or objective function values ​​based on these indicators;

[0016] Through multiple rounds of iterations, the genetic algorithm performs selection, crossover, and mutation operations, and the simulated annealing algorithm accepts inferior solutions based on probability, continuously optimizing the camera installation point combination until the preset number of iterations is reached or the convergence condition is met, and finally outputs a number of candidate installation point combinations;

[0017] From the candidate solutions, the final layout plan is determined based on the evaluation and screening of the visual field coverage, regional clarity, installation cost, and convenience of later maintenance indicators;

[0018] Use visualization tools to display the camera's field of view coverage; review the blind spots in large spaces to determine whether additional cameras are needed;

[0019] According to the final layout plan and the cameras added in the blind spots, the construction team is arranged to install cameras at the predetermined locations and adjust the camera angles to the initial angles; after the installation is completed, debug them one by one to check the image acquisition and transmission functions; test the night vision and strong light suppression effects under different lighting conditions and time periods; adjust the camera shooting angle, focal length, and white balance parameters to make the picture clear and the color restoration accurate; debug multiple cameras in a linked manner to ensure smooth switching and natural connection of vision, so as to achieve all-round monitoring without blind spots.

[0020] Furthermore, the specific operation steps for determining the final layout solution in B1 are as follows:

[0021] Comprehensive analysis is conducted through the evaluation indicators of candidate solutions, including:

[0022] Field of view coverage: Calculate the percentage of the spatial volume covered by the camera field of view in each candidate solution to the entire large space volume. This is achieved by dividing the large space into tiny cubic units and counting whether each unit is covered by at least one camera field of view. The field of view coverage is recorded as sf.

[0023] Regional clarity: For each camera’s field of view, calculate the average resolution ra, which is estimated based on the camera’s resolution parameters and its imaging effects at different distances; calculate the clarity scores around entrances and exits and important equipment using a clarity evaluation algorithm to obtain a clarity score between 0 and 1, which is recorded as qp;

[0024] Installation cost: Calculate the total cost of each candidate solution and record it as zf, including the camera equipment procurement cost, installation material cost, and installation labor cost. The equipment procurement cost is determined according to the model and quantity of the camera, and the installation material and labor costs are estimated based on the market price and installation workload;

[0025] Maintenance convenience: The accessibility scores of all cameras in each candidate solution are counted, and the mean accessibility score is calculated, denoted as hb. The accessibility score is higher if the camera is installed in an easily accessible location, and lower if it is installed in an easily accessible location. The accessibility score is set to a value between 0 and 1.

[0026] The obtained field of view coverage sf, average resolution ra, clarity score qp, total cost zf and accessibility score mean hb are normalized and entered into the following formula: The optimal evaluation value ZGM is obtained, and the optimal evaluation value ZGM is used as the evaluation standard for measuring candidate solutions. They are the preset weight coefficients of the field of view coverage sf, average resolution ra, clarity score qp, total cost zf and accessibility score mean hb;

[0027] The calculated merit values ​​ZGM of all candidate solutions are sorted by size, and the candidate solution with the largest merit value ZGM is selected as the final layout solution.

[0028] Furthermore, the specific operation steps of B2 are as follows:

[0029] Continue to receive the image streams sent back in real time from each camera in step B1, sort them according to the acquisition time and camera number information, and remove abnormal frames caused by network jitter and instantaneous strong light interference;

[0030] Adaptive filtering algorithm is used to dynamically adjust filtering parameters, remove image noise, and restore clear images. Image enhancement processing is performed simultaneously, and CLAHE is used to improve image contrast and highlight the contour details of the target object.

[0031] With the help of the camera internal and external parameters recorded in step B1, combined with the extracted two-dimensional coordinate information of the target in the image, the position of the target in three-dimensional space is inferred using the principle of triangulation; this preliminary position information is associated with the basic three-dimensional model coordinate system, and the target is preliminarily anchored in the model framework;

[0032] Introduce a multi-camera collaborative positioning mechanism; combine the same target images captured by multiple cameras with different perspectives, calculate the target's three-dimensional position respectively, and then fuse multiple sets of coordinate data through weighted averaging and least squares optimization to locate the target's spatial position in the three-dimensional model; weight distribution is dynamically adjusted based on the distance between the camera and the target, imaging clarity, and perspective stability factors;

[0033] Using Kalman filter technology, combined with the target's historical positioning trajectory and motion state estimation information, the current positioning result can be corrected in real time; based on the assumption that the target moves at a uniform speed or uniform acceleration, Kalman filter continuously iterates and updates the target position and speed parameters, smoothing the positioning jitter caused by image noise and short-term target occlusion, thereby improving positioning continuity and accuracy;

[0034] The final accurately located target location information is output in a standardized format and connected to other application systems; at the same time, target identification tags and real-time status information are added to enrich the positioning data content.

