A method for target recognition using local sparse point cloud of missile-borne / airborne laser imaging fuses

Abstract

This invention discloses a method for identifying targets in locally sparse point clouds of missile-borne / airborne laser imaging fuses, belonging to the field of laser fuse target identification technology. Based on a multi-coupled-field dynamic imaging detection model, this invention establishes a simulation system for point clouds of missile-borne / airborne linear array pushbroom laser imaging fuses to acquire locally sparse point cloud target data. According to the precise detonation strategy and neural network training and evaluation requirements, the data is labeled and divided, generating a large dataset with guiding annotation information to improve the recognition accuracy of the deep learning neural network. A cylinder pixel partitioning method is used to perform dimensionality reduction and feature transformation on the point cloud data, generating a two-dimensional feature pseudo-image. Feature processing is performed through a multi-scale balanced convolutional neural network feature extraction layer to extract and stitch together target features of different sizes, reducing target information loss. Target identification is achieved through the SSD detection method. This invention is applicable to the field of military imaging, improving the accuracy of locally sparse point cloud target identification.

Description

Technical Field

[0001] This invention relates to a method for identifying targets in local sparse point clouds of missile-borne / airborne laser imaging fuses, belonging to the field of laser fuse target identification technology. Background Technology

[0002] A fuze is a device that uses environmental, target, or platform information to control the detonation of munitions according to a predetermined strategy, while ensuring operational and launch safety. Laser fuzes have advantages such as good directionality, good coherence, small sidelobes, and strong resistance to electromagnetic interference, and are widely used in various weapon platforms. However, laser fuzes are susceptible to interference from environmental factors such as smoke, dust, rain, and snow, resulting in false alarms or missed alarms, causing premature detonation or malfunction of the munitions, and affecting their destructive effectiveness.

[0003] Laser imaging fuses acquire spatial location and echo intensity information of the target by receiving laser echoes reflected or scattered by the target, combining them into point cloud data. Point cloud data from smoke, dust, rain, and snow are characterized by nonlinearity and variable shape, while target point cloud data has clear geometric features and a fixed shape. Analysis and processing of point cloud data can effectively reduce the impact of environmental interference such as smoke, dust, rain, and snow. Unlike vehicle-mounted lidar imaging, laser fuse imaging involves variable rendezvous attitudes and is characterized by one-time, non-global imaging, making it difficult to acquire large amounts of data through experimentation. Due to cost and size limitations, laser imaging fuses have a relatively small number of emission lines, resulting in sparse point cloud data. Furthermore, to ensure optimal damage to the target and achieve a precise detonation strategy, the laser imaging fuse must output a detonation signal when acquiring relatively sparse local point cloud data of the target. Traditional point cloud target recognition methods primarily employ geometric features, grayscale variation features, keypoint neighborhood weight features, and multi-attribute combination features to identify local target features. These methods require high density in the target point cloud; for relatively sparse local point cloud data, the accuracy of traditional recognition methods is limited. Deep learning-based point cloud target recognition methods mainly use 3D voxel partitioning, point feature learning, and bird's-eye view dimensionality reduction methods to process and recognize target point clouds. While these methods offer higher recognition accuracy, they lack training datasets for laser imaging fuze application scenarios, making them difficult to use in practical applications of missile / airborne laser fuze imaging.

[0004] Therefore, effectively identifying locally sparse point cloud targets acquired by missile-borne / airborne laser imaging fuses with high recognition accuracy is of great significance for the precise damage control of weapon system platforms. Summary of the Invention

