A method for detecting unmanned aerial vehicles based on laser radar

Abstract

This invention relates to a method for detecting unmanned aerial vehicles (UAVs) based on lidar. It addresses the problem of poor UAV detection and localization capabilities in complex airspace environments found in existing technologies. The invention utilizes lidar deployed at edge nodes to continuously scan the target airspace, acquiring raw 3D point cloud data packets. These packets are then segmented at the frame level, preprocessed, and annotated with 3D targets to construct a training sample set. A DSCBAM-RCNN network model is used to learn from the training set, resulting in a UAV detection model with target recognition capabilities. This model is then used to infer the 3D location information of the target UAV in space from the tested point cloud data, and applied to downstream tasks. This invention improves the detection and localization capabilities of UAVs in complex airspace environments by using lidar-based methods. This invention belongs to the field of UAV detection technology.

Description

Technical Field

[0001] This invention relates to a method for detecting unmanned aerial vehicles (UAVs), belonging to the field of UAV detection technology. Background Technology

[0002] With the gradual opening of low-altitude airspace resources and the rapid development of drone technology, drones have been widely used in logistics transportation, environmental monitoring, and security patrols. However, with the expansion of application scenarios, the problems of illegal drone flights, malicious interference, and potential flight safety hazards in complex environments are becoming increasingly prominent, posing higher demands on public safety, protection of important targets, and maintenance of airspace order. Therefore, there is an urgent need to develop a technological means with efficient and accurate detection capabilities to achieve rapid identification and spatial positioning of small low-altitude flying targets.

[0003] Existing drone detection methods primarily rely on single-modal sensors such as radio frequency (RF), electromagnetic, acoustic, or visual sensors. RF methods struggle to distinguish small, poorly characterized drone targets; video image-based detection methods are limited by lighting conditions, occlusion, and weather conditions, especially exhibiting decreased accuracy at night or at long distances. In contrast, LiDAR (Light Detection and Ranging) offers high spatial resolution, active imaging capabilities, and anti-jamming characteristics, making it suitable for target recognition in three-dimensional scenes and providing a new approach for the spatial positioning and identification of drones.

[0004] While existing 3D target detection technologies can be applied to the detection of conventional vehicles or pedestrians, they still face the problem of insufficient adaptability when directly used for the identification of small drones. This is mainly manifested in the following aspects: weak modeling ability for sparse areas of point clouds, significant decrease in detection accuracy as the target size decreases, and limitations in feature extraction of key areas.

[0005] Therefore, it is necessary to propose a lidar detection method suitable for UAV scenarios to improve the detection and positioning capabilities of UAVs in complex airspace environments. Summary of the Invention

[0006] To address the problem of poor detection and positioning capabilities of existing technologies for unmanned aerial vehicles (UAVs) in complex airspace environments, this invention proposes a UAV detection method based on lidar.

[0007] The technical solution adopted by the present invention to solve the above problems is as follows: The steps of the present invention include: Step 1: Continuously scan the UAV flight scene using LiDAR deployed at edge nodes to obtain raw LiDAR point cloud data packets; Step 2: Perform frame-level segmentation on the original laser point cloud data packet to obtain a set of continuous single-frame point cloud sequences; Step 3: Preprocess the continuous single-frame point cloud sequence set to obtain a normalized point cloud sequence set; Step 4: Perform 3D target annotation on the normalized point cloud sequence set to generate an annotation sample set containing 3D bounding box parameters. The annotation sample set is used as the training set for the UAV verification model. Step 5: Input the training set into the DSCBAM-RCNN network model for training to obtain the drone detection model; Step 6: Input the point cloud frame to be tested into the UAV detection model and output the detection results of the UAV in the target scene; Step 7: Use the detection results to indicate the precise location of the drone in the target scene.

[0008] Furthermore, frame-level segmentation employs a time window segmentation method, based on a fixed frame rate. The original laser point cloud data packet is segmented at time intervals. satisfy .

[0009] Furthermore, in the continuous single-frame point cloud sequence set, each frame of point cloud data is independently stored as a single data unit file. The data unit file stores the three-dimensional spatial coordinates and reflection intensity of the points in a unified field order, with each point as the unit.

