A Dynamic Calibration Method for Extrinsic Parameters of Distributed Sensor Networks Based on Spatial Intelligence

Abstract

This invention relates to the fields of artificial intelligence and robot perception technology, specifically a method and system for dynamic calibration of extrinsic parameters in a distributed sensor network based on spatial intelligence. The method extracts atomic spatiotemporal events from each sensor and generates event features to construct a randomly initialized correlation field matrix. The correlation field state corresponding to the current time step, the event feature set, and the time step encoding are input into a pre-trained correlation field denoising network, iteratively obtaining the final correlation field state. High-confidence homologous event counterparts are extracted, three-dimensional point pairs are constructed, and the extrinsic parameter matrix is ​​solved using a random sampling consensus algorithm. The method is repeatedly executed over a sliding time window and combined with temporal filtering to achieve dynamic updates. This method can improve the robustness and accuracy of extrinsic parameter calibration in complex dynamic environments and under non-overlapping field-of-view conditions.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and robot perception technology, specifically to a method and system for dynamic calibration of extrinsic parameters of a distributed sensor network based on spatial intelligence. Background Technology

[0002] In applications such as smart city traffic management, automated warehousing and logistics, vehicle-road cooperation, and multi-robot collaborative operations, it is often necessary to unify the observation data collected by multiple heterogeneous and distributed sensors into a single global coordinate system for fusion processing. The foundation for achieving this goal lies in accurately obtaining the extrinsic parameter matrices between each sensor, which characterize the relative rotational and translational relationships between the sensors.

[0003] Existing external parameter calibration methods can generally be classified into the following categories.

[0004] The first category is calibration methods based on calibration objects. These methods establish correspondences between different sensors by simultaneously introducing calibration objects with known geometric structures, such as checkerboard patterns, ArUco markers, or dedicated 3D calibration targets, into the fields of view of multiple sensors, and then solving for the extrinsic parameter matrix. This type of method has the advantages of high accuracy and mature algorithms, but it usually requires manual placement and adjustment of calibration objects, making the process cumbersome. For already deployed sensor networks, recalibration often requires interrupting system operation. Furthermore, this type of method is generally difficult to meet the requirements of online calibration or dynamic continuous calibration.

[0005] The second category is calibration methods based on static scene features. These methods utilize stable static features present in the scene, such as building corners, edges, planar structures, or geometric structures in point clouds. The extrinsic parameter matrix is ​​solved by matching these static features across observations from different sensors. While these methods do not require additional calibration objects, their performance is highly dependent on the geometric richness of the scene itself. When there are few static structures, weak textures, or a significantly larger number of dynamic targets than static features in the scene, it is often difficult to obtain a sufficient number of stable matching points with a reasonable spatial distribution, leading to decreased calibration accuracy or even matching failure.

[0006] The third category is self-calibration methods based on dynamic events. These methods utilize dynamic targets in the scene as calibration cues. When the same dynamic target appears in the fields of view of different sensors at similar times, the spatiotemporal events it generates should satisfy certain consistency constraints. Based on this, a correspondence can be established and the extrinsic parameter matrix solved by matching co-originating events across sensors. This type of method does not require the introduction of dedicated calibration objects, has low dependence on static scenes, and has the potential for online self-calibration applications.

[0007] However, most existing matching methods based on dynamic events employ a local discriminative approach, where each pair of candidate events is judged separately for matching, for example, using graph neural networks (GNNs), Siamese networks, or other classification models for local discrimination. This type of method is prone to significant matching ambiguity when facing complex scenarios with dense events, similar appearances, and similar motion patterns. For example, in traffic scenarios, multiple vehicles with similar appearances and speeds may pass through an intersection simultaneously; in warehousing scenarios, multiple AGVs may perform similar transport tasks. In such cases, an event from one sensor may exhibit high similarity to multiple events from another sensor in terms of time, space, and appearance, leading to erroneous one-to-many or many-to-many associations in the local discriminative method. These erroneous associations further contaminate the subsequent pose estimation process, making it difficult for robust solution algorithms such as RANSAC to converge to the correct result, and in severe cases, even causing calibration failure.