[0035] Furthermore, the specific operation steps of B3 are as follows:

[0036] Obtain the internal and external parameters of each camera in step B1, which are used as the basis for defining the camera's field of view; use camera calibration technology to calculate and verify the internal and external parameters of the camera by placing calibration objects of known size and position in a large space and taking multiple sets of images to ensure the accuracy of the parameters;

[0037] According to the internal and external parameters of the camera, the perspective projection principle is used to calculate the field of view boundary of each camera in three-dimensional space; in the three-dimensional model coordinate system, the far and near planes of the field of view and the boundary angles in the horizontal and vertical directions are determined through geometric calculation; for each camera, its field of view range is represented as a pyramid in three-dimensional space, whose vertex is the center of the camera lens and the bottom surface is a polygon formed by the farthest boundary of the field of view;

[0038] The position coordinates of the target in the three-dimensional model determined by the spatial positioning modeling are converted to the same coordinate system as the camera's field of view; specifically, a coordinate transformation matrix is ​​used, which is calculated based on the camera's external parameters;

[0039] For the converted target position coordinates, determine whether it is inside the pyramid within the camera's field of view, specifically by determining the position relationship between the point and the plane and the angle.

[0040] Using spatial geometry algorithms, including ray and plane intersection algorithms, rays are emitted from the center of the camera lens to the target position to determine whether the rays intersect with the boundary plane of the field of view and whether the intersection position meets the definition of the field of view, so as to determine whether the target is within the field of view of the camera;

[0041] Since there are overlapping areas in the camera layout, it is necessary to identify these overlapping areas and calculate the boundaries and ranges of the overlapping areas by comparing the intersections of pyramids in different camera field of view in three-dimensional space. The spatial geometry operation library is used to calculate the intersections between pyramids to determine the position and shape of the overlapping areas in the three-dimensional model.

[0042] When the target is in the overlapping area, the priority of the camera to which it belongs is determined, and a comprehensive judgment is made based on the distance between the camera and the target, the importance of the camera, and image quality factors;

[0043] Establish a priority evaluation function, use the above factors as parameters for weighted calculation, and calculate a comprehensive priority score based on the scores of each camera on these factors. The camera with the highest score is the camera to which the target belongs in the overlapping area.

[0044] The above steps are repeated at a preset frequency to match the target position with the camera field of view. An event-driven mechanism is used to immediately trigger the target attribution judgment process when the target position changes by more than a threshold or the camera parameters change significantly, ensuring that the camera to which the target belongs can be determined in a timely and accurate manner.

[0045] Further, the analysis of the comprehensive priority score in B3 is as follows:

[0046] When the target is in the overlapping area, for camera i, i=1,2...,n, n is the number of cameras involved in the overlapping area, its comprehensive priority score is The calculation formula is as follows: ,in represents the comprehensive priority score of camera i; They are distance weight, camera importance weight, image quality weight, and ;

[0047] Represents the distance from camera i to the target, which is obtained by calculating the spatial coordinates. Let the three-dimensional coordinates of the camera be , the three-dimensional coordinates of the target are ,but ;use The form participates in the calculation, that is, the closer the distance, The larger the value of ;

[0048] Represents the importance score of camera i, with a value range of 0-1. The importance level of each camera is manually set according to the functional zoning of the large space;

[0049] Indicates the image quality score of camera i, also in the range of 0-1; the calculation method is ;in is the resolution normalized score of camera i, is the normalized score of light conditions, and are the weights of resolution and light conditions in image quality assessment, respectively, and ;

[0050] For resolution normalized score :Assume the horizontal resolution of camera i is , the vertical resolution is , the highest horizontal resolution of all cameras in the large space is , the maximum vertical resolution is ,but , to ensure that cameras with higher resolutions have higher scores in this area;

[0051] Normalized scores for light conditions : Use the light sensor to monitor the light intensity at the camera i in real time , and obtain the maximum light intensity recorded by the camera in the entire large space and minimum value ,but , that is, the greater the light intensity, The larger it is, the better the lighting conditions are;

[0052] Calculate the overall priority score of all cameras After that, select The camera with the largest value is taken as the camera to which the target belongs in the overlapping area.

[0053] Furthermore, the specific operation steps of B4 are as follows:

[0054] According to the target information determined in step B3 and the camera identification associated therewith, all the original image data of the target are retrieved in the distributed storage system; these image data are dispersedly stored in the storage nodes or storage partitions corresponding to different cameras;

[0055] Sort and organize the retrieved images according to the time sequence of image acquisition, and construct the target's image time series set for subsequent analysis of the target's movement trajectory and behavior changes to ensure the coherence and logic of image presentation;

[0056] For target images from different camera perspectives, there are geometric distortions and position deviations between images due to differences in shooting angles and distances. Use a registration method based on feature point matching or a registration algorithm based on mutual information to determine the correspondence between targets in different images and achieve geometric alignment of images, so that the positions and shapes of targets in different images can accurately correspond.

