A multi-camera cooperative positioning method based on a graph model

Abstract

This invention belongs to the field of image photometry, specifically involving a multi-camera collaborative localization method based on a graph model. The method includes the following steps: constructing a stereo vision observation network using multiple cameras to cover the area to be observed, and calibrating the intrinsic parameters of each camera; obtaining the image coordinates of the target using a target detection network, and extracting the target's appearance features using a re-identification network; calculating the target's geometric features using the calibrated intrinsic and extrinsic parameters; generating candidate spatial locations by intersecting pairs of lines of sight using the geometric features and extrinsic parameters; constructing a graph interconnecting different camera nodes using targets detected by the cameras as nodes; establishing a cost function and minimizing the cost function to obtain the optimal state estimate; collecting all nodes of the same target, and calculating the final three-dimensional coordinates of the target using the generalized weighted midpoint method combined with the observation data from multiple cameras. This invention achieves high-precision, real-time, and robust three-dimensional spatial localization of multiple targets.

Description

Technical Field

[0001] This invention belongs to the field of image photometry, and specifically relates to a multi-camera cooperative localization method based on a graph model. Background Technology

[0002] With the rapid development of applications such as intelligent monitoring, autonomous driving, and robot collaboration, obtaining high-precision and robust multi-target 3D position information in complex scenarios has become one of the core challenges in the field of visual measurement. Traditional monocular vision localization is limited by depth ambiguity and scale drift, making it difficult to meet the needs of large-scale, multi-target, and long-term observation. While simply increasing the number of cameras can introduce geometric redundancy, it also brings challenges to cross-view data correlation.

[0003] Existing technologies mainly follow two lines of thought:

[0004] The approach involves first associating and then triangulating the images. This is done by establishing target correspondences between images based on appearance features (such as Re-ID and color histograms) or motion consistency, and then using multi-view geometry to solve for the 3D coordinates. However, this method is extremely sensitive to scenes with similar appearances, frequent occlusion, or abrupt changes in lighting. If the association is incorrect, the accuracy of subsequent triangulation cannot be guaranteed.

[0005] The method first triangulates and then correlates the detection rays, directly intersecting each ray pairwise to generate a large number of spatial candidate points. Then, clustering or voting is performed based on indicators such as reprojection error. However, when rays do not intersect at a single point due to detection noise, calibration errors, or occlusion, the simple intersection strategy introduces systematic bias.

[0006] Furthermore, existing methods generally treat location calculation and association matching separately, lacking a unified optimization objective, leading to error accumulation; under complex conditions such as dynamic occlusion and motion blur, positioning accuracy drops significantly. Therefore, there is an urgent need for a multi-camera cooperative positioning scheme that can simultaneously utilize geometric and appearance information and can scale linearly with the number of cameras. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a multi-camera collaborative localization method based on a graph model. Through a two-level strategy of "single-camera detection - multi-camera localization", a graph model with interconnected nodes across viewpoints is constructed and a geometry-appearance joint cost function is established. After global optimization allocation and weighted fusion, high-precision, real-time and robust three-dimensional spatial localization of multiple targets is achieved.

[0008] This invention provides a multi-camera cooperative localization method based on a graph model, comprising the following steps:

[0009] Step S1: Construct a stereo vision observation network consisting of at least two cameras to cover the area to be observed, and calibrate the intrinsic and extrinsic parameters of each camera;

[0010] Step S2: Each camera performs single-camera detection, obtains the two-dimensional coordinates of the target in the image through the target detection network, and extracts the appearance features of the target through the re-identification network;

[0011] Step S3: Based on the camera intrinsic and extrinsic parameters obtained from the calibration, convert the two-dimensional coordinates into geometric features representing the spatial orientation of the target.

[0012] Step S4: Based on geometric features and camera extrinsic parameters, perform pairwise intersection calculations on the lines of sight of different cameras to generate multiple candidate spatial locations;

[0013] Step S5: Using the targets detected by each camera as nodes, construct a graph model, in which a connecting edge is established between any two nodes from different cameras, and nodes from the same camera are not connected to each other.