[0005] To address the issue of low accuracy in existing local sparse point cloud target recognition technologies, the main objective of this invention is to provide a method for recognizing local sparse point cloud targets in missile-borne / airborne laser imaging fuses. This method establishes a local sparse point cloud target dataset based on a laser imaging fuse point cloud acquisition simulation system. The dataset is then labeled using a completeness-based guided annotation method. Finally, a cylinder partitioning method and a multi-scale balanced convolutional neural network feature extraction layer are employed to identify local sparse point cloud targets acquired by the laser imaging fuse, thereby improving the accuracy of local sparse point cloud target recognition.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] This invention discloses a method for identifying targets in locally sparse point clouds of missile-borne / airborne laser imaging fuses. Based on a multi-coupled-field dynamic imaging detection model, a simulation system for point clouds of missile-borne / airborne linear array pushbroom laser imaging fuses is established. Dynamic parameters and laser fuse parameters are set for simulation to acquire locally sparse point cloud target data. The dataset is labeled and divided according to the requirements of precise detonation strategies and neural network training and evaluation needs. A large number of datasets with guiding annotation information are generated to improve the recognition accuracy of the deep learning neural network. A cylinder pixel partitioning method is used to perform dimensionality reduction and feature transformation on the point cloud data, generating a two-dimensional feature pseudo-map. Feature processing is performed through a multi-scale balanced convolutional neural network feature extraction layer to extract and stitch together target features of different sizes, reducing target information loss. Finally, target recognition is achieved through the SSD (Single Shot MultiBox Detector) detection method, thereby improving the accuracy of locally sparse point cloud target recognition.

[0008] This invention discloses a method for identifying targets in a local sparse point cloud of a missile-borne / airborne laser imaging fuze, comprising the following steps:

[0009] Step 1: Based on the multi-coupled field dynamic imaging detection model, establish a simulation system for acquiring laser imaging fuse point clouds and generate a local sparse point cloud target dataset.

[0010] A multi-coupled-field dynamic imaging detection model was established, including a laser emission and reception model, a target model, and a missile / aircraft target rendezvous model. A simulation system for acquiring point cloud data of missile-borne / aircraft-borne linear array pushbroom laser imaging fuses was constructed. Using laser ray tracing, path and distance information from laser emission to the target echo generation process was acquired. Through rendezvous simulation between the missile-borne / aircraft platform and the target, the laser linear array echo signal under motion conditions was obtained, generating local sparse point cloud target data for the missile-borne / aircraft-borne laser imaging fuse in a two-dimensional spatiotemporal environment. Based on the established point cloud acquisition simulation system, target point cloud data under different action positions and rendezvous attitudes were simulated and acquired by setting different rendezvous conditions, generating a local sparse point cloud target dataset.

[0011] Step 2: Label and divide the dataset according to the requirements of the precise detonation strategy and the network training and evaluation needs.

[0012] To meet the requirements of a precise detonation strategy, ensure the weapon system operates at vulnerable target locations, and achieve effective integration of the dataset with the neural network, the dataset needs to be processed and guided with annotations. The data processing and annotation methods used are as follows:

[0013] (1) Point cloud data cropping and format standardization. Point cloud cropping is performed on the target area to remove meaningless point cloud data, thereby improving the effective information density of subsequent neural network feature extraction. By reading and reorganizing the point cloud files, the single point cloud data file format of the simulation output is converted into multiple compatible formats to meet the subsequent network reading requirements.

[0014] (2) Standardization of point cloud coordinate system. The intersection angle of the target is defined differently under different coordinate systems. The coordinate system definition of the interface between point cloud data and neural network input data in the unified dataset is used. For neural network data reading interfaces with different coordinate systems, the format conversion is performed with the basic coordinate system as a reference.

[0015] (3) Ground filtering and target angle correction. Read the three-dimensional information of each point in the point cloud data, set the filtering threshold according to the target height information, filter out the ground point cloud, and obtain pure target point cloud data. Based on the target intersection angle obtained from the simulation initialization parameters, multiply the entire target point cloud by a rotation matrix to unify the target orientation. The target point cloud angle correction method reduces the subsequent target length and width estimation error and improves the labeling accuracy.