[0057] Adopt multi-resolution image fusion technology, including fusion algorithm based on Laplace pyramid decomposition; perform Laplace pyramid decomposition on the registered images to obtain image representations at different resolution levels; then, at each resolution level, according to fusion rules, including weighted average, selection strategy based on regional energy or contrast, fuse image information from different cameras, and generate a fused multi-resolution image sequence by combining the advantages of each perspective;

[0058] On the fused image, the target is annotated using the target recognition and positioning results; the annotation content includes the target type, the target unique identifier, and the target position bounding box information in the image; it is used for subsequent rapid recognition and positioning of the target in the image;

[0059] Add rich spatiotemporal metadata information to each image, including image acquisition time, acquisition location, camera number, image resolution, and frame rate; metadata information provides an important basis for image management, retrieval, and analysis, enabling location and screening when querying target images captured by time, region, or camera;

[0060] Select the storage format to store the image sequence after adding annotations and metadata, including JPEG2000 and PNG; organize and store the images according to the directory structure and naming rules.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] (1) The present invention significantly improves the accuracy and reliability of target positioning in a large space by comprehensively utilizing image recognition technology and multi-camera collaborative positioning mechanism. By combining the triangulation principle and Kalman filtering technology, it can effectively reverse the position of the target in three-dimensional space, and correct the positioning results in real time, smoothing the positioning jitter caused by image noise and temporary occlusion of the target. It not only improves the continuity of positioning, but also enhances the stability and accuracy in complex environments.

[0063] (2) The present invention optimizes the layout of cameras through algorithms, realizes full coverage and no-dead-angle monitoring coverage, and reasonably controls the overlap of images to achieve the best balance between cost and effect; uses genetic algorithms or simulated annealing algorithms to simulate and optimize camera installation points, can output a number of candidate installation point combinations, and evaluate and screen based on indicators such as field of view coverage, regional clarity, installation cost and convenience of later maintenance to determine the final layout plan; reduces the number of cameras, reduces deployment and maintenance costs, and improves monitoring efficiency;

[0064] (3) The present invention has made innovations in image data processing. Through image retrieval and annotation technology, images are retrieved and sorted according to target information, and targets are annotated and metadata is added after registration and fusion. Finally, the format is selected for storage. This not only improves the processing efficiency of image data, but also enhances the richness of target information. By adding rich spatiotemporal metadata information to each image, the present invention provides an important basis for image management, retrieval and analysis, making it possible to quickly locate and filter target images taken at a specific time, specific area or specific camera, greatly improving the practicality and convenience of large-space monitoring systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings;

[0066] Figure 1 The present invention is a flow chart of the method. DETAILED DESCRIPTION

[0067] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] It should be understood that the terms “include” and “comprising” used in the specification and claims of the present disclosure indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0069] It should also be understood that the terms used in this disclosure are only for the purpose of describing specific embodiments and are not intended to limit the disclosure. As used in this disclosure and claims, the singular forms of "a", "an", and "the" are intended to include the plural forms unless the context clearly indicates otherwise. It should also be further understood that the term "and / or" used in this disclosure and claims refers to any combination of one or more of the associated listed items and all possible combinations, including these combinations.

[0070] like Figure 1 As shown, the fast conversion method for large space positioning based on image recognition includes the following steps:

[0071] Large space image acquisition and deployment, field measurement of large space dimensions and obstacle locations, importing software to build basic 3D models, simulating the initial camera locations, determining the final layout after evaluation and screening, and installing and debugging; through multi-camera arrays, based on the structural characteristics of large spaces, using spatial topology algorithms to plan camera installation locations, ensuring all-round coverage without blind spots;

[0072] Technical personnel carry measuring equipment to conduct field measurements of large spaces, record length, width, and height data, draw spatial plans, and mark the locations of doors, windows, columns, and partitions. The dimension data and obstacle location information obtained from field measurements are imported into spatial topology modeling software, including AutoCAD Map 3D and SketchUp with topology analysis plug-ins. Based on the data, a basic three-dimensional model of the large space is constructed to restore the spatial structure and provide a visual carrier for subsequent algorithm operations.

[0073] The initial camera installation points are simulated on the spatial topological model with the help of the camera layout planning algorithm; the algorithm takes full coverage, no blind spots, and reasonable image overlap as optimization goals, inputs the large space size, obstacle distribution, camera field of view, and focal length parameters, and outputs a batch of candidate installation point combinations after multiple rounds of iterative calculations; specifically: first, determine the camera layout planning algorithm, including genetic algorithm or simulated annealing algorithm; input the size information of the large space, the marked distribution data of doors, windows, columns, partition obstacles, and the camera's field of view, focal length and other parameters into the algorithm; the algorithm randomly generates an initial camera installation site combination on the spatial topological model based on the requirements of full coverage and no blind spots, as the first generation of seeds The algorithm calculates the coverage of the camera's field of view under each site combination, evaluates its coverage rate, whether there are blind spots in monitoring, and whether the degree of overlap is reasonable, and assigns a fitness value (genetic algorithm) or an objective function value (simulated annealing algorithm) based on these indicators. Through multiple rounds of iterations, the genetic algorithm performs selection, crossover, and mutation operations, and the simulated annealing algorithm accepts inferior solutions based on probability, and continuously optimizes the camera installation site combination until the preset number of iterations is reached or the convergence condition is met. Finally, a batch of better candidate installation site combinations are output, so as to achieve the goal of all-round monitoring coverage of a large space without blind spots with a reasonable degree of overlap.

