An unmanned aerial vehicle intrusion prevention and control method and system based on a laser radar point cloud

Abstract

The application relates to the technical field of unmanned aerial vehicle invasion prevention and control, in particular to an unmanned aerial vehicle invasion prevention and control method and system based on a laser radar point cloud. The method comprises the following steps: collecting and storing a background voxel library of a target area; based on the background voxel library, filtering out background points in the laser radar point cloud in real time to obtain a foreground point cloud; clustering and feature extraction are performed on the foreground point cloud, and multi-target tracking is executed to obtain a target trajectory; unmanned aerial vehicle identification is performed based on the target trajectory, and an identification result is asynchronously output; through background modeling and online hash query, millisecond-level real-time filtering is realized, a difference-based inflation is effectively reserved in a detection airspace in combination with semantic perception, a multi-feature fusion tracking and an unsupervised identification model based on physical rules are introduced, a complete and engineering-deployable low-altitude unmanned aerial vehicle prevention and control technology closed loop from detection, tracking, identification to asynchronous output is constructed while false negatives and false alarms are significantly reduced.

Description

Technical Field

[0001] This invention relates to the field of drone intrusion prevention and control technology, and in particular to a drone intrusion prevention and control method and system based on lidar point clouds. Background Technology

[0002] Small drones are widely used in aerial photography, inspection, and logistics, but they also bring problems such as illegal intrusion, covert reconnaissance, and interference with the security of critical areas. In scenarios such as airports, energy facilities, important industrial parks, and border areas, unauthorized drone flights can easily threaten safe operations, necessitating the establishment of reliable drone intrusion monitoring and prevention technologies. Currently, common drone detection methods mainly include radar detection, radio signal monitoring, and video surveillance identification. However, video surveillance is easily affected by background interference at long distances, with small targets, and in complex lighting conditions, resulting in poor identification stability. Radio monitoring relies on drone communication signals and is difficult to effectively detect in silent flight or under signal jamming conditions. Traditional radar suffers from insufficient resolution in identifying low-altitude small targets. LiDAR, by actively emitting lasers and acquiring high-precision three-dimensional point cloud data, can provide a detailed description of the position, shape, and motion state of spatial targets, showing great potential in the field of low-altitude small target detection. Summary of the Invention

[0003] Therefore, it is necessary to provide a method and system for preventing drone intrusion based on lidar point clouds to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a drone intrusion prevention and control method based on lidar point clouds includes the following steps:

[0005] Step S1: Collect and store the background voxel library of the target area;

[0006] Step S2: Based on the background voxel library, filter out background points in the lidar point cloud in real time to obtain the foreground point cloud;

[0007] Step S3: Cluster and extract features from the foreground point cloud, and perform multi-target tracking to obtain the target trajectory;

[0008] Step S4: Identify the UAV based on the target trajectory and output the identification results asynchronously.

[0009] The present invention also provides a drone intrusion prevention and control system based on lidar point clouds, used to execute the drone intrusion prevention and control method based on lidar point clouds as described above. The drone intrusion prevention and control system based on lidar point clouds includes:

[0010] The background voxel library generation module is used to collect and store the background voxel library of the target area;

[0011] The background point filtering module is used to filter out background points in the lidar point cloud in real time based on the background voxel library to obtain the foreground point cloud.

[0012] The multi-target tracking module is used to cluster and extract features from the foreground point cloud, and perform multi-target tracking to obtain the target trajectory;

[0013] The drone identification module is used to identify drones based on target trajectories and output the identification results asynchronously.

[0014] The beneficial effects of this invention are as follows: By constructing a background voxel library of the target area offline and using this library to quickly filter the background of the LiDAR point cloud during real-time point cloud processing, fixed structural information in the environment is removed before the data enters the target analysis process, thus significantly reducing the data scale of subsequent point cloud processing; after obtaining the foreground point cloud, clustering and feature extraction operations are performed on the foreground point cloud, and a continuous target trajectory is formed by combining a multi-target tracking mechanism, so that the temporal change relationship of moving targets in space is stably expressed; on this basis, UAV identification is achieved by performing multi-dimensional feature determination on the target trajectory, and an asynchronous output mechanism is used to perform network alarm, data recording, and device linkage processing on the identification results, so that the identification calculation process and the result output process are independent of each other, thereby ensuring the continuous operation of the real-time detection process. Through the above processing flow, a complete processing link from point cloud background removal, target extraction, trajectory tracking to UAV identification and result output is realized, improving the processing efficiency and identification stability of LiDAR point cloud data in UAV intrusion prevention and control scenarios, and enabling the system to continuously complete target monitoring and intrusion identification tasks even in complex environments. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of a drone intrusion prevention and control method based on lidar point clouds.