[0016] (4) Completeness-based labeling based on overlapping area. The complete bird's-eye view area of ​​the target is calculated based on the target's basic length and width parameters. By performing rectangular fitting on the three-dimensional parameters of the point cloud data, the imaging length and width parameters of the current target point cloud are obtained, and the bird's-eye view area of ​​the current point cloud target is calculated. The completeness is defined as the percentage of the current bird's-eye view area to the complete bird's-eye view area. Selective labeling is performed based on the different completeness of the point cloud targets: to reduce the false alarm rate of the weapon system platform, point cloud targets with high completeness are effectively labeled, while point cloud targets with low completeness are not labeled to prevent false alarms and premature detonation.

[0017] (5) To meet the training and evaluation requirements of neural networks, the labeled dataset is divided into training and test sets.

[0018] Step 3: Establish a deep learning neural network based on the column element partitioning method combined with a multi-scale balanced convolutional feature extraction layer.

[0019] A cylinder-based approach is employed to process point cloud data. The point cloud data space is uniformly divided into a grid of cylinders with fixed length, width, and infinite height. Features are extracted and fused from points within each cylinder, transforming the disordered and massive 3D point cloud features into ordered and lightweight 2D feature pseudo-images. This approach preserves a significant amount of target information while avoiding computationally intensive 3D convolution operations, ensuring high real-time performance. A multi-scale balanced convolutional feature extraction layer, consisting of a downsampling module, an upsampling module, and a connection module, extracts and stitches features of different sizes from the 2D feature pseudo-images, generating high-information-content, high-information-density feature pseudo-images, guaranteeing high recognition accuracy. A single-order multi-layer SSD detection method based on anchor box generation is used to perform target recognition and regression on the feature pseudo-images.

[0020] Step 4: Train the neural network based on data augmentation and transfer learning.

[0021] Based on the established local sparse point cloud target dataset and its annotation information, a deep learning neural network is trained. Data augmentation methods are used to expand the data volume, improve training reliability, and prevent overfitting. Transfer learning methods are employed to train the recognition of target data on the locally sparse point cloud dataset, building upon the relatively mature neural network weight parameters, thereby improving recognition accuracy and training efficiency.

[0022] Step 5: For the local sparse point cloud target recognition neural network trained in Step 4, evaluate the training using the loss function value, evaluate and verify it on the test set, and evaluate the recognition accuracy.

[0023] It also includes step six, which applies the local sparse point cloud target recognition method evaluated in step five to the laser imaging fuse of the missile / airborne platform. Under the premise of high recognition accuracy, the target feature is identified in the point cloud data acquired by the laser imaging fuse of the missile / airborne platform and the output action signal is controlled to realize a precise detonation control strategy and ensure the best damage effect on the target.

[0024] Beneficial effects:

[0025] 1. The present invention discloses a method for identifying targets in local sparse point clouds of missile-borne / airborne laser imaging fuses. By performing point cloud cropping, format expansion, coordinate transformation, and annotation operations on the point cloud data generated by the laser imaging fuse point cloud acquisition simulation system, the method realizes the connection between the dataset and the neural network input interface. By generating a large dataset, the method improves the recognition accuracy of the local sparse point cloud target recognition method.

[0026] 2. The present invention discloses a method for identifying local sparse point cloud targets in missile-borne / airborne laser imaging fuses. By using a completeness-based annotation method, a deep learning neural network is trained to identify local sparse point cloud targets, thereby achieving a precise detonation control strategy and meeting the needs of practical application scenarios. Attached Figure Description

[0027] Figure 1 This is a flowchart of a method for identifying targets in a local sparse point cloud of a missile-borne / airborne laser imaging fuze, as disclosed in this invention.

[0028] Figure 2a The point cloud images obtained by the laser imaging fuse point cloud acquisition simulation system in step one of this embodiment at intersection angle θ = 0° and different action positions w are:

[0029] Figure 2b The point cloud images obtained by the laser imaging fuse point cloud acquisition simulation system in step one of this embodiment at intersection angle θ = 45° and different action positions w are:

[0030] Figure 3 This is a schematic diagram of the simulation system for acquiring the laser imaging fuse point cloud in step one of this embodiment;

[0031] Figure 4 This is a diagram showing the point cloud data annotation effect in step two of this embodiment at an intersection angle θ = 0° and different integrity γ.

