A cluster head dominant unmanned aerial vehicle target multi-camera cooperative tracking method and system

Abstract

The present application relates to a kind of cluster head dominated unmanned aerial vehicle target multi-camera cooperative tracking method and system.Method includes: each edge perception node obtains the video stream containing unmanned aerial vehicle target information and carries out target identification, and based on node perception parameter, target single point rough positioning is solved, and the initial spatial position state data of the detected target is output;According to the geographical distribution position and field of view coverage range of each edge perception node, it is divided into multiple detection clusters;And periodically dynamically elect optimal node as cluster head node in cluster;Cluster head node receives the initial spatial position state data of the detected target reported by cluster edge perception node, carries out homologous screening and multi-camera cross positioning solution, generates the three-dimensional coordinate track of the detected target;Each cluster head node uploads the three-dimensional coordinate track of the detected target to cloud center;Cloud center splices and fuses multiple three-dimensional coordinate tracks, generates the global track information of the detected target, and reports situation to client.

Description

Technical Field

[0001] This invention relates to the field of low-altitude security and visual surveillance technology, specifically to a cluster-head-led multi-camera collaborative tracking method and system for UAV targets. Background Technology

[0002] In recent years, the application of low-altitude civilian drones has rapidly become widespread, and the demand for refined management and control of drones in complex urban environments with high-rise buildings has become increasingly urgent. In order to meet the needs of low-altitude security, it is necessary to achieve low-cost, high-precision, and continuous drone target tracking under dense urban buildings and vegetation cover.

[0003] In existing technologies, panoramic camera visual detection solutions are commonly used to track low-altitude targets. This technology typically relies on a single fisheye lens or a multi-lens panoramic device to achieve 360° omnidirectional field of view coverage, enabling omnidirectional target capture without the need for gimbal rotation. Specifically, it acquires a global, blind-spot-free image using a single panoramic camera, independently performing drone target identification, coarse localization, and trajectory generation locally, and achieving continuous target tracking based on the temporal frame information of a single device. Some solutions utilize edge computing power to perform local trajectory optimization, reducing the pressure on cloud transmission.

[0004] However, the above-mentioned panoramic vision solutions have significant drawbacks in complex urban scenarios: First, they are not robust to occlusion, and a single panoramic device is easily blocked by buildings, and without multi-camera cross-positioning compensation, it is easy to lose track after occlusion; second, the hardware deployment cost is high, and it is impossible to reuse the massive amount of ordinary monitoring resources in the city; third, the existing solutions are all single-point independent operations, and have not introduced a clustered management mechanism in which a local central node (i.e., cluster head) coordinates the sub-nodes, and the computing power is easily overloaded when multiple targets are dense, making it impossible to achieve advance prediction of target flight across regions and multi-camera linkage. Summary of the Invention

[0005] The purpose of this invention is to provide a cluster-head-led multi-camera collaborative tracking method and system for UAV targets. This method can improve the robustness and continuity of low-altitude target tracking in complex high-rise building environments with low equipment investment. It solves the problem of UAV target tracking trajectory breakage caused by blind spot obstruction or high-speed target maneuvering, and optimizes the system resource scheduling efficiency. It can significantly improve the urban low-altitude security control capabilities.

[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows.

[0007] In a first aspect, the present invention provides a cluster-head-led multi-camera collaborative tracking method for UAV targets, the method comprising:

[0008] S1. Each edge sensing node acquires a video stream containing UAV target information and performs target recognition. Based on its own corresponding node sensing parameters, it performs coarse localization calculation of the target single point and outputs the initial spatial position state data of the detected target.

[0009] S2. Based on the geographical distribution and field of view coverage of each edge sensing node, divide them into multiple detection clusters; and periodically and dynamically elect the optimal node as the cluster head node within each cluster.

[0010] S3. The cluster head node receives the initial spatial position status data of the target being detected reported by the edge sensing nodes within the cluster, performs source filtering and multi-camera cross-positioning calculation, and generates the three-dimensional coordinate trajectory of the target being detected.

[0011] S4. Each cluster head node uploads the three-dimensional coordinate trajectory of the target to be detected to the cloud center; the cloud center stitches and merges multiple three-dimensional coordinate trajectories, performs relay detection and eliminates cluster positioning deviations, generates global trajectory information of the target to be detected, and reports the situation to the client.

[0012] As a preferred embodiment of the present invention, step S2, which involves periodically and dynamically electing the optimal node as the cluster head node within the cluster, specifically comprises:

[0013] Within the detection cluster, a five-dimensional weighted election model is constructed based on the real-time device status of each edge sensing node to calculate the comprehensive weight of each edge sensing node. ;

[0014] The comprehensive weight W of all edge sensing nodes within the detection cluster is calculated iteratively periodically, and the comprehensive weight is selected. The largest edge-aware node is used as the cluster head node.