[0010] Furthermore, the preprocessing steps in step 3 include: Step 301: Discrete point removal. From each frame of the continuous single-frame point cloud sequence set, remove outliers that are sparsely connected to the surrounding point clouds in space. Step 302: Ground point removal. From each frame of the continuous single-frame point cloud sequence set, remove the point cloud in the near-ground plane and retain the effective point cloud of the suspended target area in the air. Step 303: Reflection intensity standardization. The reflection intensity value of each point in each frame of the continuous single-frame point cloud sequence set is normalized to obtain a normalized point cloud sequence set.

[0011] Furthermore, the 3D bounding box parameters for 3D target annotation include: The target center coordinates represent the target's position in three-dimensional space; Target size parameters, representing the target's length, width, and height; Orientation angle indicates the direction of rotation of the target relative to the lidar coordinate system.

[0012] Furthermore, the DSCBAM-RCNN network model is an improvement upon the PointRCNN network model, specifically including: A density and semantic awareness farthest point sampling module is added, and the DS-FPS module is used to downsample the input point cloud, prioritizing the retention of representative points in sparse and semantically significant regions to reduce the overall point cloud size. An attention-enhanced point cloud feature encoding module is used to replace the original feature extraction network. The CBAM-Net++ module includes an encoding part and a decoding part. The encoding part consists of multiple point set extraction layers and channel-space attention modules stacked alternately; The decoding part consists of multiple feature propagation layers connected in series.

[0013] Furthermore, the DSCBAM-RCNN network model includes: CBAM-Net++ feature extraction module: performs multi-scale feature encoding on the keypoint set output by the DS-FPS module, and enhances the expressive power of foreground region points by fusing channels and spatial attention mechanisms, outputting high-dimensional point cloud features; 3D candidate bounding box generation module: Receives high-dimensional point cloud features output by the CBAM-Net++ module, calculates the foreground probability of each key point and predicts the target center position, and generates multiple 3D candidate bounding boxes; Candidate box fine regression module: Extracts point cloud features inside candidate boxes, performs target box parameter regression, and outputs the final UAV 3D detection results.

[0014] Furthermore, the output of the drone detection results in the target scene includes the center coordinates of the drone target's 3D bounding box, the size parameters of the drone target's 3D bounding box, the category label of the drone target, and the detection confidence score of the drone target.

[0015] The beneficial effects of this invention are: 1. This invention supports deployment and application in edge computing environments. After optimization through pruning and quantization compression, the detection model can achieve low-power, high-frame-rate inference on GPU embedded platforms, and has the ability to run offline. It is suitable for real-time low-altitude perception and linkage prevention and control systems.

[0016] 2. The UAV detection method based on LiDAR proposed in this invention integrates a point cloud density perception sampling strategy and an attention mechanism-guided feature encoding structure, which can improve the expressive ability and detection accuracy of small foreground targets in point clouds. At the same time, it has good edge deployment capability and can improve the accuracy of UAV detection in complex scenarios.

[0017] 3. This invention acquires target point cloud data using lidar and combines it with a 3D target detection algorithm to identify and determine the accurate location of the UAV. This method has high anti-interference capability, significantly improves the modeling and identification ability of UAV targets, and performs more accurately and reliably in complex airspace scenarios. Attached Figure Description

[0018] Figure 1 This is a flowchart of the present invention; Figure 2This is a schematic diagram of a scenario where a lidar deployed at an edge node collects point cloud data from a flying drone. Figure 3 This is a flowchart of the preprocessing process for a continuous single-frame point cloud sequence set; Figure 4 This is a schematic diagram of the structure of the attention-enhanced point cloud feature encoding module; Figure 5 This is a diagram of the DSCBAM-RCNN network model structure. Detailed Implementation