[0074] From the candidate solutions, the final layout plan is determined based on the evaluation and screening of the field of view coverage, regional clarity, installation cost, and convenience of later maintenance. The specific process is as follows:

[0075] Comprehensive analysis is conducted through the evaluation indicators of candidate solutions, including:

[0076] Field of view coverage: Calculate the percentage of the spatial volume covered by the camera field of view in each candidate solution to the entire large space volume. This is achieved by dividing the large space into tiny cubic units and counting whether each unit is covered by at least one camera field of view. The field of view coverage is recorded as sf. Regional clarity: For the field of view of each camera, calculate the average resolution ra, which is estimated based on the resolution parameters of the camera and its imaging effect at different distances. Count the clarity scores around entrances and exits and important equipment, and calculate them through a clarity evaluation algorithm (gradient-based clarity algorithm) to obtain a clarity score between 0 and 1, which is recorded as q. p; Installation cost: Calculate the total cost of each candidate solution and record it as zf, including the camera equipment procurement cost, installation material cost (such as brackets, cables, etc.), and installation labor cost. The equipment procurement cost is determined according to the model and quantity of the camera, and the installation material and labor costs are estimated according to the market price and installation workload; Post-maintenance convenience: Count the accessibility scores of all cameras in each candidate solution, and calculate the mean accessibility score, recorded as hb. The accessibility score is higher when the camera is installed in an easily accessible location (such as close to the channel, with a maintenance platform, etc.), otherwise it is lower. The accessibility score is set to a value between 0 and 1;

[0077] The obtained field of view coverage sf, average resolution ra, clarity score qp, total cost zf and accessibility score mean hb are normalized and entered into the following formula: The optimal evaluation value ZGM is obtained, and the optimal evaluation value ZGM is used as the evaluation standard for measuring candidate solutions. are the preset weight coefficients of field of view coverage sf, average resolution ra, clarity score qp, total cost zf and accessibility score mean hb respectively; the merit values ​​ZGM calculated for all candidate schemes are sorted according to size, and the candidate scheme with the largest merit value ZGM is selected as the final layout scheme.

[0078] Use visualization tools to demonstrate the camera field of view coverage of each solution; review the blind spot range in large spaces to determine whether additional cameras are needed; arrange construction teams to install cameras at predetermined locations according to the final layout plan and the cameras added in the blind spots, and adjust the camera angles to the initial angles; debug each camera after installation to check image acquisition and transmission functions; test night vision and strong light suppression effects under different lighting conditions and time periods; fine-tune the camera shooting angle, focal length, and white balance parameters to make the picture clear and the color reproduction accurate; debug multiple cameras in a linked manner to ensure smooth switching and natural field of view connection, so as to achieve all-round monitoring without blind spots.

[0079] Spatial positioning modeling: pre-processing and enhancement after receiving the image stream, using camera parameters and triangulation principle to locate the target, and outputting the position information through multi-camera collaboration and Kalman filter optimization;

[0080] Continuously receive the image streams sent back in real time from the cameras deployed in the large-space image acquisition, sort them according to the acquisition time and camera number information, and remove abnormal frames caused by network jitter and instantaneous strong light interference; use adaptive filtering algorithm to dynamically adjust filtering parameters, remove image noise, and restore clear images; perform image enhancement processing simultaneously, use CLAHE (contrast-limited adaptive histogram equalization) to improve the local contrast of the image, highlight the contour details of the target object, and make subsequent recognition more accurate;

[0081] With the help of the internal and external parameters of the camera recorded by the large-space image acquisition deployment (internal parameters such as focal length and pixel size; external parameters are camera position and posture), combined with the extracted two-dimensional coordinate information of the target in the image, the position of the target in three-dimensional space is reversed using the principle of triangulation; this preliminary position information is associated with the basic three-dimensional model coordinate system, and the target is preliminarily anchored within the model framework; a multi-camera collaborative positioning mechanism is introduced; the same target image captured by multiple cameras with different perspectives is integrated, and the three-dimensional position of the target is calculated separately, and then multiple sets of coordinate data are fused through weighted averaging and least squares optimization to locate the spatial position of the target in the three-dimensional model; the weight distribution is dynamically adjusted based on the distance between the camera and the target, the image clarity, and the perspective stability factors;

[0082] Kalman filtering technology is used to combine the target's historical positioning trajectory and motion state estimation information to correct the current positioning result in real time. Based on the assumption that the target moves at a uniform speed or uniform acceleration, Kalman filtering continuously iterates and updates the target position and speed parameters, smooths the positioning jitter caused by image noise and temporary target occlusion, and improves positioning continuity and accuracy. The final accurately positioned target position information is output in a standardized format (such as longitude and latitude, three-dimensional Cartesian coordinates, relative coordinates relative to a specific area, etc.) and connected to other application systems (such as security monitoring platforms, indoor navigation software, equipment operation and maintenance scheduling systems). At the same time, target identification tags and real-time status information (stationary, moving speed and direction) are added to enrich the positioning data content.