[0014] Step S6: Establish a joint cost function that includes both node cost and edge cost. The node cost is used to measure the geometric consistency between the geometric features of a node and the assigned candidate spatial position. The edge cost is used to measure the similarity between two connected nodes in terms of appearance and geometric features. By minimizing the joint cost function, the corresponding candidate spatial position is assigned to each node to form the optimal state estimate.

[0015] Step S7: Identify all nodes assigned to the same candidate spatial location as the same target, collect multi-camera observation data corresponding to these nodes, and calculate the final three-dimensional coordinates of the target under test through weighted fusion.

[0016] The beneficial effects of this invention are:

[0017] 1. Complete cross-view data association and target 3D coordinate calculation within the same framework, reduce association error rate and suppress error accumulation, and keep the positioning error in complex scenarios at the decimeter level.

[0018] 2. By using a strategy of ray generation and clustering, the number of candidate spatial points is reduced to be approximately linearly related to the number of cameras and targets, enabling real-time processing of multi-camera data.

[0019] 3. By utilizing the appearance-geometric joint cost function and generalized weighted midpoint fusion, it has natural robustness against detection noise, calibration error and partial occlusion, and maintains trajectory continuity even when the target is temporarily occluded, thus improving system stability. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the generation of candidate spatial positions in this invention;

[0021] Figure 2 This is a schematic diagram of the graphical model structure of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0024] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0025] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection, an electrical connection, a physical connection, or a wireless communication connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two elements or the interaction between two elements, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0026] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0027] like Figures 1-2 As shown, this invention provides a multi-camera cooperative localization method based on a graph model, comprising the following steps:

[0028] Step S1: Construct a stereo vision observation network consisting of at least two cameras to cover the area to be observed, and calibrate the intrinsic and extrinsic parameters of each camera; establish the mapping relationship between two-dimensional image coordinates and three-dimensional spatial coordinates through calibration, lay the geometric foundation for converting the image detection results into spatial rays in the future, and ensure positioning accuracy.

[0029] Step S2: Each camera performs single-camera detection, obtains the two-dimensional coordinates of the target in the image through the target detection network, and extracts the appearance features of the target through the re-identification network. The target detection network provides the position information of the target in the image as input for geometric calculation. The appearance features extracted by the re-identification network provide semantic information for cross-camera target matching, solving the problem of easy confusion in similar target scenes by traditional geometric methods.

[0030] Step S3: Based on the camera intrinsic and extrinsic parameters obtained from calibration, the two-dimensional coordinates are converted into geometric features representing the spatial orientation of the target; the two-dimensional image coordinates are upgraded to three-dimensional spatial rays, so that observations from different cameras can be compared and fused under a unified geometric framework, achieving dimensional leap from image detection to spatial positioning.

[0031] Step S4: Based on geometric features and camera extrinsic parameters, perform pairwise intersection calculations on the lines of sight of different cameras to generate multiple candidate spatial locations; generate candidate points by pairwise ray intersection and cluster them to reduce dimensionality, control the number of candidate spatial locations within a manageable range, avoid combinatorial explosion problem, and provide a solution space of controllable scale for subsequent graph optimization allocation.

[0032] Step S5: Using the targets detected by each camera as nodes, construct a graph model where nodes from any two different cameras are connected by edges, while nodes from the same camera are not connected to each other. Abstract the multi-camera detected targets into a graph structure. Through a specific topological structure of cross-view fully connected and single-view disconnected, the data association problem is transformed into a graph optimization problem, which is convenient for subsequent unified solution.

[0033] Step S6: Establish a joint cost function that includes both node cost and edge cost. The node cost measures the geometric consistency between the geometric features of a node and its assigned candidate spatial location, while the edge cost measures the similarity between two connected nodes in terms of appearance and geometric features. By minimizing the joint cost function, corresponding candidate spatial locations are assigned to each node, forming the optimal state estimate. By incorporating geometric consistency and appearance similarity into the same optimization framework, the synergistic effect of the two is achieved through joint optimization. Compared with the serial processing strategy, this approach can more accurately handle data association problems in complex scenarios.

[0034] Step S7: Identify all nodes assigned to the same candidate spatial location as the same target, collect multi-camera observation data corresponding to these nodes, and calculate the final three-dimensional coordinates of the target under test through weighted fusion; based on the assignment results, fuse multi-view observation data, suppress the influence of detection noise and calibration error through weighted averaging, and output high-precision three-dimensional spatial coordinates.