[0019] Specific Implementation Method 1: The steps of the UAV detection method based on lidar described in this implementation method include: Step 1: Continuously scan the UAV flight scene using LiDAR deployed at edge nodes to obtain raw LiDAR point cloud data packets; Step 2: Perform frame-level segmentation on the original laser point cloud data packet to obtain a set of continuous single-frame point cloud sequences; Frame-level segmentation uses a time window segmentation method, based on a fixed frame rate. The original laser point cloud data packet is segmented at time intervals. satisfy ; Each frame of point cloud data in the continuous single-frame point cloud sequence set is stored independently as a single data unit file. The data unit file stores the three-dimensional spatial coordinate values ​​of the points in a unified field order, with each point as the unit. and reflection intensity ; Step 3: Preprocess the continuous single-frame point cloud sequence set to obtain a normalized point cloud sequence set; The preprocessing steps include: Step 301: Discrete point removal. From each frame of the continuous single-frame point cloud sequence set, remove outliers that are sparsely connected to the surrounding point clouds in space. Step 302: Ground point removal. From each frame of the continuous single-frame point cloud sequence set, remove the point cloud in the near-ground plane and retain the effective point cloud of the suspended target area in the air. Step 303: Reflection intensity standardization. The reflection intensity value of each point in each frame of the continuous single-frame point cloud sequence set is normalized to obtain a normalized point cloud sequence set. Step 4: Perform 3D target annotation on the normalized point cloud sequence set to generate an annotation sample set containing 3D bounding box parameters. The annotation sample set is used as the training set for the UAV verification model. The 3D bounding box parameters for 3D target annotation include: Target center coordinates , indicating the position of the target in three-dimensional space; Target size parameters , representing the length, width, and height of the target; Orientation Angle , indicating the rotation direction of the target with respect to the lidar coordinate system; Step 5: Input the training set into the DSCBAM-RCNN network model for training to obtain the drone detection model; The DSCBAM-RCNN network model is an improvement upon the PointRCNN network model, including: A density and semantic awareness farthest point sampling module (DS-FPS module) is added. The DS-FPS module is used to downsample the input point cloud, and prioritizes the retention of representative points in sparse regions and semantically significant regions to reduce the overall point cloud size. An attention-enhanced point cloud feature encoding module (CBAM-Net++ module) is used to replace the original feature extraction network. The CBAM-Net++ module includes an encoding part and a decoding part. The encoding part consists of multiple point set extraction layers (SA layers) and channel-space attention modules (CBAM layers) stacked alternately; The decoding section consists of multiple feature propagation layers (FP layers) connected in series; The DSCBAM-RCNN network model consists of the following modules connected in sequence: DS-FPS module: Selects representative sparse keypoints from the original point cloud in the training set, and determines the reserved region by combining point density estimation and semantic response score; CBAM-Net++ feature extraction module: performs multi-scale feature encoding on the key point set output by the DS-FPS module, and enhances the expressive power of foreground region points by fusing channels and spatial attention mechanisms, and outputs high-dimensional point cloud features; 3D candidate bounding box generation module: Receives the high-dimensional point cloud features output by the CBAM-Net++ module, calculates the foreground probability of each key point and predicts the target center position, and generates multiple 3D candidate bounding boxes; Candidate box fine regression module: Extracts the point cloud features inside the candidate boxes, performs target box parameter regression, and outputs the final UAV 3D detection results; Step 6: Input the point cloud frame to be tested into the UAV detection model and output the detection results of the UAV in the target scene; The output of the drone detection results in the target scene includes: the center coordinates of the 3D bounding box of the drone target. 3D bounding box size parameters of UAV targets Category labels for drone targets and detection confidence scores for drone targets; Step 7: Use the detection results to indicate the precise location of the drone in the target scene.

[0020] The accurate position of the drone in the target scene is obtained through the center coordinates of the three-dimensional bounding box in the detection results, and is further applied to navigation, tracking or other downstream application tasks.

[0021] In this embodiment, the drone detection model has undergone pruning and quantization compression, making it suitable for low-power, high-frame-rate inference on edge device GPUs, and it also has the ability to run offline.

[0022] Example like Figure 1 As shown, the steps of a drone detection method based on lidar include: S101. The drone flight scene is continuously scanned by the lidar deployed at the edge node to obtain the original lidar point cloud data packet; like Figure 2 As shown in the figure, in this embodiment, the lidar shown in 101B is a rotating or solid-state laser scanning device, which is installed on a height platform shown in 101D. The horizontal viewing angle coverage is not less than 270°, which can realize periodic scanning of UAV targets within a radius of 100 meters, and the scanning frame rate is Hz. Flight scenarios include open areas, such as autonomous flying drones, low-altitude crossing drones, or targets in hovering states, as shown in 101A; In this embodiment, the lidar device is deployed on a fixed edge computing node as shown in 101C. The edge node integrates an embedded processing unit with GPU acceleration capabilities and has continuous power supply, environmental protection and local data storage capabilities. In this embodiment, the original laser point cloud data packet is encapsulated in the ROS standard message format and synchronously recorded as a ROS data packet file (Robot Operating System Bag, rosbag).