[0083] Target attribution judgment: obtain camera parameters to calculate the field of view boundary, convert the target coordinates to determine whether it is within the field of view, and determine the camera priority based on multiple factors when processing overlapping areas;

[0084] Obtain the internal parameters (such as focal length, pixel size, optical center position) and external parameters (camera position and posture in space, including translation and rotation information) of each camera in the large-space image acquisition deployment. These parameters are the basis for defining the camera's field of view; use camera calibration technology to place calibration objects of known size and position (such as a chessboard) in a large space, take multiple sets of images to calculate and verify the camera's internal and external parameters to ensure the accuracy of the parameters;

[0085] According to the internal and external parameters of the camera, the perspective projection principle is used to calculate the field of view boundary of each camera in three-dimensional space; in the three-dimensional model coordinate system, the far and near planes of the field of view and the boundary angles in the horizontal and vertical directions are determined by geometric calculation; for each camera, its field of view range is represented as a pyramid in three-dimensional space, whose vertex is the center of the camera lens and the bottom surface is a polygon formed by the farthest boundary of the field of view; the position coordinates of the target in the three-dimensional model determined by the spatial positioning modeling are converted to the same coordinate system as the camera field of view range; specifically, the coordinate transformation matrix is ​​used, which is calculated based on the external parameters of the camera; ensuring the coordinate system of the target position coordinates and the camera field of view range is a key step, otherwise it is impossible to accurately determine whether the target is within the camera field of view;

[0086] For the converted target position coordinates, determine whether it is located inside the pyramid of the camera's field of view, specifically through the position relationship between the point and the plane (such as the distance and direction from the target position to the field of view boundary plane) and angle judgment (whether the horizontal and vertical angles of the target position relative to the center of the camera lens are within the field of view angle range); use spatial geometry algorithms, such as the ray and plane intersection algorithm, to emit rays from the center of the camera lens to the target position, and determine whether the rays intersect with the field of view boundary plane, and whether the intersection position meets the definition of the field of view range, so as to determine whether the target is within the camera's field of view;

[0087] Since there may be overlapping areas in the camera layout, it is necessary to identify these overlapping areas and calculate the boundaries and ranges of the overlapping areas by comparing the intersections of pyramids in different camera field of view in three-dimensional space. Use spatial geometry operation libraries (such as CGAL-ComputationalGeometryAlgorithmsLibrary) to calculate the intersections between pyramids and determine the position and shape of the overlapping areas in the three-dimensional model.

[0088] When the target is in the overlapping area, determine the priority of the camera to which it belongs, and make a comprehensive judgment based on factors such as the distance between the camera and the target (cameras with closer distance have higher priority), the importance of the camera (such as cameras in key areas have higher priority), and image quality (cameras with high resolution and good lighting conditions have higher priority). Establish a priority evaluation function, use the above factors as parameters for weighted calculation, and calculate a comprehensive priority score based on the score of each camera on these factors. The camera with the highest score is the camera to which the target belongs in the overlapping area. When the target is in the overlapping area, for camera i (i=1,2...,n, n is the number of cameras involved in the overlapping area), its comprehensive priority score The calculation formula is as follows: ,in represents the comprehensive priority score of camera i. The higher the score, the higher the priority of the camera in target attribution judgment. They are distance weight, camera importance weight, image quality weight, and ;

[0089] Represents the distance from camera i to the target, which is obtained by calculating the spatial coordinates. Assume that the three-dimensional coordinates of the camera are , the three-dimensional coordinates of the target are ,but ; In order to give the closer camera an advantage in weighted calculation, use The form participates in the calculation, that is, the closer the distance, The larger the value of ;

[0090] Indicates the importance score of camera i, with a value range of 0-1. The importance level of each camera is set manually according to the functional zoning of the large space, including cameras in key areas such as key passages, entrances and exits, and above valuable equipment. ; In relatively minor areas, the values ​​are assigned in descending order according to the importance. At the same time, the fuzzy evaluation method or expert scoring method can also be used to determine the importance of each camera. value;

[0091] Indicates the image quality score of camera i, and the value range is also 0-1. The calculation method is .in is the resolution normalized score of camera i, is the normalized score of light conditions, and are the weights of resolution and light conditions in image quality assessment, respectively, and ; For resolution normalized score :Assume the horizontal resolution of camera i is , the vertical resolution is , the highest horizontal resolution of all cameras in the large space is , the maximum vertical resolution is ,but , ensuring that cameras with higher resolutions get higher scores; scores are normalized for lighting conditions : Use the light sensor to monitor the light intensity at the camera i in real time , and obtain the maximum light intensity recorded by the camera in the entire large space and minimum value ,but , that is, the closer the light intensity is to the maximum value, The closer it is to 1, the better the lighting conditions are; calculate the comprehensive priority score of all cameras After that, select The camera with the largest value is taken as the camera to which the target belongs in the overlapping area.

[0092] Repeat the above steps at a certain frequency (such as several times per second, determined according to the requirements of the actual application scenario and system performance) to match the target position with the camera field of view. Use an event-driven mechanism to immediately trigger the target attribution judgment process when the target position changes by more than a certain threshold (such as moving a certain distance or entering a new area) or the camera parameters change significantly, to ensure that the camera to which the target belongs can be determined in a timely and accurate manner.