[0016] Figure 2 This is a schematic diagram of a drone intrusion prevention and control system based on lidar point clouds.

[0017] Figure 3 A flowchart illustrating the laser radar-based drone control process;

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0021] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] To achieve the above objectives, please refer to Figures 1 to 3 A method for preventing drone intrusion based on lidar point clouds includes the following steps:

[0023] All specific values ​​involved in this embodiment are exemplary parameters used to clearly illustrate the technical operation process and are not the only limitation of the present invention.

[0024] Step S1: Collect and store the background voxel library of the target area;

[0025] Step S2: Based on the background voxel library, filter out background points in the lidar point cloud in real time to obtain the foreground point cloud;

[0026] Step S3: Cluster and extract features from the foreground point cloud, and perform multi-target tracking to obtain the target trajectory;

[0027] Step S4: Identify the UAV based on the target trajectory and output the identification results asynchronously.

[0028] In this embodiment of the invention, a rotating lidar is deployed around the perimeter of the target area to continuously scan the monitored area and collect multi-frame environmental point cloud data under conditions where no UAVs are flying. The collected point clouds are uniformly converted to the same spatial coordinate system, and the space is divided into a regular three-dimensional voxel grid according to a preset voxel size. Spatial statistics are performed on the point cloud within each voxel, recording the voxel center coordinates, number of points, and average reflection intensity. Spatial overlay is performed on multiple consecutive point clouds. Voxels that are occupied in all sampled frames are marked as stable background voxels, and voxel dilation is performed on these stable background voxels to expand their spatial coverage, forming a background voxel set. The background voxel set is written into a storage unit in the form of voxel coordinate indexes to obtain the background voxel library of the target area, which serves as the query basis for real-time point cloud background filtering in step S2.

[0029] During the monitoring and operation phase, the LiDAR continuously outputs real-time 3D point cloud data, and each frame of point cloud is projected onto the voxel space coordinate system established in step S1. The voxel index of each point cloud point is calculated according to the same voxel size as the background voxel library, and the background voxel library is accessed based on the voxel index. If the corresponding index exists in the background voxel library, the point is determined to be a background point and removed from the current point cloud; if the corresponding index is not recorded as a background voxel, the point is retained. By performing voxel index query and filtering operations point by point, a foreground point cloud set containing only dynamic object points is obtained. The foreground point cloud set is output to step S3 in chronological order for target clustering and trajectory construction.

[0030] Spatial density clustering is performed on the foreground point cloud obtained in step S2. A neighborhood radius parameter and a minimum point count threshold are set. For each point in the foreground point cloud, its neighboring points are searched. When the number of neighboring points reaches the threshold, a point cloud cluster is established and continuously expanded until no new neighboring points appear, thus obtaining multiple independent candidate target point cloud clusters. For each point cloud cluster, the centroid coordinates, 3D bounding box size, and point density are calculated to form a target feature vector. For the target feature vectors in continuous time frames, a matching cost matrix is ​​calculated based on positional distance and size difference. The matching and updating of candidate targets and historical trajectories is completed according to the minimum cost principle, thereby obtaining target trajectory data in continuous time series, and the trajectory information is passed to step S4.

[0031] The target trajectory data generated in step S3 is read, and the velocity sequence, altitude change sequence, and trajectory direction change are calculated for continuous position points in the trajectory. The trajectory duration, spatial movement range, and corresponding point cloud cluster size are also statistically analyzed. Based on preset scoring rules, scores are calculated for velocity range, altitude interval, trajectory smoothness, motion cycle characteristics, point cloud cluster size, and point density. These scores are then accumulated using fixed weights to obtain a comprehensive confidence score. When the comprehensive confidence score exceeds the identification threshold, the target trajectory is marked as a UAV target. The identification result is encapsulated into an output data frame containing a timestamp, trajectory coordinates, and confidence score, and written to an output queue. An independent processing thread reads the data frame from the queue and executes network alarm information transmission, trajectory data file writing, and camera device linkage control, thereby completing the asynchronous output of the UAV intrusion identification result.