[0008] Therefore, there is a need for a distributed sensor network extrinsic dynamic calibration method and system that can handle cross-sensor event correlations at the global level, suppress one-to-many and many-to-many erroneous matching, and continuously output stable extrinsic parameter estimation results over a sliding time window. Summary of the Invention

[0009] The purpose of this invention is to provide a dynamic calibration method for extrinsic parameters of distributed sensor networks based on spatial intelligence, in order to solve the problems of insufficient robustness of existing self-calibration methods based on dynamic events when facing matching ambiguity, easy generation of one-to-many or many-to-many incorrect matches, and thus affecting the accuracy of extrinsic parameter solution.

[0010] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0011] This invention provides a method for dynamic calibration of extrinsic parameters in a distributed sensor network based on spatial intelligence, the method comprising the following steps:

[0012] Each sensor in the distributed sensor network processes the sensing data it collects to extract atomic spatiotemporal events. Each atomic spatiotemporal event includes at least a timestamp, a three-dimensional position in the local coordinate system of the corresponding sensor, and a semantic category, and may further include at least one of instantaneous velocity, acceleration, and appearance features.

[0013] For the first event set corresponding to the first sensor to be calibrated and the second event set corresponding to the second sensor, an association field matrix is ​​constructed, and the association field matrix is ​​initialized to the initial association field state. The initial associated field state is a random noise state.

[0014] The associated field state corresponding to time step t The system uses the event feature sets corresponding to the first and second event sets, along with the encoded input of the current time step into a pre-trained association field denoising network, to predict the association field state at time step t. The noise in the data is analyzed, and the associated field state at time step t is predicted based on the noise. Update to obtain the associated field state of the previous time step. From the initial associated field state Begin, following the preset time step sequence. Iteratively update the associated field; when the update at t=1 is completed, the final associated field state is obtained. .

[0015] For the final associated field state Normalization is performed in both the row and column dimensions. Event pairs that simultaneously satisfy the maximum correlation probability in the row direction and the maximum correlation probability in the column direction, and whose correlation probability is higher than a preset threshold, are extracted as high-confidence homologous event counterparts.

[0016] A three-dimensional point pair is constructed based on the three-dimensional positions of each event in the high-confidence homogeneous event correspondence, and the extrinsic parameter matrix between the first sensor and the second sensor is solved using a random sampling consensus algorithm.

[0017] Repeat the above steps over a sliding time window and perform time-series filtering on the continuously solved extrinsic parameter matrix sequence to obtain dynamically updated extrinsic parameter estimation results.

[0018] Preferably, the correlation field matrix is:

[0019]

[0020] Where M is the number of events in the first event set, N is the number of events in the second event set, and the matrix elements are... Used to represent the first event in the first event set The first event and the second event set Each event is a correlation value of a common source event; the random noise state is initialized by sampling the matrix elements of the correlation field matrix from a standard normal distribution.

[0021] Preferably, the associated field denoising network is a neural network based on the Transformer architecture, and the network input includes the associated field state at time step t. The network outputs a prediction result for the associated field noise, consisting of an event feature set and a time step encoding. The associated field denoising network includes a self-attention module and a cross-attention module. The self-attention module is used to model the global context relationship between candidate matching pairs, and the cross-attention module is used to inject feature information from the event feature set into the denoising prediction process of the associated field.

[0022] Preferably, the correlation field denoising network is trained through a forward denoising process before deployment; the forward denoising process uses the truth correlation matrix. Starting with a predetermined variance, a series of noisy training samples are generated by progressively injecting Gaussian noise with a predetermined variance. The training objective is to minimize the difference between the network's predicted noise and the actual injected noise.

[0023] Preferably, the time-series filtering is a Kalman filter, used to smooth the extrinsic parameter matrix obtained in a continuous time window and update its state.

[0024] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0025] First, this invention does not isolate the discrimination of individual candidate event pairs, but processes the relationships of all candidate events as a whole on a unified correlation field matrix, and forms a clear correlation structure through iterative denoising, thereby effectively suppressing one-to-many and many-to-many erroneous matching.

[0026] Second, this invention uses the associated field state corresponding to time step t. Together with the event feature set as conditional input, the denoising process can comprehensively utilize the global contextual relationships across events as well as the event's own temporal, spatial, semantic, motion, and appearance information, thereby improving the robustness of association in complex dynamic environments.