[0093] Image retrieval and fusion annotation: retrieve and sort images according to target information, annotate targets and add metadata after registration and fusion, and finally select format for storage;

[0094] According to the target information determined by the target attribution judgment and the camera identification associated with it, all the original image data that have been photographed by the target are quickly retrieved in the distributed storage system; these image data may be scattered and stored in the storage nodes or storage partitions corresponding to different cameras; the retrieved images are sorted and organized according to the time sequence of image acquisition, and the image time series set of the target is constructed for subsequent analysis of the target's movement trajectory and behavior changes to ensure the consistency and logic of the image presentation. For example, if the target is a customer moving in a large shopping mall, arranging the images in chronological order can clearly show its path from entering the mall to various areas;

[0095] For target images from different camera perspectives, there are geometric distortions and position deviations between images due to differences in shooting angles, distances and other factors. Use image registration algorithms, such as registration methods based on feature point matching (feature point extraction algorithms such as SIFT and SURF combined with the RANSAC algorithm to eliminate mismatches) or registration algorithms based on mutual information to determine the correspondence between targets in different images, achieve geometric alignment of images, and make the positions and shapes of targets in different images accurately correspond; use multi-resolution image fusion technology, such as a fusion algorithm based on Laplace pyramid decomposition. Perform Laplace pyramid decomposition on the registered images to obtain image representations at different resolution levels. Then, at each resolution level, according to certain fusion rules (such as weighted averaging, selection strategies based on regional energy or contrast), fuse the image information from different cameras, and combine the advantages of each perspective to generate a fused multi-resolution image sequence. This can not only retain the detailed information of the target, but also improve the clarity and recognition of the overall image, especially when the target is in a complex background or partially blocked, it can better present the overall picture of the target;

[0096] On the fused image, the target is accurately annotated using the target recognition and positioning results. The annotation content includes the target type (such as personnel, vehicles, equipment, etc.), the target unique identifier (if any), the target position bounding box information in the image, etc. This helps to quickly identify and locate the position of the target in the image in the future, and facilitates image analysis and query. For example, for customer images in a shopping mall, the customer's general outline, gender, clothing characteristics and other information are annotated to facilitate security personnel or shopping mall operators to track people and analyze their behavior; rich spatiotemporal metadata information is added to each image, such as image acquisition time (accurate to milliseconds), acquisition location (corresponding to the specific coordinate position or area identification in a large space), camera number, image resolution, frame rate, etc. These metadata information provides an important basis for image management, retrieval and analysis, so that target images taken at a specific time, specific area or specific camera can be quickly located and filtered; the image sequence after the annotation and metadata are added is stored in a suitable storage format. Common storage formats include JPEG2000, PNG, etc. These formats have a good balance between image compression ratio and image quality, and can effectively reduce storage space while ensuring the integrity of image information. Images are organized and stored according to a certain directory structure and naming rules. For example, the target type and target identifier are used as the first-level directory, and the acquisition time is used as the sub-directory. Images are stored in an orderly manner to facilitate subsequent management and access.