[0032] Please refer to [link / reference needed] for further information. Figure 3 The flowchart illustrates the process of LiDAR drone control, which involves building a background voxel library, filtering background points in real time to obtain foreground point clouds, clustering and extracting features to complete target tracking, and finally identifying drones based on multi-dimensional trajectory scoring, and asynchronously outputting results for linked early warning.

[0033] Preferably, step S1 includes the following steps:

[0034] Step S11: Collect and process multi-frame background point clouds of the target area without drones to obtain a clean background point cloud set;

[0035] Step S12: Cluster the pure background point cloud and use a classification model to identify the tree cluster;

[0036] Step S13: Perform large-coefficient dilation on the voxels corresponding to the identified tree clusters and small-coefficient dilation on the voxels corresponding to the non-tree clusters to generate a background voxel library.

[0037] In this embodiment of the invention, a rotating scanning lidar is deployed around the perimeter of the target area to continuously scan the monitored airspace. Multiple frames of 3D point cloud data are collected during the time period when no UAV flying targets are confirmed within the airspace. Each frame of point cloud data is uniformly converted to a fixed spatial coordinate system, and the continuously collected point clouds are overlaid with a time series. The frequency of point occurrences at each spatial coordinate location is counted. When a point is recorded at a certain spatial location in all collected frames, that spatial location is determined as a stable background point. Subsequently, the stable background points are merged according to time frames, and duplicate points are removed to form a point cloud set containing only static environmental structures, thus obtaining a clean background point cloud set. This clean background point cloud set is then input into step S12 for spatial clustering processing.

[0038] Spatial density clustering is performed on the clean background point cloud obtained in step S11. A spatial neighborhood radius and a minimum point count threshold are set. A neighborhood search is performed on each point. When the number of points in the neighborhood reaches the threshold, a point cloud cluster is established and continuously expanded until no new points are added to the neighborhood, resulting in multiple independent point cloud clusters. The height range, point cloud density, vertical dispersion, and spatial distribution morphology parameters are calculated for each point cloud cluster. When the height range is greater than a preset threshold and the vertical dispersion is higher than the density threshold, the point cloud cluster is marked as a tree cluster; when the height range is lower than the threshold and the spatial distribution is planar, the point cloud cluster is marked as a non-tree cluster. After classification and labeling, a set of point cloud clusters with category labels is obtained and passed to step S13 for voxel construction.

[0039] The point cloud clusters with category labels obtained in step S12 are mapped to a 3D voxel grid space. The monitoring area is divided into 3D voxel units according to a preset voxel side length, and the index number of the voxel containing all points within each point cloud cluster is calculated. For point cloud clusters labeled as tree clusters, multiple layers of voxel units are expanded in six directions in the 3D grid to form a large-coefficient expansion region, with the voxel contained in the cluster as the central voxel. For point cloud clusters labeled as non-tree clusters, a single layer of voxels is expanded only around the central voxel to form a small-coefficient expansion region. All voxel indices obtained after expansion are summarized and a voxel index table is established, thereby generating a background voxel library covering the static structure of the target area. This background voxel library is written into a storage unit for subsequent background filtering steps.

[0040] Preferably, step S12 includes the following steps:

[0041] Step S121: Cluster the pure background point cloud to obtain multiple point cloud clusters;

[0042] Step S122: Extract the geometric, reflectivity, and spatial distribution features of each point cloud cluster to form a feature vector;

[0043] Step S123: Input the feature vector into the preset XGBoost model, output the probability that each point cloud cluster is a tree, and identify tree clusters based on the probability threshold.