[0015] As a preferred embodiment of the present invention, in step S3, the multi-camera cross-positioning calculation is performed based on the three-dimensional decoupled Kalman filter method, and the specific steps are as follows:

[0016] Based on the line-of-sight orientation and pitch angle of the edge sensing node, a rotation matrix is ​​constructed to transform the observations, states, and covariance in the ENU coordinate system to the local coordinate system of the noise principal axis.

[0017] After decoupling each axis in the local space and performing Kalman iteration to configure the observation noise differently for each axis, the state and covariance are restored to the global ENU coordinate system through the inverse transformation of the original rotation matrix.

[0018] As a preferred embodiment of the present invention, step S4 further includes a collaborative guided tracking sub-step:

[0019] Based on the global trajectory information, the movement speed and activity range of the detected target, the cloud center performs quantitative prediction, triggers preset collaborative command strategies in a differentiated manner, and issues scheduling commands to drive the corresponding edge perception nodes to continuously track the target.

[0020] As a preferred embodiment of the present invention, the preset instruction coordination strategy includes a local autonomous coordination strategy, a neighborhood-oriented coordination strategy, and a cross-domain cloud coordination strategy.

[0021] As a preferred embodiment of the present invention, the local autonomous collaboration strategy is generated based on a dual-mode collaboration mechanism of edge autonomy priority and cluster head optimization post-processing, and is superimposed with anti-shaking constraints.

[0022] As a preferred embodiment of the present invention, the neighborhood-oriented cooperative strategy is triggered when the detected target is about to fly out of the control boundary of the current detection cluster; the cross-domain cloud cooperative strategy is triggered when the detected target crosses multiple detection clusters within the speed time threshold.

[0023] Secondly, the present invention also provides a cluster-head-led multi-camera collaborative tracking system for UAV targets, the system comprising:

[0024] The edge perception module includes multiple edge perception nodes. The edge perception nodes are used to acquire video streams containing UAV target information and perform target recognition, and perform coarse localization calculation of the target single point based on their corresponding node perception parameters, and output the initial spatial position state data of the detected target.

[0025] The cluster head control module includes cluster head nodes that are periodically and dynamically elected within each detection cluster. These nodes are used to receive the initial spatial position status data and perform source filtering and multi-camera cross-positioning calculations to generate the three-dimensional coordinate trajectory of the detected target.

[0026] The cloud center is used to receive the three-dimensional coordinate trajectories uploaded by each cluster head node, and to stitch and fuse multiple three-dimensional coordinate trajectories to perform relay detection and eliminate cluster positioning deviations in order to generate global trajectory information of the detected target, and report the situation to the client.

[0027] As a preferred embodiment of the present invention, the cloud center includes a pre-division unit for detection clusters; the pre-division unit for detection clusters is used to divide and generate multiple detection clusters according to the geographical distribution location and field of view coverage of each edge sensing node.

[0028] As a preferred embodiment of the present invention, the cloud center includes a collaborative guidance and tracking unit; the collaborative guidance and tracking unit is used to make quantitative predictions based on the global trajectory information, the movement rate and activity range of the detected target, trigger preset collaborative command strategies in a differentiated manner, and issue scheduling commands to drive the corresponding edge perception nodes to continuously track the target.

[0029] In summary, the present invention has the following beneficial effects:

[0030] This invention constructs a three-level hierarchical distributed architecture of edge layer - cluster head layer - cloud center layer. At the edge layer, by reusing the city's existing visible light monitoring resources, continuous tracking of UAV targets can be achieved without the need for additional professional monitoring equipment, which significantly reduces hardware deployment costs. In addition, running some light tasks on edge sensing nodes balances the computing load and data transmission pressure of the entire network, which is conducive to the large-scale deployment of low-altitude UAV detection and tracking systems.

[0031] In this invention, by dynamically electing the optimal cluster head node within the cluster and performing multi-camera cross-positioning calculations, and simultaneously using a three-dimensional decoupled Kalman filter method to transform the global ENU coordinate system to the local coordinate system of the noise principal axis for split-axis filtering, noise disturbances caused by monocular visual ranging are accurately identified and eliminated. This significantly enhances the anti-interference capability of target three-dimensional trajectory fitting and position state calculation in complex occlusion environments, and has excellent spatial precision positioning accuracy.