[0035] This invention completes cross-view data association and target 3D coordinate calculation within the same framework. Through graph model construction, joint cost function optimization, and generalized weighted fusion, it achieves the following beneficial effects:

[0036] (1) High-precision positioning minimizes the impact of factors such as detection noise, calibration error and obstruction on positioning accuracy, and the error can be controlled at the decimeter level;

[0037] (2) Real-time processing capability: Through the three-stage framework of "candidate generation-clustering-optimized allocation", the computational complexity is reduced from exponential to linear.

[0038] (3) It is robust and can maintain the continuity of target matching even if there are missed detections, false detections or short-term occlusions.

[0039] (4) It has good versatility and can be widely used in various scenarios such as drone swarm recovery, intelligent video surveillance, and autonomous driving collaborative perception.

[0040] In one embodiment, in step S3, the two-dimensional coordinates are converted into geometric features representing the spatial orientation of the target as follows:

[0041] For any camera The geometric features of any detected target Calculated using the following formula:

[0042] ;

[0043] in, For camera The intrinsic parameter matrix, For camera The rotation matrix represents the absolute rotation of the camera coordinate system relative to the world coordinate system. Let be the homogeneous coordinates of the target object in the image;

[0044] Through the intrinsic parameter matrix and rotation matrix The inverse transformation transforms the homogeneous coordinates of the image. The transformation is precisely converted into a spatial ray direction vector in the world coordinate system. This transformation is based on a rigorous camera imaging geometry model, ensuring the mapping accuracy from 2D detection to 3D space, and providing accurate geometric input for subsequent candidate location generation and graph optimization allocation.

[0045] Geometric features This represents the direction vector of the spatial ray pointing from the camera's optical center towards the target. Geometric features Expressing the ray direction vector concisely simplifies the complex camera imaging model into a unified vector representation. Compared to using the original image coordinates or projection matrix for calculations, this representation can be directly used in subsequent pairwise ray intersections, common perpendicular segment calculations, and nodal cost functions, avoiding repetitive coordinate transformations and reducing computational complexity.

[0046] In one embodiment, the specific method for generating multiple candidate spatial locations in step S4 includes:

[0047] Based on the calibration parameters of each camera, the geometric features corresponding to each target are... The mapping is a spatial ray originating from the camera's optical center, and the world coordinates of the camera's optical center are represented as follows:

[0048] ;

[0049] in, For camera The coordinates of the optical center in the world coordinate system. This represents the absolute translation of the camera coordinate system relative to the world coordinate system;

[0050] This step follows the camera calibration results from step S1 and the geometric feature calculations from step S3, using the extrinsic parameters obtained from the calibration to calculate the geometric features. ;

[0051] This is mapped to a spatial ray originating from the camera's optical center. (Using the formula...) The coordinates of the camera's optical center in the world coordinate system are accurately calculated, and together with the direction vector calculated in step S3, they form a complete ray representation. This mapping unifies the observations of each camera to the same world coordinate system, providing a unified reference framework for subsequent cross-camera geometric intersections. It serves as a bridge connecting step S1 calibration, step S2 detection, step S3 geometric transformation, and subsequent spatial position generation.

[0052] For any two different cameras, pairwise intersections are performed, and the midpoint of the common perpendicular segment of the two rays is calculated to obtain a series of candidate points. Due to detection and calibration errors in actual measurements, the lines of sight to the same target from different cameras often do not intersect precisely at a single point. This step obtains the midpoint of the closest position in space between the two lines of sight by calculating the midpoint of the common perpendicular segment, serving as candidate points for that ray pair. Compared to directly solving for the intersection point, the midpoint of the common perpendicular segment is inherently robust to errors, providing stable candidate point positions even in the presence of noise, thus providing reliable input data for subsequent clustering. Simultaneously, pairwise intersections are performed for all different camera combinations to ensure that every pair of rays that might observe the same target generates candidate points, avoiding the possibility of missing the true target.