[0027] Third, the present invention repeatedly solves the extrinsic parameter matrix over a sliding time window and combines time-series filtering to smooth the sequence of extrinsic parameter matrices, thereby outputting stable, continuous and dynamically updated extrinsic parameter estimation results. Attached Figure Description

[0028] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0029] Figure 1This is a schematic diagram of atomic spatiotemporal event extraction and representation in an embodiment of the present invention. The diagram shows that atomic spatiotemporal events are generated from the original data stream of the sensor, after target detection, target tracking and event definition, and further formed into event feature vectors including timestamps, three-dimensional positions, semantic categories, motion features and appearance features.

[0030] Figure 2 This is a schematic diagram of the iterative denoising process of the correlation field matrix in an embodiment of the present invention; the diagram shows the process by which the correlation field gradually evolves from a random noise state to a clear structure state through back diffusion iteration.

[0031] Figure 3 This is a schematic diagram of the structure of the associated field denoising network in an embodiment of the present invention. The diagram shows that the associated field state, event feature set and time step encoding corresponding to time step t are input together into the network backbone formed by stacking multiple Transformer blocks, and the noise prediction result is obtained through the output projection layer.

[0032] Figure 4 This is a schematic diagram of the training process of the correlation field denoising network in an embodiment of the present invention; the diagram shows the training process based on the true correlation matrix, obtaining noisy training samples by forward noise addition, and then inputting them into the network for noise prediction and loss calculation.

[0033] Figure 5 This is a schematic diagram of the system architecture in an embodiment of the present invention; the figure shows the data flow relationship between multiple distributed sensors, edge computing units and central computing units, wherein the edge computing units are used to generate atomic spatiotemporal event streams, and the central computing units are used to perform correlation denoising, dynamic extrinsic parameter calculation and dynamic updating.

[0034] Figure 6 This is a flowchart of the overall method in an embodiment of the present invention; the flowchart shows the complete process from data acquisition, atomic spatiotemporal event extraction and characterization, correlation field matrix construction, correlation field denoising, extraction of corresponding entities of homologous events, solving of extrinsic parameter matrices to dynamic updating and temporal filtering. Detailed Implementation

[0035] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Example 1: System Overall Architecture and Methodology

[0037] like Figure 5As shown, the application scenario of this embodiment can be a smart warehousing scenario. The system includes multiple distributed sensors deployed in different locations, edge computing units that are communicatively connected to the sensors, and a central computing unit. The distributed sensors can be cameras, LiDAR, or other sensors capable of collecting environmental perception data. The edge computing units are used to preprocess the local perception data and generate atomic spatiotemporal events; the central computing unit is used to perform correlation field matrix construction, correlation field denoising, extraction of high-confidence homologous event counterparts, extrinsic parameter matrix solving, and dynamic update processing. Figure 6 The overall flow of the method is shown.

[0038] The method may include the following steps.

[0039] Step S1: Atomic Spatiotemporal Event Extraction and Event Feature Generation

[0040] First, each sensor processes the raw sensing data it acquires. For cameras, this involves processing video stream data; for LiDAR, it involves processing point cloud sequence data. Edge computing units can perform target detection and tracking on the sensor side to extract dynamic targets from the raw sensing data.

[0041] Specifically, target detection can employ YOLO, Faster R-CNN, or other target detection models to identify dynamic targets of interest, such as AGVs, workers, or vehicles, within image frames or point cloud frames. Target tracking can utilize DeepSORT, FairMOT, or other multi-target tracking algorithms to assign a unique target identifier to the detected target within the current sensor field of view and continuously track its trajectory changes.

[0042] When the state of the tracked target undergoes a preset change, that change is defined as an atomic spatiotemporal event. State changes can include, for example, the target entering or leaving the field of view for the first time, passing through a predefined region, a sudden change in velocity, or a sudden change in direction. Each atomic spatiotemporal event can be recorded as a structured data object.

[0043] To enable subsequent networks to process atomic spatiotemporal events, each event needs to be converted into a corresponding event feature vector. Specifically, the event feature vector can be obtained by concatenating one or more of the following features:

[0044] 1. Spatiotemporal characteristics. Spatiotemporal characteristics include the event timestamp. and the three-dimensional position of the event in the corresponding sensor's local coordinate system. Among them, timestamp It can be converted into a high-dimensional vector, three-dimensional position, through sinusoidal positional encoding. It can be used directly or after being encoded by a small multilayer perceptron (MLP).