[0032] In this invention, relying on the full-domain flight track information and the quantitative prediction of target movement speed, a differentiated three-level collaborative command strategy of local, neighboring, and cross-domain is triggered to realize the adaptive switching of the tracking network in complex maneuvering conditions such as target blind spot waiting, smooth cross-boundary relay, and high-speed large-span jump-level deployment, which significantly improves the ability and scheduling efficiency of urban low-altitude security control. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a flowchart of the method of the present invention;

[0035] Figure 2 This is a system structure block diagram of the present invention;

[0036] Figure 3 This is a schematic diagram illustrating the communication between the cloud center, cluster head node, and edge sensing node in a two-detector cluster scenario.

[0037] Figure 4 This is a schematic diagram illustrating the three-step execution of the present invention;

[0038] Figure 5 This is a schematic diagram illustrating the blind spots in the predicted location in the embodiment;

[0039] Figure 6This is a schematic diagram of target tracking within the cluster for blind spot filling in this embodiment;

[0040] Figure 7 This is a schematic diagram of target tracking corresponding to the neighborhood-oriented cooperative strategy in the embodiment;

[0041] Figure 8 This is a schematic diagram of target tracking corresponding to the cross-domain cloud collaboration strategy in the embodiment. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be explained and described below with reference to the accompanying drawings. However, the following embodiments are only preferred embodiments of the present invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of the present invention.

[0043] It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and are not intended to limit the scope, applicability, or examples set forth in the claims. The function and arrangement of the elements discussed may be changed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. For example, the described methods may be performed in a different order than described, and steps may be added, omitted, or combined. Furthermore, features described in some examples may be combined in other examples.

[0044] This invention provides a cluster-head-led multi-camera collaborative tracking system for UAV targets. At the hardware deployment level, it employs visual perception, eliminating the need for high-cost dedicated sensing and detection equipment such as panoramic cameras, radar, infrared, and photoelectric direction finding. The underlying perception and tracking layer is directly deployed on a vast number of ordinary surveillance cameras (bullet or PTZ) within urban areas. These cameras are upgraded to edge perception nodes by adding lightweight edge computing units (such as edge AI chips or embedded computing boards) or by upgrading the front-end firmware. These edge perception nodes are connected to the mid-layer cluster head nodes via network switches or wireless gateways according to spatial and geographical logic, ultimately converging at the top-layer cloud center.

[0045] Example 1

[0046] refer to Figure 1 , Figure 3 and Figure 4 As shown in the figure, this embodiment further elaborates on the cluster-head-led multi-camera collaborative tracking method for UAV targets provided by the present invention. The method includes:

[0047] S1. Each edge sensing node acquires a video stream containing UAV target information and performs target recognition. Based on its own corresponding node sensing parameters, it performs coarse localization calculation of the target single point and outputs the initial spatial position state data of the detected target.

[0048] During system operation, step S1 is executed independently by each edge sensing node. Each sensing node pulls the visible light video stream from the local ordinary surveillance camera in real time and decodes it into an image frame sequence. Using a deep learning target detection algorithm (such as a lightweight YOLO network) deployed in a lightweight edge computing unit, the low-altitude drone target in the image frame is extracted and identified in real time, and a bounding box containing the target category and pixel coordinates is output.

[0049] Once the edge sensing node identifies the drone target, it immediately combines the camera's inherent intrinsic parameters (such as focal length, principal point coordinates, and distortion coefficients) with the node sensing parameters calibrated during installation (including the camera's three-dimensional latitude, longitude, and altitude coordinates, real-time gimbal heading / azimuth angle, and pitch angle) to perform coarse localization calculation of the target single point using a line-of-sight vector projection algorithm. This transforms the pixels in the image coordinate system into initial position coordinates in three-dimensional geographic space, outputting the initial spatial position status data of the detected target.

[0050] After calculating the initial spatial position state data, the edge sensing node only reports the quantized three-dimensional coordinate data and the current state parameters of the camera (such as the current PTZ angle and computing load) to the cluster head node. It does not transmit any original video image streams between nodes within the detection cluster. Therefore, it can significantly reduce the amount of data transmitted over the network, free up network bandwidth, and thus provide physical feasibility for ultra-large-scale networking of the entire city.

[0051] S2. Based on the geographical distribution and field of view coverage of each edge sensing node, divide them into multiple detection clusters; and periodically and dynamically elect the optimal node as the cluster head node within each cluster.

[0052] The action of dividing the detection clusters can be achieved based on the detection cluster pre-division unit inside the cloud center. The detection cluster pre-division unit adaptively divides the massive number of edge sensing nodes into multiple independent detection clusters according to the geographical distribution location of the global edge sensing nodes, the camera axis, and the geometric coverage of the field of view, following the quantitative principles of geographical proximity, overlapping fields of view, and complementary spatial coverage.