[0053] Cluster all candidate points and use each cluster center as a candidate spatial location. The same real target will generate multiple similar candidate points (from the intersection results of different camera pairs). This step merges these candidate points through clustering, with each cluster center representing a candidate spatial location. This clustering operation has a dual effect: on the one hand, it significantly reduces the number of candidate locations, reducing the computational complexity from traditional exhaustive search to approximately linear correlation with the number of cameras and targets, creating a feasible computational premise for the graph model construction in step S5 and the optimized allocation in step S6; on the other hand, the cluster center, as a comprehensive representation of the intersection results of multiple ray pairs, has higher reliability and noise resistance than the midpoint of a single ray pair, improving the quality of candidate locations.

[0054] This embodiment realizes the transformation from ray representation to candidate positions. Through three-level processing of ray mapping, midpoint of common perpendicular segment, and clustering, it not only solves the practical problem that rays cannot intersect precisely under error conditions, but also effectively controls the computational scale of subsequent optimization through clustering dimensionality reduction.

[0055] refer to Figure 1 , Figure 1 This is a schematic diagram illustrating the generation of candidate spatial locations in this invention. For example... Figure 1 As shown, each camera emits a spatial ray pointing towards the target from its optical center. The midpoint of the common perpendicular segment of any two rays from different cameras is calculated to obtain a series of candidate points. Then, clustering is used to treat each cluster center as a candidate spatial location. By generating candidate points through the intersection of pairs of rays and reducing dimensionality through clustering, the number of candidate spatial locations is controlled within a manageable range, avoiding the combinatorial explosion problem and providing a solution space of manageable scale for subsequent graph optimization allocation.

[0056] In one embodiment, step S5, the specific method for constructing a graph model includes:

[0057] Each target detected by each camera is treated as a node. Each node contains corresponding appearance features. and node geometry features Node geometric features The geometric features characterizing the target space orientation are used to abstract each target detected in step S2 into a node in the graph model, and the appearance features extracted in step S2 are used to represent the target space orientation. and the node geometric features calculated in step S3 As attribute information of nodes, each node simultaneously carries semantic information (appearance features) and geometric information (ray direction), providing a complete data foundation for the node cost and edge cost in the subsequent joint cost function. Compared to traditional node definitions that only contain location information or only contain appearance information, the node design of this invention can support richer and more accurate matching decisions.

[0058] Set up a camera Detected If there are 1 objective, then the total number of nodes in the graph model is:

[0059] ;

[0060] in, N The total number of cameras;

[0061] The scale of the graphical model is clearly defined by the formula, thus quantifying the total detection results of the multi-camera system. This total number of nodes reflects the number of targets observed by all cameras at the current moment, and is related to the scale of the camera observation network constructed in step S1 (number of cameras). N ) and the single-camera detection capability in step S2 (number of targets detected by each camera) This is directly related to the graph optimization problem. This quantitative expression provides a clear variable space for subsequent graph optimization solutions and also lays the foundation for computational complexity analysis.

[0062] For any two nodes from different cameras, an edge is created between them. Nodes within the same camera are not connected to each other. The total number of edges in the graph model is... for:

[0063] .

[0064] This step defines the connection rules for the graph model, namely a specific topology of "fully connected across viewpoints and disconnected within a single viewpoint." This rule has the following technical effects: First, establishing edges between cross-camera nodes enables graph optimization to discover and match the same target observed from different viewpoints, which is the core task of multi-camera cooperative localization; second, nodes within the same camera are not connected to each other, avoiding invalid matching between different targets within the same image and reducing the number of redundant edges and computational cost; finally, the number of edges is precisely quantified by a formula, significantly reducing computational cost compared to fully connected edges.

[0065] refer to Figure 2 , Figure 2 This is a schematic diagram of the graphical model structure of the present invention. For example... Figure 2 As shown, each target detected by a camera is abstracted as a node, which includes appearance features and node geometry features; nodes between different cameras are fully connected (shown by solid lines in the figure), while nodes within the same camera are not connected to each other. This specific topology of "fully connected across views and disconnected within a single view" transforms the data association problem into a graph optimization problem, facilitating a unified solution later.

[0066] In one embodiment, in step S6, the joint cost function is specifically:

[0067] ;

[0068] in, , , For nodes Assigned candidate spatial location coordinates, and It is a node Connecting nodes Includes appearance features and node geometry features. , , , For node weights, The edge weight is denoted as .