[0045] 2. Semantic Features. Semantic features include the category label of the target corresponding to the event. Category labels can be mapped to category embedding vectors through a learnable embedding layer.

[0046] 3. Motion characteristics. Motion characteristics may include the instantaneous velocity of the target at the time the event occurs. and acceleration .

[0047] 4. Appearance Features. When the sensor is a camera, appearance feature vectors can be extracted from slices of the target image, for example, using a pre-trained Re-ID network.

[0048] Therefore, each event Each can correspond to an event feature vector. ,event eigenvectors It is the concatenation of all the feature embedding vectors mentioned above, that is The feature vectors of each event together constitute the event feature set F.

[0049] Step S2: Construction and initialization of the correlation field matrix

[0050] Assume that the first and second sensors to be calibrated generate the first event set respectively. Second event set For the first set of events and the second set of events, construct the correlation field matrix. , where matrix elements Used to represent the first event in the first event set The first event and the second event set Each event is a related value of a source event.

[0051] Initially, the system is uncertain about the true origin relationship, so the correlation field matrix is ​​initialized to a random noise state. Specifically, the initial correlation field state corresponding to time step T can be denoted as... and make Each matrix element independently follows a standard normal distribution. Obtained through sampling. This initialization method corresponds to a completely uncertain state of the association relationship. Figure 2 The right side reveals the state of random noise. Figure 6 The process node of "constructing the initial random correlation field" was also disclosed.

[0052] Step S3: Backdiffusion Iteration Based on Correlation Field Denoising Network

[0053] This step is the core step of the present invention, used to determine the associated field state corresponding to time step t. Gradually restore it to a clear relational structure.

[0054] From the initial associated field state Begin, following the preset time step sequence. Iteratively update the associated field. At each time step t, update the associated field state corresponding to time step t. The event feature set F and the encoded input of the current time step t are associated with the field denoising network. Output from the network and Noise prediction matrix of the same dimension:

[0055]

[0056] Subsequently, the associated field state at time step t is determined based on the predicted noise. Update to obtain the associated field state of the previous time step. Its update process can be represented as:

[0057]

[0058] After the update at t=1 is completed, the final associated field state is obtained. .in, Figure 2 The t=0 annotation on the left indicates the termination state number, corresponding to the final associated field. .

[0059] In one implementation, the update function can use the backsampling formula of DDPM:

[0060]

[0061]

[0062]

[0063] in, It is a predefined time step From 1 to Monotonically increasing noise scheduling sequences (e.g., from) linearly increase to ); It is the standard deviation that controls the randomness of sampling, usually taken as... Or a related quantity; It is a newly sampled standard Gaussian noise matrix; if ,but .

[0064] To improve inference efficiency, the deterministic sampling method of DDIM can also be used. The formula simplifies to Direct prediction and interpolation.

[0065] After a predetermined number of iterations, the correlation field matrix changes from the initial correlation field state. Gradually evolves into the final associated field state At this point, the matrix elements representing true homology relationships will be significantly higher than the matrix elements representing mismatch relationships, thus forming an association result that approximates a unimodal matching structure. Figure 2 The evolution process from a random noise state to a clear structure state is disclosed.

[0066] Step S4: Extraction of high-confidence homologous event correspondences

[0067] After obtaining the final associated field state Subsequently, to extract high-confidence homologous event counterparts for final pose estimation, the final correlation field state can be analyzed. Normalization is performed on both the row and column dimensions, for example, by performing a Softmax operation, to obtain the normalized association probabilities.

[0068] Based on this, event pairs that simultaneously meet the following conditions are extracted as high-confidence homologous event counterparts:

[0069] 1. It has the highest correlation probability in the corresponding row;

[0070] 2. It has the highest correlation probability in the corresponding column;

[0071] 3. Its correlation probability is higher than the preset threshold.

[0072] By adopting the above-mentioned bidirectional constraint plus threshold filtering method, one-to-many or many-to-many erroneous matching can be further suppressed, thereby improving the reliability of subsequent extrinsic parameter solution.