[0053] Within a probe cluster, the system introduces a five-dimensional weighted election model based on the real-time device status of each edge sensing node. The optimal node is periodically (e.g., every 5 seconds) dynamically elected as the cluster head node. The specific election process is as follows:

[0054] The comprehensive weight W of all edge sensing nodes within the detection cluster is calculated periodically and iteratively:

[0055]

[0056] In the formula: , , , , For the normalized weight coefficients, satisfying:

[0057]

[0058] The effective field of view coverage after deducting environmental obstructions;

[0059] Normalized scoring of historical collaborative positioning accuracy for nodes;

[0060] Confidence coefficient for target tracking;

[0061] This is a network stability coefficient calculated based on transmission delay and packet loss rate;

[0062] This is the real-time computing load coefficient. The higher the load, the larger the value. It is used as a negative penalty term in the weighting formula.

[0063] The edge sensing node with the largest comprehensive weight W in the current period is automatically selected as the cluster head node of the detection cluster. Since the hardware is completely identical, the edge sensing node selected as the cluster head node retains its basic identification, localization and autonomous tracking capabilities while activating its exclusive local dominant control function.

[0064] The computing load coefficient L serves as a negative penalty term. Combined with the dynamic replacement of cluster head nodes using multi-dimensional indicators, it can eliminate the single-point failure risk of traditional fixed central node architecture. In complex urban environments, once the current cluster head node encounters sudden occlusion, local network fluctuations, or computing overload due to dense targets, its comprehensive weight W will decrease in the next cycle iteration. This allows the system to seamlessly and smoothly transfer local command to the idle sensing node with the best indicators in the cluster within one to two cycles.

[0065] S3. The cluster head node receives the initial spatial position status data of the target being detected reported by the edge sensing nodes within the cluster, performs source filtering and multi-camera cross-localization calculation, and generates the three-dimensional coordinate trajectory of the target being detected.

[0066] The newly elected cluster head node serves as the regional scheduling core, receiving initial spatial position status data reported in real time from all active edge sensing nodes within its detection cluster. Since different cameras may simultaneously observe multiple targets or be affected by environmental interference such as birds and clutter, the cluster head node first runs data association algorithms, such as the Mahalanobis distance threshold determination algorithm, to perform spatiotemporal homogeneity filtering and clustering of multi-source data, grouping coarse positioning data pointing to the same UAV target into the same data pool. Subsequently, the cluster head node performs multi-camera stereo cross-positioning calculations on the homogeneous data using a 3D decoupled Kalman filter method to generate a high-precision 3D coordinate trajectory of the detected target.

[0067] Because ordinary civilian monocular cameras exhibit strong anisotropy when detecting at long distances, their ranging noise diverges laterally along the camera's line of sight. If filtering is performed directly in the global northeast-northeast ENU coordinate system, the state covariance matrix will not converge. Therefore, based on the above-mentioned three-dimensional decoupled Kalman filtering method, multi-camera stereo cross-positioning calculation is performed on the same source data to generate a high-precision three-dimensional coordinate trajectory of the detected target.

[0068] Specifically, the cluster head node utilizes the real-time line-of-sight azimuth angle of the camera in the reported sensing data. and pitch angle Construct a composite rotation matrix of lines of sight :

[0069]

[0070] in, , The operators are Z-axis azimuth and X-axis pitch and rotation, respectively.

[0071] Using a line-of-sight composite rotation matrix Forward transformation inverse matrix The observation vector in the global ENU coordinate system State vector With covariance matrix Project and transform the data to the local coordinate system of the noise principal axis, and convert it into the corresponding transformation parameters. , and :

[0072]

[0073]

[0074]

[0075] In the transformed local space, the observation noise covariance matrix along each axis This achieves decoupling, ensuring that the observation errors of each axis are independent. At this point, the filter configures the observation noise covariance matrix differently for each axis in the local noise coordinate system and performs conventional Kalman time-series iterations.

[0076]

[0077] in, Here is the state transition matrix. For process noise array, The observation noise covariance matrix for each axis , For this step's Kalman gain, For prior estimation, This is a posterior estimate.

[0078] After iteration, the original line-of-sight composite rotation matrix is ​​used. The inverse transformation formula restores the state and covariance of the local space to the global ENU coordinate system:

[0079]

[0080]

[0081] Finally, a high-precision and unbroken three-dimensional coordinate trajectory of the target was fitted.