[0032] Figure 5 This is a schematic diagram of the column element division principle in step three of this embodiment;

[0033] Figure 6 This is a schematic diagram illustrating the principle of the multi-scale balanced convolutional neural network feature extraction layer in step three of this embodiment;

[0034] Figure 7 This is a graph of the neural network training loss function in step five of this embodiment;

[0035] Figure 8 This is a graph showing the neural network evaluation and recognition accuracy in step five of this embodiment;

[0036] Figure 9 This is a diagram showing the neural network target recognition effect of step five in this embodiment at different operating positions w with an intersection angle θ = 45°. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The technical problems solved by the present invention and its beneficial effects are also described. It should be noted that the described embodiments are only intended to facilitate understanding of the present invention and do not constitute any limitation thereof.

[0038] The target recognition method for local sparse point cloud of missile-borne / airborne laser imaging fuse disclosed in this invention is applicable to dynamic platforms such as loitering munitions and UAVs that do not have a roll flight attitude. This embodiment uses loitering munitions as the optimal application for illustration.

[0039] This embodiment discloses a method for target identification of local sparse point clouds in missile-borne / airborne laser imaging fuses, such as... Figure 1 As shown, it includes the following steps:

[0040] Step 1: Based on the multi-coupled field dynamic imaging detection model, establish a point cloud acquisition simulation system to generate a local sparse point cloud target dataset. Based on the missile / airborne linear array laser imaging fuze point cloud acquisition simulation system, define the target's bottom center as the origin, use the Y-axis pointing upwards, and the XZ plane as the ground's left-handed coordinate system. The angle between the target's orientation vector and the positive Z-axis in a clockwise direction is the intersection angle θ. The loitering munition starts at a Z-axis coordinate of -7m and stops at a Z-axis coordinate of w. The intersection parameters of the loitering munition and the target, and the laser imaging fuze parameters are set as shown in Table 1. The simulated point cloud image obtained under the conditions of a miss distance of 5m and an attack angle of 30° is shown in Table 1. Figure 2a , Figure 2b As shown, the principle of the point cloud acquisition simulation system is as follows: Figure 3 As shown, the process of laser imaging fuze detecting a target is simulated through the physical engine of the virtual platform, acquiring 1296 sets of target point cloud data and establishing a local sparse point cloud target dataset.

[0041] Table 1. Parameter Settings for Point Cloud Acquisition Simulation System

[0042]

[0043] Step 2: Label and divide the dataset according to the requirements of the precise detonation strategy and the network training and evaluation needs.

[0044] Based on the established local sparse point cloud target dataset, the dataset processing and annotation steps are as follows:

[0045] (1) Point cloud data cropping and format standardization. The obtained simulation point cloud data is coarsely cropped to a size of 40m×48m×7m to remove redundant point cloud data, increase the information content of the target point cloud, and enhance the training efficiency of the subsequent neural network. The point cloud data in txt file type output by the simulation system is read and the format is expanded according to the requirements to output multiple types of point cloud files such as pcd and bin, which are convenient for the subsequent neural network to read directly. In this embodiment, only bin file conversion is used.

[0046] (2) Point Cloud Coordinate System Standardization. Referring to the data format of the point cloud output by the simulation system, a left-handed coordinate system is defined as the standard coordinate system. The Y-axis points upwards, the XZ plane represents the ground, and the angle between the target orientation and the positive Z-axis (clockwise) is the intersection angle θ. If the coordinate system definition within the neural network differs from the standard coordinate system, a coordinate standardization transformation is required. In this embodiment, for input into the neural network for training, a coordinate standardization transformation is necessary:

[0047]

[0048] In the formula, X, Y, Z, and θ are parameters of the standard coordinate system, and X′, Y′, Z′, and γ are parameters of the neural network coordinate system.