[0044] In this embodiment of the invention, spatial density clustering is performed on the clean background point cloud obtained in step S11. All point cloud coordinate data are read according to a unified coordinate system, and the neighborhood search radius is set to 0.5 meters, with a minimum point count threshold of 20 points. Taking any unmarked point in the point cloud as the starting point, the Euclidean distance from that point to surrounding points is calculated in three-dimensional space. When the distance is less than the neighborhood radius, the point is recorded as a neighboring point. When the number of neighboring points reaches the minimum point count threshold, a point cloud cluster is established, and each neighboring point is used as a new search center to continue neighborhood expansion until no new neighboring points appear. The same processing procedure is performed on all point clouds, dividing the spatially continuous point set into multiple independent point cloud clusters, and recording the point coordinate set contained in each point cloud cluster to form a point cloud cluster list, which is then passed to step S122.

[0045] For each point cloud cluster obtained in step S121, multidimensional feature parameters are calculated. The minimum bounding box is calculated based on the three-dimensional coordinates of all points in the point cloud cluster, and the length, width, and height of the bounding box are recorded as geometric features. Then, the laser reflection intensity values ​​of all points within the point cloud cluster are statistically analyzed, and the average reflectivity and standard deviation of reflectivity are calculated as reflectivity features. Next, the spatial dispersion and vertical height distribution density are calculated based on the distribution range of each point in the point cloud cluster in the horizontal and vertical directions as spatial distribution features. The bounding box size, average reflectivity, standard deviation of reflectivity, spatial dispersion, and vertical distribution density are combined in a fixed order to form a numerical sequence, thereby forming the feature vector of the corresponding point cloud cluster. This feature vector is then input into step S123 for tree recognition calculation.

[0046] Read the feature vectors of the point cloud clusters generated in step S122 and input them into the pre-trained XGBoost classification structure. This classification structure consists of multiple decision trees. Each decision tree determines the node based on the geometric dimensions, reflectance statistics, and spatial dispersion parameters in the feature vector and outputs a tree category probability value. The probability values ​​output by all decision trees are weighted and accumulated to obtain the final tree probability. The obtained probability value is compared with a preset threshold of 0.6. When the probability value is greater than or equal to the threshold, the corresponding point cloud cluster is marked as a tree cluster; when the probability value is lower than the threshold, the corresponding point cloud cluster is marked as a non-tree cluster. After completing the probability calculation for all point cloud clusters, output a set of point cloud clusters with category labels and pass it to the subsequent voxel dilation processing step.

[0047] Preferably, step S2 includes the following steps:

[0048] Step S21: Load the background voxel library and build the hash index;

[0049] Step S22: Map each point of the real-time LiDAR point cloud to voxel space;

[0050] Step S23: Determine whether each voxel belongs to the background voxel through a hash query, filter out voxels belonging to the background, and output the foreground point cloud.

[0051] In this embodiment of the invention, the background voxel library generated in step S13 is read from the storage unit. The background voxel library records the three-dimensional voxel index coordinates and corresponding numbers of each background voxel. A three-dimensional voxel grid is established according to a preset voxel side length of 0.5 meters, and the integer coordinate indices of each background voxel are encoded in a fixed order. The voxel indices in the three directions are combined to form a unique integer key value. This key value is written into a hash index table, and a correspondence between the key value and the background voxel label is established in the hash table, so that each background voxel has a unique hash index. After all voxel key values ​​are written, a background voxel hash index structure is formed, and this index structure is loaded into memory for use in the real-time point cloud processing stage.

[0052] During the monitoring and operation phase, the LiDAR outputs real-time 3D point cloud data according to a fixed scanning cycle. The spatial coordinates of each point in the current frame's point cloud are read, and the space is discretized according to the voxel side length of 0.5 meters used in the background voxel library. By performing integer division operations on the voxel spacing scale for each point coordinate, the corresponding voxel integer index coordinates are obtained, and this voxel index is used as the location identifier of the voxel containing that point. Points within the same voxel are merged and recorded to obtain all voxel indices and their corresponding point sets in the current frame's point cloud. All voxel indices are sequentially passed to step S23 for background voxel query and filtering processing.

[0053] Read the voxel index set obtained in step S22 and convert the voxel coordinates into hash keys using the same index encoding method as in step S21. Use this key to perform a lookup operation in the background voxel hash index table. If a corresponding key exists in the hash table, mark all points within that voxel as background points and delete them from the current point cloud data; if no corresponding key exists in the hash table, retain the points within that voxel and record them as dynamic points. After performing the above query operation sequentially on all voxel indices in the current frame, summarize the remaining point set to form the foreground point cloud and output it in chronological order to the subsequent target clustering and trajectory analysis processing steps.