[0082] The above-described rotational decoupling filtering process eliminates the inter-axis coupling between the radial and azimuth / elevation angle observation errors in ranging; it can rely on the local principal axis coordinate system to differentiate the observation noise of each axis, thereby suppressing radial ranging noise while weakening the angular component filtering.

[0083] S4. Each cluster head node uploads the three-dimensional coordinate trajectory of the target to be detected to the cloud center; the cloud center stitches and merges multiple three-dimensional coordinate trajectories, performs relay detection and eliminates cluster positioning deviations, generates the global trajectory information of the target to be detected, and reports the situation to the client.

[0084] In step S4, the cluster head node of each detection cluster uploads the high-precision 3D coordinate trajectory calculated for its local area to the cloud center in real time. The cloud center aggregates the data reported by all detection clusters across the entire area, and stitches and fuses multiple local 3D coordinate trajectories at overlapping boundaries or temporal connection points to achieve a relay detection capability similar to existing panoramic eagle-eye cameras. This eliminates systematic cluster positioning deviations caused by physical errors in camera installation or local terrain filtering, ultimately generating a large-scale, all-area flight path information covering the low-altitude airspace of the entire city, which is then pushed to the situational awareness client.

[0085] In addition to generating global trajectory information, to address the issues of blind spots caused by tall buildings in urban areas and high-speed targets falling out of sight, the system utilizes cloud center or cluster head nodes to dynamically trigger a collaborative guided tracking sub-step. This sub-step performs quantitative prediction based on global trajectory information, the real-time movement rate and activity range of the detected target, and triggers a pre-set command collaboration strategy in a differentiated manner, issuing drive commands to orient the PTZ orientation of the PTZ camera.

[0086] The preset command coordination strategies include local autonomous coordination strategy, neighborhood-oriented coordination strategy, and cross-domain cloud coordination strategy.

[0087] Local Autonomous Collaboration Strategy: This strategy is applicable to guiding and tracking UAV targets at low speeds and within a single cluster.

[0088] When a target moves within a single detection cluster, the cluster head node allocates cameras based on a multi-target globally optimal allocation strategy. Its global optimization objective is to minimize the global comprehensive matching cost f.

[0089]

[0090] in, The sum of distances for all target-camera pairings. For global distance cost, The cost of observing the included angle of the configuration, The cost of switching solutions due to jitter.

[0091] It should be noted that the above optimal solution solution process is subject to four constraints: full coverage, distance, positioning configuration, and anti-shake. The specific constraints are as follows:

[0092] • Full Coverage Constraint: Prioritize ensuring that all detected targets are assigned to at least one camera for tracking. This serves as a prerequisite hard constraint for optimal solution, preventing target loss and ensuring the integrity of global tracking.

[0093] Distance constraint: Under the premise of full coverage, prioritize matching cameras that are spatially closer to each other to minimize the global distance cost. :

[0094]

[0095] In the formula: Let c be the three-dimensional coordinates of the k-th camera (the subscript c represents the camera, and k is the camera number). Let t be the 3D coordinates of the t-th target (subscript o represents the target object, t is the target number); the optimization process continuously reduces Overall numerical values ​​are used to achieve the optimal global average observation distance, thereby improving overall recognition accuracy.

[0096] • Positioning constraints: Distribute camera resources as widely as possible to avoid multiple cameras clustered together tracking a single target. Prioritize assigning two cameras to track the target, and ensure the angle between the two cameras crosses to satisfy the optimal geometric constraints.

[0097]

[0098] In the formula: The angle between the observation lines of the two cameras relative to the target is denoted as θ. This constraint ensures that each target can obtain stereo cross-observation conditions, which greatly improves the accuracy of multi-camera collaborative positioning calculation.

[0099] • Anti-jitter constraint: To suppress tracking jitter caused by frequent switching, the system introduces a scheme difference cost. Lazy constraints are applied to global solution updates;

[0100] Define the difference between the optimal allocation schemes in the two rounds as: The global pairing strategy is only updated when the following condition is met:

[0101]

[0102] In the formula: To determine the overall extent of the changes to the plan, This is a preset stability threshold.

[0103] This constraint is used to avoid frequent jumps in the optimal solution that cause the PTZ gimbal to rotate repeatedly and the target to briefly go out of sight. While ensuring global optimality, it can greatly improve the stability of long-term continuous tracking of multiple targets.

[0104] In this strategy, to address the inherent computational latency of cluster head multi-machine fusion and collaborative localization, which cannot adapt to the requirements of high-speed real-time tracking, this embodiment preferably designs a dual-mode architecture with edge autonomy priority and cluster head node optimization as a post-processing step:

[0105] • Edge Autonomous Priority Mode: After the edge node identifies the target, it immediately generates high-precision gimbal commands based on its own azimuth and pitch data, starts local self-tracking, ensures millisecond-level response, and avoids the target from falling out of sight while waiting for cluster head commands.