[0097] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A fast conversion method for large space positioning based on image recognition, characterized in that: The following steps are involved: B1: Large space image acquisition and deployment, on-site measurement of large space dimensions and obstacle locations, importing software to build a basic 3D model, simulating the initial position of the camera, determining the final layout plan after evaluation and screening, and installing and debugging the camera; The process of B1 is as follows: Technicians carry measuring equipment to conduct on-site measurements of large spaces, record length, width, and height data, draw spatial plans, and mark the locations of doors, windows, columns, partitions, and fixed obstacles; Import the dimension data and obstacle location information obtained from field measurements into spatial topology modeling software, including AutoCAD Map 3D and SketchUp with topology analysis plug-ins; construct a basic three-dimensional model of the large space based on the data, restore the spatial structure, and provide a visual carrier for subsequent algorithm operations; The initial camera installation points are simulated on the spatial topology model using the camera layout planning algorithm. The algorithm takes full coverage, no blind spots, and reasonable image overlap as optimization goals. It inputs large space dimensions, obstacle distribution, camera field of view, and focal length parameters. After multiple rounds of iterative calculations, it outputs several candidate installation point combinations. Specifically: First, determine the camera layout planning algorithm, including genetic algorithm or simulated annealing algorithm; input the size information of the large space, the distribution data of marked doors, windows, columns, partition obstacles, and the field of view and focal length parameters of the camera into the algorithm; the algorithm randomly generates the initial camera installation site combination on the spatial topology model as the initial population or initial solution based on the requirements of full coverage and no blind spots; then calculate the camera's field of view coverage under each site combination, evaluate its coverage rate, whether there are monitoring blind spots, and whether the degree of image overlap meets the standards, and assign fitness values ​​or objective function values ​​based on these indicators; Through multiple rounds of iterations, the genetic algorithm performs selection, crossover, and mutation operations, and the simulated annealing algorithm accepts inferior solutions based on probability, continuously optimizing the camera installation point combination until the preset number of iterations is reached or the convergence condition is met, and finally outputs a number of candidate installation point combinations; From the candidate solutions, the final layout plan is determined based on the evaluation and screening of the visual field coverage, regional clarity, installation cost, and convenience of later maintenance indicators; Use visualization tools to display the camera's field of view coverage; review the blind spots in large spaces to determine whether additional cameras are needed; Arrange the construction team to install cameras at the predetermined locations according to the final layout plan and the cameras added in the blind spots, and adjust the camera angles to the initial angles; After installation, debug each camera one by one to check image acquisition and transmission functions; test night vision and strong light suppression effects under different lighting conditions and time periods; adjust the camera shooting angle, focal length, and white balance parameters to make the picture clear and the color restoration accurate; debug multiple cameras in conjunction to ensure smooth switching and natural field of view connection, so as to achieve all-round monitoring without blind spots; The specific steps for determining the final layout solution in B1 are as follows: Comprehensive analysis is conducted through the evaluation indicators of candidate solutions, including: Field of view coverage: Calculate the percentage of the spatial volume covered by the camera field of view in each candidate solution to the entire large space volume. This is achieved by dividing the large space into tiny cubic units and counting whether each unit is covered by at least one camera field of view. The field of view coverage is recorded as sf. Regional clarity: For each camera’s field of view, calculate the average resolution ra, which is estimated based on the camera’s resolution parameters and its imaging effects at different distances; calculate the clarity scores around entrances and exits and important equipment using a clarity evaluation algorithm to obtain a clarity score between 0 and 1, which is recorded as qp; Installation cost: Calculate the total cost of each candidate solution and record it as zf, including the camera equipment procurement cost, installation material cost, and installation labor cost. The equipment procurement cost is determined according to the model and quantity of the camera, and the installation material and labor costs are estimated based on the market price and installation workload; Maintenance convenience: The accessibility scores of all cameras in each candidate solution are counted, and the mean accessibility score is calculated, denoted as hb. The accessibility score is higher if the camera is installed in an easily accessible location, and lower if it is installed in an easily accessible location. The accessibility score is set to a value between 0 and 1. The obtained field of view coverage sf, average resolution ra, clarity score qp, total cost zf and accessibility score mean hb are normalized and entered into the following formula: The optimal evaluation value ZGM is obtained, and the optimal evaluation value ZGM is used as the evaluation standard for measuring candidate solutions. They are the preset weight coefficients of the field of view coverage sf, average resolution ra, clarity score qp, total cost zf and accessibility score mean hb; Sort the ZGM values ​​calculated for all candidate solutions by size, and select the candidate solution with the largest ZGM value as the final layout solution; B2: Spatial positioning modeling: After receiving the image stream, it performs preprocessing and enhancement, locates the target using camera parameters and triangulation principles, and outputs position information through multi-camera collaboration and Kalman filter optimization; B3: Target attribution judgment, obtain camera parameters to calculate the field of view boundary, convert the target coordinates to determine whether it is within the field of view, and determine the camera priority based on multiple factors when processing overlapping areas; B4: Image retrieval and fusion annotation: retrieve and sort images according to target information, annotate targets and add metadata after registration and fusion, and finally select format for storage.

2. The fast conversion method for large space positioning based on image recognition according to claim 1 is characterized in that: The specific operation steps of B2 are as follows: Continue to receive the image streams sent back in real time by each camera in step B1, sort them according to the acquisition time and camera number information, and remove abnormal frames caused by network jitter and instantaneous strong light interference; Adaptive filtering algorithm is used to dynamically adjust filtering parameters, remove image noise, and restore clear images. Image enhancement processing is performed simultaneously, and CLAHE is used to improve image contrast and highlight the contour details of the target object. With the help of the camera internal and external parameters recorded in step B1, combined with the extracted two-dimensional coordinate information of the target in the image, the position of the target in three-dimensional space is inferred using the principle of triangulation; this preliminary position information is associated with the basic three-dimensional model coordinate system, and the target is preliminarily anchored in the model framework; Introduce a multi-camera collaborative positioning mechanism; combine the same target images captured by multiple cameras with different perspectives, calculate the target's three-dimensional position respectively, and then fuse multiple sets of coordinate data through weighted averaging and least squares optimization to locate the target's spatial position in the three-dimensional model; The weight distribution is dynamically adjusted based on the distance between the camera and the target, the image clarity, and the viewing angle stability. Using Kalman filtering technology, combined with the target's historical positioning trajectory and motion state estimation information, the current positioning result can be corrected in real time; Based on the assumption that the target moves at a uniform speed or uniform acceleration, the Kalman filter continuously iterates and updates the target position and speed parameters, smoothing the positioning jitter caused by image noise and temporary occlusion of the target, thereby improving the continuity and accuracy of positioning. The final accurately located target location information is output in a standardized format and connected to other application systems; at the same time, target identification tags and real-time status information are added to enrich the positioning data content.