[0054] Preferably, step S3 includes the following steps:

[0055] Step S31: Cluster the foreground point clouds to obtain multiple candidate target point cloud clusters;

[0056] Step S32: Extract the centroid coordinates, bounding box size, and shape features of each candidate target point cloud cluster to generate a target point cloud cluster feature vector;

[0057] Step S33: Based on the feature vector of the target point cloud cluster, perform trajectory prediction and data association on the candidate target, and output the target trajectory.

[0058] In this embodiment of the invention, the foreground point cloud data output in step S23 is read, and spatial density clustering is performed on all points according to a unified spatial coordinate system. The neighborhood search radius is set to 0.6 meters, and the minimum number of points is set to 15. Starting with an unlabeled point, the three-dimensional spatial distance between that point and its surrounding points is calculated. When the distance is less than the neighborhood radius, the neighboring point is added to the same set, and the neighborhood search continues for the newly added point until no new neighboring points appear. A spatially continuous set of points is formed through continuous expansion, and this set is marked as an independent point cloud cluster. This processing procedure is repeated for all points in the foreground point cloud, ultimately resulting in multiple spatially independent candidate target point cloud clusters. The point sets of each point cloud cluster are then passed to step S32 for feature calculation.

[0059] Geometric feature calculations are performed on each candidate target point cloud cluster obtained in step S31. The three-dimensional coordinates of all points within the point cloud cluster are statistically analyzed, and the minimum and maximum coordinate values ​​in each of the three directions are calculated. The difference between the maximum and minimum values ​​determines the target's bounding box length, width, and height. Subsequently, the coordinates of all points within the point cloud cluster are summed and divided by the number of points to obtain the centroid coordinates of the point cloud cluster. The aspect ratio and height ratio are then calculated based on the dimensions of the bounding box in the three directions as shape features. The centroid coordinates, bounding box dimensions, and shape ratios are combined into a numerical sequence in a fixed order to generate the feature vector of the corresponding candidate target point cloud cluster. This feature vector is then input into step S33 for trajectory association processing.

[0060] A target trajectory record table is established for continuous time frames, recording the centroid position, estimated velocity, and corresponding number of each trajectory in the previous frame. The target point cloud cluster feature vector generated in step S32 is read, and its centroid coordinates are extracted as the target position in the current frame. The predicted position for the current frame is calculated based on the trajectory position and velocity information recorded in the previous frame, and the spatial distance between the predicted position and the centroid position in the current frame is calculated. A matching cost table is constructed according to the distance, and the target point cloud cluster with the smallest distance is associated with and updated with its corresponding trajectory number, while simultaneously recording the new centroid coordinates and timestamp. When a target maintains an association relationship across multiple consecutive frames, the centroid coordinates of the target in each frame are connected in chronological order to form a complete target trajectory, which is then output to the subsequent UAV recognition and processing steps.

[0061] Preferably, step S31 includes the following steps:

[0062] Step S311: Set clustering parameters and execute the DBSCAN clustering algorithm on the foreground point cloud;

[0063] Step S312: Extract each independent point cloud cluster from the clustering results as a candidate target point cloud cluster.

[0064] In this embodiment of the invention, the foreground point cloud data output in step S23 is read, and the three-dimensional coordinates of each point in the point cloud are uniformly numbered. DBSCAN clustering parameters are set, with the neighborhood search radius set to 0.6 meters and the minimum neighboring point threshold set to 15 points. All points are traversed in numerical order, and the three-dimensional spatial distance between each point and its surrounding points is calculated. The number of neighboring points with a distance less than 0.6 meters is counted. When the number of neighboring points reaches 15, the point is marked as a core point, and this core point and its neighboring points are grouped into the same cluster set. Simultaneously, the same neighborhood search expansion process is performed on the neighboring points until no new neighboring points appear. The clustering of the foreground point cloud is completed through the above density expansion process, resulting in a point cloud set data with cluster numbers.