[0049] (3) Ground filtering and target angle correction. Read the 3D coordinate data of the point cloud, referencing the target's 4m height information, and set a 0.5m height filtering threshold to remove ground point cloud data, obtaining a pure target point cloud. Based on the intersection angle θ, multiply the entire point cloud by a rotation matrix:

[0050]

[0051] Adjust the target angle so that the target is uniformly oriented in the positive Z-axis direction.

[0052] (4) Integrity-based labeling based on overlapping area. In this embodiment, the target's dimensions are 12m × 8m × 4m, and the target's complete bird's-eye view area S is 96m². 2 By performing rectangular fitting on the pure target point cloud obtained from the previous step with corrected angles, the bird's-eye view area S′ of the current target point cloud is calculated based on the length and width obtained from the fitting result. This leads to the target integrity γ = S′ / S × 100%. Based on different imaging integrity levels, the target point cloud is divided into six levels: 0%, 30%, 50%, 70%, 90%, and 100%. Target point cloud data with integrity of 50% or higher are effectively labeled, while target point cloud data with integrity below 50% are not labeled, achieving precise control of the explosion points. The labeling effect is as follows: Figure 4 As shown.

[0053] The dataset was divided into 900 sets of labeled target point cloud data as the training set and 396 sets of labeled target point cloud data as the test set, thus meeting the needs of subsequent neural network training and evaluation.

[0054] Step 3: Establish a deep learning neural network based on a voxel partitioning method combined with a multi-scale balanced convolutional feature extraction layer. A voxel grid partitioning method is used to divide the original point cloud image into 0.1m × 0.1m voxels. Feature transformation converts the 3D features of the point cloud into 2D voxel features. The principle of voxels is as follows: Figure 5As shown. The input feature pseudomap size of the feature extraction layer of the multi-scale balanced convolutional neural network is (4, 64, 496, 432). A three-layer downsampling module is used, reducing the feature pseudomap size by half with each downsampling layer. A three-layer upsampling module unifies the feature pseudomap size, and a concatenation module is used to stitch the feature pseudomaps together. The principle of the feature extraction layer is as follows: Figure 6 As shown, the recognition box generation adopts the single-stage multilayer detector (SSD) method, and the recognition box regression is performed based on anchor boxes of 12m×8m×4m.

[0055] Step 4: Train the neural network based on data augmentation and transfer learning.

[0056] The training parameters used are shown in Table 2. Based on conventional training, the number of training samples was increased and data utilization was improved through virtual object sampling, random translation, random rotation, and local and global data augmentation methods with Gaussian noise. Transfer learning was then performed on existing, well-established vehicle target recognition parameters to shorten the network training cycle and improve learning stability and recognition accuracy.

[0057] Table 2 Training parameters for deep learning neural networks

[0058]

[0059] Step 5: Training Process and Evaluation. The training progress and regression performance of the neural network are evaluated using the loss function graph during the training process. The loss function graph is shown below. Figure 7 As shown, the training basically converges after 60 generations. The trained local sparse point cloud target recognition neural network is evaluated on the test set, the number of correct and incorrect recognitions is counted, and the recognition accuracy is calculated as follows: Figure 8 As shown, the final accuracy reached 93.35%, with a single-frame point cloud processing time of 0.016s, and the recognition effect was as follows. Figure 9 As shown.

[0060] Step 6: Apply the local sparse point cloud target recognition method after training and evaluation to the laser imaging fuse of the loitering munition. Based on the point cloud data obtained by the laser imaging fuse of the loitering munition, target feature recognition is performed. Under the conditions of good real-time performance and high recognition accuracy, the best damage effect of the loitering munition is achieved.