[0069] This step integrates node costs and edge costs into a unified joint cost function through a weighted summation. and The adjustable weighting coefficients are used to balance the importance of geometric consistency constraints and appearance similarity constraints in optimization. Compared to the existing technology of using geometric verification and appearance matching sequentially (such as appearance matching first and then geometric verification for screening, or geometric matching first and then appearance confirmation), the parallel weighted fusion method of this invention can achieve synergistic optimization of the two, avoiding the problem of early decision errors in sequential processing leading to subsequent uncorrectable issues, and improving the accuracy and robustness of data association.

[0070] Furthermore, by using weighted summation with adjustable weights, distance penalties based on negative logarithmic Gaussian distribution, and joint constraints in product form, synergistic optimization of geometric consistency constraints and appearance similarity constraints is achieved. Compared to existing technologies (such as epipolar geometry methods that rely solely on geometry, pure ReID methods that rely solely on appearance, and serial combination methods that are prone to error accumulation), the joint cost function of this invention can more accurately handle data association problems in complex scenarios. It can maintain a high matching accuracy even when targets have similar appearances, geometric ambiguity, or occlusion, providing core technical support for achieving high-precision and robust localization.

[0071] In one embodiment, in step S6, the joint cost function is minimized, and corresponding candidate spatial positions are assigned to each node to form the optimal state estimate Z:

[0072] ;

[0073] in, , Indicates camera The Each detected node is assigned to a candidate space location. .

[0074] In this embodiment, by solving... This invention obtains a node-candidate position allocation scheme that achieves the joint optimality of geometric consistency and appearance similarity. Compared to the serial processing in existing technologies (such as appearance matching followed by geometric verification, or geometric matching followed by appearance confirmation), the global optimization method of this invention avoids the problem of early decision errors in serial strategies leading to subsequent uncorrectable errors, and directly obtains the globally optimal data association result. Simultaneously, it utilizes 0-1 variables... The clear definition provides a clear input for the fusion calculation in step S7, realizing a complete technical loop from "detection" to "association" and then to "localization".

[0075] In one embodiment, step S7 involves collecting multi-camera observation data corresponding to these nodes as follows: For the first... One goal: to collect all The nodes are used to obtain a set of camera position coordinates corresponding to the nodes. and geometric features .

[0076] In this embodiment, by collecting all The nodes are identified, and the corresponding camera position coordinates are extracted. and geometric features This provides accurate and complete input data for the weighted fusion calculation in step S7. This collection method is a key bridge connecting "data association" (step S6) and "3D localization" (step S7), ensuring that the final 3D coordinates of each target can be fused based on all its valid observations. Compared with the existing technology that uses nearest neighbor matching or greedy matching followed by direct fusion, the collection method of this invention is based on the globally optimized allocation results, which has higher matching accuracy and stronger robustness, providing a reliable data foundation for achieving high-precision 3D localization.

[0077] In one embodiment, in step S7, the final three-dimensional coordinates The calculation method is as follows:

[0078] First, calculate the depth of the target along each line of sight. Then, take points along each line of sight. Finally, the final three-dimensional spatial coordinates are obtained by weighted averaging. The weight satisfy .

[0079] In this embodiment, observation information distributed across different cameras and lines of sight is fused into a high-precision 3D coordinate system. Compared to simple triangulation or equal-weighted averaging in existing technologies, the weighted fusion method of this invention has the following advantages: First, depth optimization suppresses detection errors from a single line of sight; second, weighted averaging allows high-quality observations to play a greater role; finally, the designability of weights enables the system to adaptively adjust the fusion strategy according to the actual scene, maintaining stable positioning accuracy even when the target is partially occluded or the detection confidence level changes. As the final output of this invention, the results of all the preceding steps (calibration, detection, geometric transformation, candidate generation, graph optimization, and node collection) are transformed into directly applicable 3D positioning results, forming a complete technical closed loop.

[0080] The above description is merely an embodiment and does not constitute any limitation on the present invention. Any person skilled in the art can make many possible variations, modifications, or alterations to the technical solutions of the present invention without departing from the scope of the present invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the present invention, should fall within the protection scope of the present invention.