[0073] Step S5: Solving the extrinsic parameter matrix based on the correspondences of high-confidence homologous events

[0074] For the high-confidence homogeneous event counterparts obtained in step S4, their 3D positions in the first sensor local coordinate system and the second sensor local coordinate system are extracted respectively, constructing a set of 3D point pairs. Let the sets of 3D point pairs be respectively... and The goal of solving the extrinsic matrix is ​​to obtain the rotation matrix R and the translation vector t such that:

[0075]

[0076] Since a small number of false matches may still exist in the correspondences of high-confidence homologous events, RANSAC is used for robust solving of the extrinsic parameter matrix. The specific process may include:

[0077] 1. Within a preset number of iterations, randomly select the smallest set of point pairs that can satisfy the solution conditions for the three-dimensional rigid body transformation from the set of three-dimensional point pairs; for the solution of the three-dimensional rigid body transformation, 3 pairs of non-collinear points can be used;

[0078] 2. Based on the minimum set of point pairs, use SVD decomposition or Kabsch algorithm to obtain the candidate extrinsic parameter matrix;

[0079] 3. Verify all 3D point pairs using the candidate extrinsic matrix and calculate the transformation residuals for each point pair; when the residuals are less than a preset threshold, the corresponding point pair is recorded as an interior point;

[0080] 4. Record the candidate extrinsic matrix that receives the most interior point support;

[0081] 5. After all iterations are completed, the final extrinsic parameter matrix is ​​obtained by resolving the problem based on the set of interior points corresponding to the best candidate extrinsic parameter matrix.

[0082] Step S6: Dynamic Updates and Timing Filtering on the Sliding Time Window

[0083] To achieve dynamic continuous calibration, steps S1 to S5 can be performed periodically over a sliding time window. For example, atomic spatiotemporal events extracted within the most recent time window can be processed at preset intervals to obtain a time-varying sequence of extrinsic parameter estimates. .

[0084] Since the quality of events may fluctuate within different time windows, the directly obtained extrinsic parameter estimation sequence may exhibit jitter. Therefore, a time-series filter can be further introduced to smooth the extrinsic parameter estimation sequence. In one implementation, a Kalman filter can be used.

[0085] Specifically, the state vector of the Kalman Filter can be defined as a Lie algebraic representation of the extrinsic parameters, including rotation and translation vectors. In the prediction phase, an approximate static model can be used, and process noise is introduced during prediction to allow for slow drift of the extrinsic parameters. In the update phase, each newly obtained extrinsic parameter matrix is ​​used as an observation to update the filter state. The observation noise covariance can be set based on the number of inliers and the fitting residuals obtained during the RANSAC process; the more inliers and the smaller the fitting residuals, the smaller the observation noise covariance can be set. Through the above processing, the system can output smooth, stable, and continuously updated extrinsic parameter estimation results.

[0086] Example 2: Correlation Field Denoising Network Structure and Training Process

[0087] like Figure 3 As shown, the correlation field denoising network It can be built using the Transformer architecture. The network is used to represent the associated field state at a given time step t. Predict the associated field noise at the current time step, given the conditions of (noisy state), event feature set, and time step encoding. Figure 4 The online training process has been made public.

[0088] In one implementation, the input to the correlated field denoising network includes the correlated field state at time step t. The event feature set F and the encoding of time step t. The associated field state corresponding to time step t. It can be viewed as a matrix input, or it can be flattened into a one-dimensional sequence before input; the event feature set F and the time step encoding are input into the network as conditional information.

[0089] The network backbone can be formed by stacking multiple Transformer blocks. Each Transformer block may include:

[0090] 1. Self-attention module. The self-attention module is used to establish global associations between candidate matching pairs, enabling a certain position in the network to pay attention to all other positions, thereby learning the overall constraint relationship between candidate matching pairs;

[0091] 2. Cross-Attention Module. The cross-attention module is used to inject information from the event feature set into the association field denoising process; in one implementation, the internal representation of the association field can be used as the Query, and the event feature set can be used as the Key and Value;

[0092] 3. Feedforward network. A feedforward network is used to perform a non-linear mapping on the output of the attention module;

[0093] 4. Output Projection Layer. The output projection layer is used to map the high-dimensional representation inside the network back to the same dimension as the correlation field matrix to obtain the noise prediction result.