[0106] • Cluster Head Optimization Post-Processing Mode: The cluster head node continuously and asynchronously completes data fusion, global allocation, and blind zone prediction in the background to generate a global tracking scheme. Even in scenarios where no cross-location conditions are formed and only single-device data is relied upon, the cluster head node can issue coarse driving commands to drive devices that have not observed targets to turn to the predicted target position, avoiding device stagnation and improving the probability of target re-acquisition; when the cross-location conditions of multiple devices are met, the driving commands are then corrected with high-precision global positioning results to achieve a smooth transition.

[0107] • Dual-mode collaborative logic: Edge nodes always prioritize local high-precision self-guidance, while cluster head node guidance serves as a supplement and correction, without interrupting the local tracking process, achieving a seamless connection of "ensuring local real-time performance first, and then optimizing the global system later".

[0108] This mechanism leverages the low latency and high precision advantages of single-device self-guidance while avoiding passive waiting when devices have no instructions through fallback guidance from cluster head nodes. It balances real-time performance and global coordination capabilities without interrupting tracking.

[0109] Under this strategy, an additional preferred PTZ stabilization constraint adapted to civilian cameras is selected:

[0110] like Figure 5 and 6 As shown, there is a blind spot in the predicted location. This constraint is used to schedule idle PTZ cameras to turn to fill the blind spot in advance.

[0111] To address the issues of slow PTZ adjustment speed and inability to accurately identify targets during rotation, anti-shake constraints are set for both cluster head guidance and local node guidance:

[0112] Cluster head guidance constraint: After a camera completes one PTZ rotation, it must stop rotating for a preset time before receiving the next drive command; each time the global tracking scheme is updated, it must be ensured that at least one camera / camera in the tracking queue of each target has completed PTZ positioning before the scheme switch is executed to prevent target tracking from being interrupted.

[0113] • Local guidance constraint at nodes: When the target is about to exceed the field of view boundary, PTZ adjustment is started again. After rotation, the target is docked at the reserved edge of the field of view, leaving a margin to prevent the target from turning around or overshooting. If a single camera observes multiple targets at the same time, the tracking object is not switched unless the currently tracked target is completely outside the field of view, to avoid frequent rotation.

[0114] Neighborhood-oriented collaborative: This strategy is suitable for guiding and tracking drone targets that are about to fly into clusters and cross boundaries at low speeds.

[0115] All existing monitoring devices within the cluster are deployed together to form an overall control and detection boundary with an elliptical outline. The cluster head node calculates the remaining time for the target to fly out of the enclosed area by quantification, matches the unique neighboring cluster corresponding to the target's flight direction, and sends the target trajectory, predicted path, and flight parameters only to the head node of that neighboring cluster, without linking with unrelated neighboring clusters. The neighboring cluster head node schedules its idle devices in advance to complete the tracking relay preparation, achieving seamless connection across cluster boundaries, avoiding target loss at the boundary, and eliminating the resource waste caused by network-wide broadcasting.

[0116] This embodiment provides an algorithm for quantizing the fly-away time of a target cluster boundary:

[0117] Set the current position of the target velocity vector This cluster controls the center of the elliptical boundary. The major semi-axis a and minor semi-axis b of the ellipse, and the equation of the ellipse boundary:

[0118]

[0119] in, Given the shortest distance from the target along its own direction of motion to the boundary of the ellipse, the predicted time for the target to leave this cluster is:

[0120]

[0121] In the formula, Let be the Euclidean norm of the target's three-dimensional velocity vector, representing the target's velocity.

[0122] Cluster head node based on Determine the urgency of the target across clusters. The smaller the value, the higher the pre-scheduling priority.

[0123] This embodiment also provides a neighborhood cluster optimal orientation search algorithm:

[0124] Define an inter-cluster spatial correlation function to precisely match unique neighborhood clusters:

[0125]

[0126] In the formula: The directional matching degree between cluster i and candidate neighboring cluster j;

[0127] The shortest distance between the boundaries of the two clusters; This is the maximum distance from cluster i to all candidate neighboring clusters;

[0128] The angle between the target's direction of motion and the line connecting the two clusters;

[0129] , For the normalized weight coefficients, satisfying: .

[0130] Traverse all candidate clusters and select The corresponding cluster serves as the unique neighboring relay cluster, sending the target trajectory, predicted path, and flight parameters only to the head node of that neighboring cluster, without linking with unrelated neighboring clusters.