[0065] Read the point cloud dataset with cluster numbers generated in step S311, and group the points according to the cluster numbers. Extract the point set corresponding to each cluster number, and merge all points belonging to the same number to form an independent point cloud cluster. Count the number of points in each point cloud cluster, and remove clusters with fewer than 10 points. Create a point cloud cluster record table for the remaining point sets according to the cluster numbers. Each record contains the point cloud cluster number and the 3D coordinate data of all points in the cluster. After extracting all cluster numbers, multiple independent point cloud clusters are obtained, and each point cloud cluster is output as a candidate target point cloud cluster to step S32 for feature extraction processing.

[0066] Preferably, step S33 includes the following steps:

[0067] Step S331: Predict the state of the existing trajectory and construct a comprehensive matching cost matrix;

[0068] Step S332: Based on the comprehensive matching cost matrix, the candidate target is associated with the predicted trajectory and the status of the successfully matched trajectory is updated.

[0069] In this embodiment of the invention, a target trajectory record table already established in historical frames is read. The trajectory record table stores the centroid coordinates, velocity vector, and timestamp of each trajectory in one frame. The velocity component is calculated based on the difference between the centroid coordinates of two consecutive frames, and the velocity component is used to linearly extrapolate the position of the current frame to obtain the predicted position coordinates of each trajectory at the current moment. Subsequently, the centroid coordinates of the candidate target point cloud cluster generated in step S32 are read, and the three-dimensional spatial distance between the centroid of each candidate target and the position of each predicted trajectory is calculated, while the bounding box size difference is also calculated. The spatial distance and size difference are numerically combined according to a fixed weight to generate the matching cost value between the trajectory and the candidate target, and arranged according to the trajectory number and the candidate target number to form a comprehensive matching cost matrix.

[0070] Read the comprehensive matching cost matrix generated in step S331 and perform matching processing in ascending order of cost value. For each predicted trajectory, find the candidate target number with the minimum cost value in the corresponding row of the matrix. When the cost value is less than the preset matching threshold, establish an association between the candidate target and the corresponding trajectory number, and write the centroid coordinates of the candidate target into the trajectory record table as the current frame position. At the same time, recalculate the velocity vector based on the current frame position and the previous frame position. For unmatched candidate targets, create a new trajectory number in the trajectory record table and record the initial centroid coordinates and time information. Save the centroid coordinate sequence of trajectories that have completed association updates in multiple consecutive frames in chronological order to form a complete target trajectory and output it to the UAV recognition processing step.

[0071] Preferably, step S4 includes the following steps:

[0072] Step S41: Based on the target trajectory, calculate the velocity score, altitude score, trajectory smoothness score, motion pattern score, size score, point cloud density score, and signal-to-noise ratio score, and then perform a weighted summation to obtain the overall confidence score of the UAV.

[0073] Step S42: When the overall confidence level exceeds a preset threshold, the target trajectory is determined to be a drone;

[0074] Step S43: Put the drone identification result into the output queue, and execute network alarm, data storage or camera linkage operation asynchronously by an independent thread.

[0075] In this embodiment of the invention, the target trajectory data output in step S33 is read, and the centroid coordinate sequence of consecutive frames in the trajectory is extracted. The displacement per unit time is calculated based on the difference in centroid coordinates between adjacent frames, the average velocity is calculated, and mapped to a velocity score according to a preset velocity range. The height components of all centroid coordinates in the trajectory are read, the average height is calculated, and a height score is generated according to a preset height range. Direction change statistics are performed on the trajectory coordinate sequence, and a trajectory smoothness score is calculated based on the change in the angle between adjacent motion vectors. A motion mode score is calculated based on the trend of trajectory movement direction and height change. A size score is generated by reading the bounding box size obtained in step S32. A point cloud density score is generated by counting the number of point cloud clusters in each frame. A signal-to-noise ratio score is generated by counting the dispersion of point cloud reflection intensity. All scores are accumulated according to fixed weights to obtain the overall confidence level of the UAV.