[0094] The correlation field denoising network can be trained using a forward noise addition method. The specific process is as follows.

[0095] First, prepare training samples. Training samples consist of paired sets of events and their corresponding ground truth matching relationships. Training samples can come from a simulation environment, such as CARLA or Isaac Sim. In the simulation environment, global identifiers of the targets can be obtained, thereby determining the true homology relationships between cross-sensor events. True homology relationships can be represented as a ground truth correlation matrix. In one implementation, the truth correlation matrix It can be represented in binary form, that is, when the... The event and the first When the events are of the same origin, the matrix elements ;otherwise .

[0096] Secondly, the truth-valued incidence matrix Perform forward noise addition. For any time step t, the associated field state at time step t. It can be obtained from the following formula Direct generation:

[0097]

[0098] in, To and A standard Gaussian noise matrix of the same size.

[0099] During training, a training sample, a time step t, and a noise matrix can be randomly sampled. Based on the above formula, the associated field state corresponding to time step t is generated. As a noisy sample, and The event feature set F and the time step t are used as input correlation field denoising networks. The predicted noise is obtained. The loss function can be defined as the mean square error between the predicted noise and the actual noise:

[0100]

[0101] By minimizing the loss function, the network can progressively learn to recover the true association structure from the association field state corresponding to time step t under different noise levels.

[0102] Example 3: System Implementation

[0103] The present invention also provides a dynamic calibration system for extrinsic parameters of a distributed sensor network based on spatial intelligence. The system includes at least one processor and a memory coupled to the at least one processor. The memory stores computer program instructions, which, when executed by the at least one processor, cause the system to perform the method of any of the foregoing embodiments.

[0104] In one implementation, the system further includes multiple sensors, edge computing units that are communicatively connected to each sensor, and a central computing unit. The edge computing units are used to perform atomic spatiotemporal event extraction and event feature generation; the central computing unit is used to perform correlation field matrix construction, correlation field denoising, extraction of high-confidence homologous event counterparts, extrinsic parameter matrix solving, and dynamic updating of extrinsic parameters. Figure 5 A structural diagram illustrating the implementation of this system is provided.

[0105] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic calibration of extrinsic parameters in a distributed sensor network based on spatial intelligence, characterized in that, The method includes the following steps: S1. Each sensor in the distributed sensor network processes the sensing data it collects and extracts atomic spatiotemporal events. Each atomic spatiotemporal event includes at least a timestamp, a three-dimensional position in the local coordinate system of the corresponding sensor, and a semantic category, and further includes at least one of instantaneous velocity, acceleration, and appearance features. S2. For the first event set corresponding to the first sensor to be calibrated and the second event set corresponding to the second sensor, construct an association field matrix and initialize the association field matrix to the initial association field state. The initial associated field state is a random noise state; S3. Associativity of the field state corresponding to time step t The system uses the event feature sets corresponding to the first and second event sets, along with the encoded input of the current time step into a pre-trained association field denoising network, to predict the association field state at time step t. The noise in the data is analyzed, and the associated field state at time step t is predicted based on the noise. Update to obtain the associated field state of the previous time step. From the initial associated field state Begin, following the preset time step sequence. Iteratively update the associated field; when the update at t=1 is completed, the final associated field state is obtained. ; S4. Regarding the final associated field state Normalization is performed in both the row and column dimensions to extract event pairs that simultaneously satisfy the maximum correlation probability in the row direction and the maximum correlation probability in the column direction, and whose correlation probability is higher than a preset threshold. These are then used as high-confidence homologous event counterparts. S5. Construct three-dimensional point pairs based on the three-dimensional positions of each event in the high-confidence homogeneous event correspondence, and use the random sampling consensus algorithm to solve the extrinsic parameter matrix between the first sensor and the second sensor; S6. Repeat steps S1 to S5 over the sliding time window, and perform time-series filtering on the continuously solved extrinsic parameter matrix sequence to obtain dynamically updated extrinsic parameter estimation results.