Claims

1. A multi-camera cooperative localization method based on a graph model, characterized in that, Includes the following steps: Step S1: Construct a stereo vision observation network consisting of at least two cameras to cover the area to be observed, and calibrate the intrinsic and extrinsic parameters of each camera; Step S2: Each camera performs single-camera detection, obtains the two-dimensional coordinates of the target in the image through the target detection network, and extracts the appearance features of the target through the re-identification network; Step S3: Based on the camera intrinsic and extrinsic parameters obtained from the calibration, convert the two-dimensional coordinates into geometric features representing the spatial orientation of the target. Step S4: Based on geometric features and camera extrinsic parameters, perform pairwise intersection calculations on the lines of sight of different cameras to generate multiple candidate spatial locations; Step S5: Using the targets detected by each camera as nodes, construct a graph model, in which a connecting edge is established between any two nodes from different cameras, and nodes from the same camera are not connected to each other. Step S6: Establish a joint cost function that includes both node cost and edge cost. The node cost is used to measure the geometric consistency between the geometric features of a node and the assigned candidate spatial position. The edge cost is used to measure the similarity between two connected nodes in terms of appearance and geometric features. By minimizing the joint cost function, the corresponding candidate spatial position is assigned to each node to form the optimal state estimate. Step S7: Identify all nodes assigned to the same candidate spatial location as the same target, collect multi-camera observation data corresponding to these nodes, and calculate the final three-dimensional coordinates of the target under test through weighted fusion. For any two different cameras, perform pairwise intersections of rays and calculate the midpoint of the common perpendicular segment of the two rays to obtain a series of candidate points; Cluster all candidate points and use each cluster center as a candidate spatial location. ; In step S6, the joint cost function is specifically as follows: in, , , For nodes Assigned candidate spatial location coordinates, and It is a node Connecting nodes Includes appearance and geometric features, , , , For node weights, The edge weight; In step S6, minimizing the cost function yields the optimal state estimate Z as follows: in, , Indicates camera The Each detected node is assigned to a candidate space location. .

2. The multi-camera cooperative localization method based on a graph model as described in claim 1, characterized in that, In step S3, the two-dimensional coordinates are converted into geometric features representing the spatial orientation of the target as follows: For any camera The geometric features of any detected target Calculated using the following formula: in, For camera The intrinsic parameter matrix, For camera The rotation matrix represents the absolute rotation of the camera coordinate system relative to the world coordinate system. Let be the homogeneous coordinates of the target object in the image; Geometric features This represents the direction vector of the spatial ray pointing from the camera's optical center toward the target.

3. The multi-camera cooperative localization method based on a graph model as described in claim 2, characterized in that, In step S4, the specific methods for generating multiple candidate spatial locations include: Based on the calibration parameters of each camera, the geometric features corresponding to each target are... The mapping is a spatial ray originating from the camera's optical center, and the world coordinates of the camera's optical center are represented as follows: in, For camera The coordinates of the optical center in the world coordinate system. This represents the absolute translation of the camera coordinate system relative to the world coordinate system.

4. The multi-camera cooperative localization method based on a graph model as described in claim 3, characterized in that, In step S5, the specific methods for constructing a graph model include: Each target detected by each camera is treated as a node. Each node contains corresponding appearance features. and geometric features ; Set up a camera Detected If there are 1 objective, then the total number of nodes in the graph model is: in, N The total number of cameras; For any two nodes from different cameras, an edge is created between them. Nodes within the same camera are not connected to each other. The total number of edges in the graph model is... for:

5. The multi-camera cooperative localization method based on a graph model as described in claim 1, characterized in that, In step S7, the method for collecting multi-camera observation data corresponding to these nodes is as follows: For the first node... One goal: to collect all The nodes are used to obtain a set of camera position coordinates corresponding to the nodes. and geometric features .

6. The multi-camera cooperative localization method based on a graph model as described in claim 5, characterized in that, In step S7, the final three-dimensional coordinates The calculation method is as follows: First, calculate the depth of the target along each line of sight. ; Then, take points along each line of sight. Finally, the final three-dimensional spatial coordinates are obtained by weighted averaging. The weight satisfy .

Images