[0131] Specifically, such as Figure 7 As shown, the cluster head node of the first detection cluster only sends target information to the second detection cluster in the target flight direction, without initiating linkage with other detection clusters, thus achieving precise cross-cluster relay.

[0132] Cross-domain cloud collaboration: This strategy is suitable for guiding and tracking drone targets in high-speed, long-distance flight scenarios.

[0133] For drones flying at high speeds and traversing multiple detection clusters in a short time, the current cluster head node reports the target information to the cloud center. The cloud center predicts the long-distance flight path based on the target's motion vector and directly sends the target information to the distant cluster head nodes along the flight path, skipping the process of passing the information through intermediate clusters. The cluster head nodes along the target path pre-arrange the deployment of equipment within their groups to achieve long-distance early tracking, solving the problem of target loss caused by untimely relay of neighboring domains when high-speed targets cross large areas.

[0134] To address the short-term, multi-cluster movement characteristics of high-speed UAVs, the local cluster head node only uploads the target's high-speed motion status to the cloud center; the cloud center constructs a cross-domain collaborative cluster optimal matching model to achieve accurate long-distance cluster location.

[0135]

[0136] In the formula: k represents the source detection cluster currently reporting target information, and m represents the candidate remote detection cluster to be screened;

[0137] The score is the matching score between the target flight path in source cluster k and the distant candidate cluster m.

[0138] The spatial overlap length between the target flight path and cluster m; Total flight distance for the entire journey to the target;

[0139] The angle between the target heading and the azimuth of candidate cluster m;

[0140] Let m be the number of intermediate clusters that the target needs to pass through on its journey from source cluster k to candidate cluster m.

[0141] , , These are dynamically normalized weighting coefficients, with a sum of 1, adaptively configured based on the target speed, satisfying the following:

[0142]

[0143] The cloud center selects the cluster with the highest score as the cross-domain pre-scheduling target, skips the intermediate step-by-step transmission process, and directly issues pre-scheduling instructions to the remote clusters on the target's flight path. The remote cluster head nodes complete equipment deployment and tracking preparation in advance, realizing long-distance, large-span advance collaborative tracking and solving the problem of high-speed target cross-domain loss.

[0144] Specifically, such as Figure 8 As shown, for high-speed flying targets, the cloud center directly sends the target information to the remote detection clusters along the flight path, bypassing intermediate clusters and transmitting it step by step, thus achieving long-distance advance deployment and tracking.

[0145] Example 2

[0146] like Figure 2 As shown, this embodiment further elaborates on the specific architecture of a cluster-head-led multi-camera collaborative tracking system for UAV targets provided by the present invention. The system includes:

[0147] The edge perception module includes multiple edge perception nodes. The edge perception nodes are used to acquire video streams containing UAV target information and perform target recognition. They also perform coarse localization calculation of the target single point based on their corresponding node perception parameters and output the initial spatial position state data of the detected target.

[0148] The cluster head control module includes cluster head nodes that are periodically and dynamically elected from each detection cluster. These nodes are used to receive initial spatial position status data and perform source filtering and multi-camera cross-positioning calculations to generate the three-dimensional coordinate trajectory of the detected target.

[0149] The cloud center receives the 3D coordinate trajectories uploaded by each cluster head node, stitches and merges multiple 3D coordinate trajectories, performs relay detection and eliminates cluster positioning deviations to generate global trajectory information of the detected target, and reports the situation to the client.

[0150] In addition, the cloud center includes a pre-division unit for detector clusters; the pre-division unit is used to divide and generate multiple detector clusters based on the geographical distribution and field of view coverage of each edge sensing node.

[0151] Meanwhile, the cloud center includes a collaborative guidance and tracking unit; the collaborative guidance and tracking unit is used to make quantitative predictions based on the global trajectory information, the movement rate and activity range of the detected target, trigger preset collaborative command strategies in a differentiated manner, and issue scheduling commands to drive the corresponding edge perception nodes to continuously track the target.

[0152] While several specific implementation details are included in the foregoing discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

Claims

1. A cluster-head-led multi-camera collaborative tracking method for UAV targets, characterized in that the method... include: S1. Each edge sensing node acquires a video stream containing UAV target information and performs target recognition. Based on its own corresponding node sensing parameters, it performs coarse localization calculation of the target single point and outputs the initial spatial position state data of the detected target. S2. Based on the geographical distribution and field of view coverage of each edge sensing node, divide them into multiple detection clusters; and periodically and dynamically elect the optimal node as the cluster head node within each cluster. S3. The cluster head node receives the initial spatial position status data of the target being detected reported by the edge sensing nodes within the cluster, performs source filtering and multi-camera cross-positioning calculation, and generates the three-dimensional coordinate trajectory of the target being detected. S4. Each cluster head node uploads the three-dimensional coordinate trajectory of the target to be detected to the cloud center; the cloud center stitches and merges multiple three-dimensional coordinate trajectories, performs relay detection and eliminates cluster positioning deviations, generates global trajectory information of the target to be detected, and reports the situation to the client.