[0023] S102. Perform frame-level segmentation on the original laser point cloud data packet to obtain a continuous single-frame point cloud sequence set; The rosbag format file obtained from S101 is segmented into frames using the time window segmentation method, and then segmented at a fixed frame rate. The original laser point cloud data packet is segmented, with time intervals satisfying... ; After segmentation, a continuous single-frame point cloud sequence set is obtained, in the format of point cloud data (pcd) or binary data file (bin).

[0024] S103: Preprocess the continuous single-frame point cloud sequence set to obtain a normalized point cloud sequence set; like Figure 3 As shown, the preprocessing includes: Read a frame of point cloud data from a set of consecutive single-frame point cloud sequences; Step 301: Remove outliers from the point cloud frame that are sparsely connected to the surrounding point cloud in space. Step 302: Ground point removal. From the point cloud frame, remove the point cloud in the near-ground plane and retain the effective point cloud of the suspended target area in the air. Step 303: Normalize the reflection intensity. Normalize the reflection intensity value of each point in the point cloud frame and output a normalized point cloud sequence. Determine if the last frame has been processed. If not, continue reading one frame of point cloud data from the continuous single-frame point cloud sequence set; if so, output the normalized point cloud sequence set.

[0025] S104: Perform 3D target annotation on the normalized point cloud sequence set to generate an annotation sample set containing 3D bounding box parameters. The annotation sample set is used as the training set for the UAV detection model. In this embodiment, the 3D target annotation of the normalized point cloud sequence set is performed using a manual annotation method; In this embodiment, the 3D bounding box parameters include: Target center coordinates , indicating the position of the target in three-dimensional space; Target size parameters , representing the length, width, and height of the target; Orientation Angle , indicating the rotation direction of the target with respect to the lidar coordinate system; In this embodiment, the parameters generated after annotation are stored in a text file (txt).

[0026] S105: Input the training set into the DSCBAM-RCNN network model for training to obtain the UAV detection model; In this embodiment, the DSCBAM-RCNN network model is an improvement based on the PointRCNN network model. The main improvements are as follows: A density and semantic awareness farthest point sampling module (DS-FPS module) is added to downsample the input point cloud, prioritizing the retention of representative points in sparse and semantically significant regions to reduce the overall point cloud size; An attention-enhanced point cloud feature encoding module (CBAM-Net++ module) is used to replace the original feature extraction network. The CBAM-Net++ module includes an encoding part and a decoding part. like Figure 4 As shown, the encoding part consists of multiple SA layers and CBAM layers stacked alternately; the decoding part consists of multiple FP layers connected in series.

[0027] like Figure 5 As shown, The DSCBAM-RCNN network model includes: DS-FPS module: Selects representative sparse keypoints from the original point cloud in the training set, and determines the reserved region by combining point density estimation and semantic response score; CBAM-Net++ feature extraction module: performs multi-scale feature encoding on the key point set output by the DS-FPS module, and enhances the expressive power of foreground region points by fusing channels and spatial attention mechanisms, and outputs high-dimensional point cloud features; 3D candidate bounding box generation module: Receives the high-dimensional point cloud features output by the CBAM-Net++ module, calculates the foreground probability of each key point and predicts the target center position, and generates multiple 3D candidate bounding boxes; Candidate box fine regression module: Extracts the point cloud features inside the candidate boxes, performs target box parameter regression, and outputs the final UAV 3D detection results.

[0028] S106: Input the point cloud frame to be tested into the UAV detection model and output the detection result of the UAV in the target scene; In this embodiment, the drone detection model outputs the detection results of the drone in the target scene, including: the center coordinates of the three-dimensional bounding box of the drone target. 3D bounding box size parameters of UAV targets Category labels for drone targets and detection confidence scores for drone targets.

[0029] S107: The detection results are used to indicate the precise location of the drone in the target scene; In this embodiment, the accurate position of the UAV in the target scene is obtained through the center coordinates of the three-dimensional bounding box in the detection result, and is further applied to navigation, tracking or other downstream application tasks.

[0030] In this embodiment, the drone detection model undergoes pruning and quantization compression, making it suitable for low-power, high-frame-rate inference on edge device GPUs and capable of offline operation.