2. The method for dynamic calibration of extrinsic parameters of a distributed sensor network based on spatial intelligence according to claim 1, characterized in that, The correlation field matrix mentioned in step S2 is: Where M is the number of events in the first event set, N is the number of events in the second event set, and the matrix elements are... Used to represent the first event in the first event set The first event and the second event set Each event is a correlation value of a common source event; the random noise state is initialized by sampling the matrix elements of the correlation field matrix from a standard normal distribution.

3. The method for dynamic calibration of extrinsic parameters of a distributed sensor network based on spatial intelligence according to claim 1, characterized in that, The update process in step S3 is a back diffusion process, following a preset time step sequence. The update process is performed iteratively and satisfies the following conditions: in, The associated field state corresponding to time step t, The associated field state corresponding to time step t-1, For correlation field denoising network, For the set of event features, For the current time step, This is an update function that calculates the state of the previous time step based on the current state and the predicted noise.

4. The method for dynamic calibration of extrinsic parameters of a distributed sensor network based on spatial intelligence according to claim 1 or 3, characterized in that, The associated field denoising network is a neural network based on the Transformer architecture, and its input includes the associated field state at time step t. event feature set and current time step The encoding outputs the prediction results of noise in the associated field.

5. The method for dynamic calibration of extrinsic parameters of a distributed sensor network based on spatial intelligence according to claim 4, characterized in that, The association field denoising network includes a self-attention module and a cross-attention module; wherein, the self-attention module is used to capture the global contextual relationship between candidate matching pairs, and the cross-attention module is used to inject feature information from the event feature set into the denoising prediction process of the association field.

6. The method for dynamic calibration of extrinsic parameters of a distributed sensor network based on spatial intelligence according to claim 1, characterized in that, The correlation field denoising network is trained through a forward denoising process before deployment; the forward denoising process uses the truth correlation matrix. Starting with a predetermined variance, a series of noisy training samples are generated by progressively injecting Gaussian noise with a predetermined variance. The training objective is to minimize the difference between the network's predicted noise and the actual injected noise.

7. The method for dynamic calibration of extrinsic parameters of a distributed sensor network based on spatial intelligence according to claim 6, characterized in that, The truth correlation matrix This is a matrix representing the correspondence between real events, where the element values ​​at positions corresponding to events of the same origin are higher than the element values ​​at positions corresponding to events of different origins.

8. The method for dynamic calibration of extrinsic parameters of a distributed sensor network based on spatial intelligence according to claim 1, characterized in that, Step S5 includes: From the high-confidence homogeneous event correspondences, a set of minimum point pairs for solving the three-dimensional rigid body transformation is randomly selected, and the candidate extrinsic parameter matrix is ​​calculated. The candidate extrinsic matrix is ​​used to verify all three-dimensional point pairs, and three-dimensional point pairs with errors less than a preset threshold are determined as interior points; The candidate extrinsic matrix that receives the most inlier support is selected as the optimal candidate extrinsic matrix, and the final extrinsic matrix is ​​recalculated based on the corresponding inlier.

9. The method for dynamic calibration of extrinsic parameters of a distributed sensor network based on spatial intelligence according to claim 1, characterized in that, The temporal filtering in step S6 is a Kalman Filter; the temporal filtering uses the extrinsic parameter matrix obtained at each time step as the observation for state update, and sets the observation noise covariance according to the number of interior points and / or the fitting residual in the RANSAC solution process.

10. A dynamic calibration system for extrinsic parameters of a distributed sensor network based on spatial intelligence, characterized in that, The system includes at least one processor and a memory coupled to the at least one processor, the memory storing computer program instructions, which, when executed by the at least one processor, cause the system to perform the spatial intelligence-based distributed sensor network extrinsic dynamic calibration method as described in any one of claims 1 to 9.

11. The spatial intelligence-based distributed sensor network extrinsic parameter dynamic calibration system according to claim 10, characterized in that, The system also includes: Multiple sensors are used to collect sensing data; Edge computing units, which are communicatively connected to each of the aforementioned sensors, are used to extract atomic spatiotemporal events and generate event features; The central computing unit is used to perform correlation field matrix construction, correlation field iterative denoising, high-confidence homologous event correspondence extraction, extrinsic parameter matrix solving, and dynamic updating of the extrinsic parameter matrix sequence.

Images