2. The cluster-head-led multi-camera collaborative tracking method for UAV targets according to claim 1, characterized in that, In step S2, the specific steps for periodically and dynamically electing the optimal node as the cluster head node within the cluster are as follows: Within the detection cluster, a five-dimensional weighted election model is constructed based on the real-time device status of each edge sensing node to calculate the comprehensive weight of each edge sensing node. ; The comprehensive weight W of all edge sensing nodes within the detection cluster is calculated iteratively periodically, and the comprehensive weight is selected. The largest edge-aware node is used as the cluster head node.

3. The cluster-head-led multi-camera collaborative tracking method for UAV targets according to claim 1, characterized in that, In step S3, multi-camera cross-localization calculation is performed based on the three-dimensional decoupled Kalman filter method. The specific steps are as follows: Based on the line-of-sight orientation and pitch angle of the edge sensing node, a rotation matrix is ​​constructed to transform the observations, states, and covariance in the ENU coordinate system to the local coordinate system of the noise principal axis. After decoupling each axis in the local space and performing Kalman iteration to configure the observation noise differently for each axis, the state and covariance are restored to the global ENU coordinate system through the inverse transformation of the original rotation matrix.

4. The cluster-head-led multi-camera collaborative tracking method for UAV targets according to claim 1, characterized in that, Step S4 also includes a collaborative guided tracking sub-step: Based on the global trajectory information, the movement speed and activity range of the detected target, the cloud center performs quantitative prediction, triggers preset collaborative command strategies in a differentiated manner, and issues scheduling commands to drive the corresponding edge perception nodes to continuously track the target.

5. The cluster-head-led multi-camera collaborative tracking method for UAV targets according to claim 4, characterized in that, The preset instruction coordination strategies include local autonomous coordination strategy, neighborhood-oriented coordination strategy, and cross-domain cloud coordination strategy.

6. The cluster-head-led multi-camera collaborative tracking method for UAV targets according to claim 5, characterized in that, The autonomous collaboration strategy of this domain is generated based on a dual-mode collaboration mechanism that prioritizes edge autonomy and optimizes cluster heads, and is superimposed with anti-shake constraints.

7. The method for collaborative tracking of UAV targets by multiple cameras under cluster head leadership according to claim 6, characterized in that, When the target being detected is about to fly out of the control boundary of the current detection cluster, the neighborhood-oriented cooperative strategy is triggered; when the target being detected crosses multiple detection clusters within the speed time threshold, the cross-domain cloud cooperative strategy is triggered.

8. A cluster-head-led multi-camera collaborative tracking system for unmanned aerial vehicles (UAVs), characterized in that, The system includes: The edge perception module includes multiple edge perception nodes. The edge perception nodes are used to acquire video streams containing UAV target information and perform target recognition, and perform coarse localization calculation of the target single point based on their corresponding node perception parameters, and output the initial spatial position state data of the detected target. The cluster head control module includes cluster head nodes that are periodically and dynamically elected within each detection cluster. These nodes are used to receive the initial spatial position status data and perform source filtering and multi-camera cross-positioning calculations to generate the three-dimensional coordinate trajectory of the detected target. The cloud center is used to receive the three-dimensional coordinate trajectories uploaded by each cluster head node, and to stitch and fuse multiple three-dimensional coordinate trajectories to perform relay detection and eliminate cluster positioning deviations in order to generate global trajectory information of the detected target, and report the situation to the client.

9. A cluster-head-led multi-camera collaborative tracking system for UAV targets according to claim 8, characterized in that, The cloud center includes a pre-division unit for detection clusters; the pre-division unit is used to divide and generate multiple detection clusters based on the geographical distribution and field of view coverage of each edge sensing node.

10. A cluster-head-led multi-camera collaborative tracking system for UAV targets according to claim 8, characterized in that, The cloud center includes a collaborative guidance and tracking unit; the collaborative guidance and tracking unit is used to make quantitative predictions based on the global trajectory information, the movement rate and activity range of the detected target, trigger preset collaborative command strategies in a differentiated manner, and issue scheduling commands to drive the corresponding edge perception nodes to continuously track the target.

Images