[0076] Read the overall confidence score of the UAV generated in step S41 and establish a target trajectory recognition judgment table. Set the overall confidence score threshold to 0.7, and compare the overall confidence score of each trajectory with this threshold. When the overall confidence score is greater than 0.7, mark the corresponding trajectory number as a UAV target in the trajectory recognition judgment table, and record the start time, current time, and trajectory centroid coordinate sequence of the trajectory. When the overall confidence score is less than or equal to 0.7, keep the unrecognized mark in the trajectory recognition judgment table. After completing the threshold comparison processing for all trajectories, output the trajectory number marked as a UAV target and its trajectory data as the UAV recognition result and pass it to step S43 for result output processing.

[0077] The drone identification results generated in step S42 are read, and identification result data frames are constructed according to the trajectory number, timestamp, trajectory centroid coordinate sequence, and comprehensive confidence level. The identification result data frames are written into a first-in-first-out (FIFO) output queue, and an independent processing thread is established for the output queue. The independent processing thread continuously reads the identification result data frames in the queue and performs three types of processing operations on each frame: encapsulating the identification results into network data packets and sending them to the monitoring center server via a UDP port; writing the trajectory data and time information to a local log file for persistent storage; and sending control commands to the gimbal camera device, causing the camera device to perform rotation and tracking shooting based on the trajectory centroid coordinates, thereby completing the asynchronous output of the drone identification results.

[0078] Preferably, step S43 includes the following steps:

[0079] Step S431: Encapsulate the drone identification result into an output frame and push it into the output queue;

[0080] Step S432: Retrieve output frames from the output queue based on an independent output processor thread;

[0081] Step S433: The output processor thread executes at least one of the following operations in parallel: sending the result to the network, saving the data to a file, and triggering camera linkage.

[0082] In this embodiment of the invention, the UAV identification result output in step S42 is read, and the corresponding target trajectory number, comprehensive confidence score, trajectory centroid coordinate sequence, and identification timestamp are extracted. An output frame is constructed according to a fixed data structure. The output frame fields include a trajectory number field, a timestamp field, a comprehensive confidence score field, and a trajectory coordinate sequence field. A length identifier and a verification identifier are generated for the output frame. After field filling is completed, the output frame is written to the tail of the output queue of the circular buffer structure. Simultaneously, the queue write pointer is updated, so that the output queue forms a sequence of output frames arranged in chronological order, serving as the data source for subsequent output processing threads.

[0083] An independent output processor thread is created in the data processing system, and a fixed memory buffer is allocated to this thread. The output processor thread continuously polls the read pointer of the output queue. When an unread output frame is detected in the output queue, the corresponding output frame is read from the head of the queue and copied to the thread buffer, while the queue read pointer is updated. After reading is completed, the trajectory number, timestamp, comprehensive confidence score, and trajectory coordinate sequence in the output frame are parsed into structured data, and an output task identifier to be executed is generated inside the output processor thread, thereby providing input data for the parallel output processing in step S433.

[0084] After parsing the output frame, the output processor thread generates three types of output tasks. The first task constructs a network data packet based on the output frame content and sends the identification result data to a designated port on the monitoring center server via the Ethernet interface. The second task creates a data file name according to the trajectory number and writes the timestamp, overall confidence score, and trajectory coordinate sequence to a log file on the local storage device. The third task calculates the horizontal and vertical angles of the gimbal camera based on the trajectory centroid coordinates and sends control commands via a serial communication interface to cause the camera to rotate towards the trajectory position. These three tasks are executed sequentially in a parallel task queue within the output processor thread, thus completing the output processing of the UAV identification results.

[0085] The present invention also provides a drone intrusion prevention and control system based on lidar point clouds, used to execute the drone intrusion prevention and control method based on lidar point clouds as described above. The drone intrusion prevention and control system based on lidar point clouds includes:

[0086] Background voxel library generation module 101 is used to collect and store the background voxel library of the target area;

[0087] Background point filtering module 102 is used to filter out background points in the lidar point cloud in real time based on the background voxel library to obtain the foreground point cloud.

[0088] The multi-target tracking module 103 is used to cluster and extract features from the foreground point cloud, and perform multi-target tracking to obtain the target trajectory;

[0089] The drone identification module 104 is used to identify drones based on target trajectories and output the identification results asynchronously.