[0031] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for detecting unmanned aerial vehicles (UAVs) based on lidar, characterized in that, The specific steps include: Step 1: Continuously scan the UAV flight scene using LiDAR deployed at edge nodes to obtain raw LiDAR point cloud data packets; Step 2: Perform frame-level segmentation on the original laser point cloud data packet to obtain a set of continuous single-frame point cloud sequences; Step 3: Preprocess the continuous single-frame point cloud sequence set to obtain a normalized point cloud sequence set; In the continuous single-frame point cloud sequence set, each frame of point cloud data is independently stored as a single data unit file. The data unit file stores the three-dimensional spatial coordinate values ​​of points in a unified field order, with each point as the unit. and reflection intensity ; Step 4: Perform 3D target annotation on the normalized point cloud sequence set to generate an annotation sample set containing 3D bounding box parameters. The annotation sample set is used as the training set for the UAV verification model. Step 5: Input the training set into the DSCBAM-RCNN network model for training to obtain the drone detection model; Step 6: Input the point cloud frame to be tested into the UAV detection model and output the detection results of the UAV in the target scene; Step 7: Use the detection results to indicate the precise location of the drone in the target scene.

2. The method for detecting unmanned aerial vehicles based on lidar according to claim 1, characterized in that, Frame-level segmentation uses a time window segmentation method, based on a fixed frame rate. The original laser point cloud data packet is segmented at time intervals. satisfy .

3. The method for detecting unmanned aerial vehicles based on lidar according to claim 1, characterized in that, Each frame of point cloud data in the continuous single-frame point cloud sequence set is stored independently as a single data unit file. The data unit file stores the three-dimensional spatial coordinates and reflection intensity of the points in a unified field order, with each point as the unit.

4. The method for detecting unmanned aerial vehicles based on lidar according to claim 1, characterized in that, The preprocessing steps in step 3 include: Step 301: Discrete point removal. From each frame of the continuous single-frame point cloud sequence set, remove outliers that are sparsely connected to the surrounding point clouds in space. Step 302: Ground point removal. From each frame of the continuous single-frame point cloud sequence set, remove the point cloud in the near-ground plane and retain the effective point cloud of the suspended target area in the air. Step 303: Reflection intensity standardization. The reflection intensity value of each point in each frame of the continuous single-frame point cloud sequence set is normalized to obtain a normalized point cloud sequence set.

5. The method for detecting unmanned aerial vehicles based on lidar according to claim 1, characterized in that, The 3D bounding box parameters for 3D target annotation include: The target center coordinates represent the target's position in three-dimensional space; Target size parameters, representing the target's length, width, and height; Orientation angle indicates the direction of rotation of the target relative to the lidar coordinate system.

6. The method for detecting unmanned aerial vehicles based on lidar according to claim 1, characterized in that, The DSCBAM-RCNN network model is an improvement upon the PointRCNN network model, specifically including: A density and semantic awareness farthest point sampling module is added, and the DS-FPS module is used to downsample the input point cloud, prioritizing the retention of representative points in sparse and semantically significant regions to reduce the overall point cloud size. An attention-enhanced point cloud feature encoding module is used to replace the original feature extraction network. The CBAM-Net++ module includes an encoding part and a decoding part. The encoding part consists of multiple point set extraction layers and channel-space attention modules stacked alternately; The decoding part consists of multiple feature propagation layers connected in series.

7. The method for detecting unmanned aerial vehicles based on lidar according to claim 1, characterized in that, The DSCBAM-RCNN network model includes: CBAM-Net++ feature extraction module: performs multi-scale feature encoding on the keypoint set output by the DS-FPS module, and enhances the expressive power of foreground region points by fusing channels and spatial attention mechanisms, outputting high-dimensional point cloud features; 3D candidate bounding box generation module: Receives high-dimensional point cloud features output by the CBAM-Net++ module, calculates the foreground probability of each key point and predicts the target center position, and generates multiple 3D candidate bounding boxes; Candidate box fine regression module: Extracts point cloud features inside candidate boxes, performs target box parameter regression, and outputs the final UAV 3D detection results.

8. The method for detecting unmanned aerial vehicles based on lidar according to claim 1, characterized in that, The output of the drone detection results in the target scene includes the center coordinates of the drone target's 3D bounding box, the size parameters of the drone target's 3D bounding box, the category label of the drone target, and the detection confidence score of the drone target.

Images