3. The fast conversion method for large space positioning based on image recognition according to claim 1 is characterized in that: The specific operation steps of B3 are as follows: Obtain the internal and external parameters of each camera in step B1, which are used as the basis for defining the camera's field of view; use camera calibration technology to calculate and verify the internal and external parameters of the camera by placing calibration objects of known size and position in a large space and taking multiple sets of images to ensure the accuracy of the parameters; According to the internal and external parameters of the camera, the perspective projection principle is used to calculate the field of view boundary of each camera in three-dimensional space; In the three-dimensional model coordinate system, the far and near planes of the field of view and the boundary angles in the horizontal and vertical directions are determined through geometric calculations; for each camera, its field of view is represented as a pyramid in three-dimensional space, with its vertex being the center of the camera lens and its bottom being a polygon formed by the farthest boundary of the field of view; The position coordinates of the target in the three-dimensional model determined in step B2 are converted to a coordinate system that is the same as the camera's field of view; specifically, a coordinate transformation matrix is ​​used, which is calculated based on the camera's external parameters; For the converted target position coordinates, determine whether it is inside the pyramid within the camera's field of view, specifically by determining the position relationship between the point and the plane and the angle. Using spatial geometry algorithms, including ray and plane intersection algorithms, rays are emitted from the center of the camera lens to the target position to determine whether the rays intersect with the boundary plane of the field of view and whether the intersection position meets the definition of the field of view, so as to determine whether the target is within the field of view of the camera; Since there are overlapping areas in the camera layout, it is necessary to identify these overlapping areas and calculate the boundaries and ranges of the overlapping areas by comparing the intersections of pyramids in different camera field of view in three-dimensional space. The spatial geometry operation library is used to calculate the intersections between pyramids to determine the position and shape of the overlapping areas in the three-dimensional model. When the target is in the overlapping area, the priority of the camera to which it belongs is determined, and a comprehensive judgment is made based on the distance between the camera and the target, the importance of the camera, and image quality factors; Establish a priority evaluation function, use the above factors as parameters for weighted calculation, and calculate a comprehensive priority score based on the scores of each camera on these factors. The camera with the highest score is the camera to which the target belongs in the overlapping area. Repeat the above steps at a preset frequency to determine the match between the target position and the camera field of view; An event-driven mechanism is adopted. When the target position changes beyond the threshold or the camera parameters change significantly, the target ownership judgment process is immediately triggered to ensure that the camera to which the target belongs can be judged in a timely and accurate manner.

4. The rapid conversion method for large space positioning based on image recognition according to claim 3 is characterized in that: The analysis of the comprehensive priority score in B3 is as follows: When the target is in the overlapping area, for camera i, i=1,2...,n, n is the number of cameras involved in the overlapping area, its comprehensive priority score is The calculation formula is as follows: ,in represents the comprehensive priority score of camera i; They are distance weight, camera importance weight, image quality weight, and ; Represents the distance from camera i to the target, which is obtained by calculating the spatial coordinates. Let the three-dimensional coordinates of the camera be , the three-dimensional coordinates of the target are ,but ;use The form participates in the calculation, that is, the closer the distance, The larger the value of ; Represents the importance score of camera i, with a value range of 0-1. The importance level of each camera is manually set according to the functional zoning of the large space; Indicates the image quality score of camera i, also in the range of 0-1; the calculation method is ;in is the resolution normalized score of camera i, is the normalized score of light conditions, and are the weights of resolution and light conditions in image quality assessment, respectively, and ; For resolution normalized score :Assume the horizontal resolution of camera i is , the vertical resolution is , the highest horizontal resolution of all cameras in a large space is , the maximum vertical resolution is ,but , to ensure that the higher the resolution of the camera, the higher the resolution of the normalized score The higher; Normalized scores for light conditions : Use the light sensor to monitor the light intensity at the camera i in real time , and obtain the maximum light intensity recorded by the camera in the entire large space and minimum value ,but , that is, the greater the light intensity, The larger it is, the better the lighting conditions are; Calculate the overall priority score of all cameras After that, select The camera with the largest value is taken as the camera to which the target belongs in the overlapping area.

5. The fast conversion method for large space positioning based on image recognition according to claim 1 is characterized in that: The specific operation steps of B4 are as follows: According to the target information determined in step B3 and the camera identification associated therewith, all the original image data of the target are retrieved in the distributed storage system; these image data are dispersedly stored in the storage nodes or storage partitions corresponding to different cameras; Sort and organize the retrieved images according to the time sequence of image acquisition, and construct the target's image time series set for subsequent analysis of the target's movement trajectory and behavior changes to ensure the coherence and logic of image presentation; For target images from different camera perspectives, there are geometric distortions and position deviations between images due to differences in shooting angles and distances. Use a registration method based on feature point matching or a registration algorithm based on mutual information to determine the correspondence between targets in different images and achieve geometric alignment of images, so that the positions and shapes of targets in different images can accurately correspond. Adopt multi-resolution image fusion technology, including fusion algorithm based on Laplace pyramid decomposition; perform Laplace pyramid decomposition on the registered images to obtain image representations at different resolution levels; Then, at each resolution level, according to the fusion rules, including weighted average, selection strategy based on regional energy or contrast, the image information from different cameras is fused, and the advantages of each perspective are combined to generate a fused multi-resolution image sequence; On the fused image, the target is annotated using the target recognition and positioning results; the annotation content includes the target type, the target unique identifier, and the target position boundary box information in the image; Used for subsequent rapid identification and location of the target in the image; Add rich spatiotemporal metadata information to each image, including image acquisition time, acquisition location, camera number, image resolution, and frame rate; Metadata information provides an important basis for image management, retrieval and analysis, making it possible to locate and filter target images when querying time, area or camera images; Select the storage format to store the image sequence after adding annotations and metadata, including JPEG2000 and PNG; organize and store the images according to the directory structure and naming rules.

Images