[0090] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for preventing unmanned aerial vehicle (UAV) intrusion based on lidar point clouds, characterized in that, Includes the following steps: Step S1: Collect and store the background voxel library of the target area; Step S2: Based on the background voxel library, filter out background points in the lidar point cloud in real time to obtain the foreground point cloud; Step S3: Cluster and extract features from the foreground point cloud, and perform multi-target tracking to obtain the target trajectory; Step S4: Identify the UAV based on the target trajectory and output the identification results asynchronously.

2. The UAV intrusion prevention and control method based on lidar point clouds according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Collect and process multi-frame background point clouds of the target area without drones to obtain a clean background point cloud set; Step S12: Cluster the pure background point cloud and use a classification model to identify the tree cluster; Step S13: Perform large-coefficient dilation on the voxels corresponding to the identified tree clusters and small-coefficient dilation on the voxels corresponding to the non-tree clusters to generate a background voxel library.

3. The UAV intrusion prevention and control method based on lidar point clouds according to claim 2, characterized in that, Step S12 includes the following steps: Step S121: Cluster the pure background point cloud to obtain multiple point cloud clusters; Step S122: Extract the geometric, reflectivity, and spatial distribution features of each point cloud cluster to form a feature vector; Step S123: Input the feature vector into the preset XGBoost model, output the probability that each point cloud cluster is a tree, and identify tree clusters based on the probability threshold.

4. The UAV intrusion prevention and control method based on lidar point clouds according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Load the background voxel library and build the hash index; Step S22: Map each point of the real-time LiDAR point cloud to voxel space; Step S23: Determine whether each voxel belongs to the background voxel through a hash query, filter out voxels belonging to the background, and output the foreground point cloud.

5. The method for preventing and controlling unmanned aerial vehicle (UAV) intrusion based on lidar point clouds according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Cluster the foreground point clouds to obtain multiple candidate target point cloud clusters; Step S32: Extract the centroid coordinates, bounding box size, and shape features of each candidate target point cloud cluster to generate a target point cloud cluster feature vector; Step S33: Based on the feature vector of the target point cloud cluster, perform trajectory prediction and data association on the candidate target, and output the target trajectory.

6. The UAV intrusion prevention and control method based on lidar point clouds according to claim 5, characterized in that, Step S31 includes the following steps: Step S311: Set clustering parameters and execute the DBSCAN clustering algorithm on the foreground point cloud; Step S312: Extract each independent point cloud cluster from the clustering results as a candidate target point cloud cluster.

7. The UAV intrusion prevention and control method based on lidar point clouds according to claim 5, characterized in that, Step S33 includes the following steps: Step S331: Predict the state of the existing trajectory and construct a comprehensive matching cost matrix; Step S332: Based on the comprehensive matching cost matrix, the candidate target is associated with the predicted trajectory and the status of the successfully matched trajectory is updated.

8. The method for preventing and controlling unmanned aerial vehicle (UAV) intrusion based on lidar point clouds according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Based on the target trajectory, calculate the velocity score, altitude score, trajectory smoothness score, motion pattern score, size score, point cloud density score, and signal-to-noise ratio score, and then perform a weighted summation to obtain the overall confidence score of the UAV. Step S42: When the overall confidence level exceeds a preset threshold, the target trajectory is determined to be a drone; Step S43: Put the drone identification result into the output queue, and execute network alarm, data storage or camera linkage operation asynchronously by an independent thread.

9. The UAV intrusion prevention and control method based on lidar point clouds according to claim 8, characterized in that, Step S43 includes the following steps: Step S431: Encapsulate the drone identification result into an output frame and push it into the output queue; Step S432: Retrieve output frames from the output queue based on an independent output processor thread; Step S433: The output processor thread executes at least one of the following operations in parallel: sending the result to the network, saving the data to a file, and triggering camera linkage.

10. A drone intrusion prevention and control system based on lidar point clouds, characterized in that, For executing the drone intrusion prevention and control method based on lidar point clouds as described in claim 1, the drone intrusion prevention and control system based on lidar point clouds includes: The background voxel library generation module is used to collect and store the background voxel library of the target area; The background point filtering module is used to filter out background points in the lidar point cloud in real time based on the background voxel library to obtain the foreground point cloud. The multi-target tracking module is used to cluster and extract features from the foreground point cloud, and perform multi-target tracking to obtain the target trajectory; The drone identification module is used to identify drones based on target trajectories and output the identification results asynchronously.

Images