[0061] In summary, this invention realizes a method for identifying locally sparse point cloud targets in missile-borne / airborne laser imaging fuses, which has a high recognition accuracy for locally sparse point cloud targets and can achieve precise detonation control.

[0062] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for target identification using a locally sparse point cloud of a missile-borne / airborne laser imaging fuze, characterized in that: Includes the following steps, Step 1: Based on the multi-coupled field dynamic imaging detection model, establish a simulation system for acquiring laser imaging fuse point clouds and generate a local sparse point cloud target dataset. Step 2: Label and divide the dataset according to the requirements of the precise detonation strategy and the needs of network training and evaluation; The implementation method for step two is as follows: The data processing and annotation methods used are as follows: (1) Point cloud data cropping and format unification: Point cloud cropping is performed on the target area to remove meaningless point cloud data, improve the effective information density of subsequent neural network feature extraction, and the single point cloud data file format of the simulation output is converted into multiple compatible formats by reading and reorganizing the point cloud file to meet the subsequent network reading requirements. (2) Point cloud coordinate system unification; the definition of the intersection angle of the target is different under different coordinate systems. The coordinate system definition of the interface between the point cloud data and the neural network input data in the unified dataset is used. For neural network data reading interfaces with different coordinate systems, the format conversion is performed with the basic coordinate system as a reference. (3) Ground filtering and target angle correction: Read the three-dimensional information of each point in the point cloud data, set the filtering threshold according to the target height information, filter out the ground point cloud, and obtain pure target point cloud data. According to the target intersection angle obtained by the simulation initialization parameters, multiply the entire target point cloud by the rotation matrix to unify the target orientation. Reduce the subsequent target length and width estimation error and improve the labeling accuracy by using the target point cloud angle correction method. (4) Integrity-based labeling based on overlapping area; The complete bird's-eye view area of ​​the target is calculated based on the target's basic length and width parameters. By performing rectangular fitting on the three-dimensional parameters of the point cloud data, the imaging length and width parameters of the current target point cloud are obtained, and the bird's-eye view area of ​​the current point cloud target is calculated. The completeness is defined as the percentage of the current bird's-eye view area to the complete bird's-eye view area. Selective labeling is performed according to the different completeness of the point cloud targets: in order to reduce the false alarm rate of the weapon system platform, point cloud targets with high completeness are effectively labeled, while point cloud targets with low completeness are not labeled in order to prevent false alarms and premature detonation. (5) Divide the labeled dataset into training and testing sets; Step 3: Establish a deep learning neural network based on the column element partitioning method combined with a multi-scale balanced convolutional feature extraction layer; Step 4: Train the neural network based on data augmentation and transfer learning; Step 5: For the local sparse point cloud target recognition neural network trained in Step 4, evaluate the training using the loss function value, evaluate and verify it on the test set, and evaluate the recognition accuracy.

2. The method for identifying targets in a local sparse point cloud of a missile-borne / airborne laser imaging fuze as described in claim 1, characterized in that: It also includes step six, which applies the local sparse point cloud target recognition method evaluated in step five to the laser imaging fuze of the missile / airborne platform. Under the premise of balancing real-time performance and high recognition accuracy, the target feature is identified and the output action signal is controlled based on the point cloud data acquired by the laser imaging fuze of the moving platform, so as to realize a precise detonation control strategy and ensure the best damage effect on the target.

3. The method for identifying targets in a local sparse point cloud of a missile-borne / airborne laser imaging fuze as described in claim 1, characterized in that: The implementation method for step one is as follows: A multi-coupled-field dynamic imaging detection model was established, including a laser emission and reception model, a target model, and a missile / aircraft target rendezvous model. A simulation system for acquiring point cloud data of missile-borne / aircraft-borne linear array pushbroom laser imaging fuses was constructed. Using the laser ray tracing method, path and distance information from laser emission to the emission of the target and the generation of the echo were obtained. Through rendezvous simulation between the missile-borne / aircraft-borne platform and the target, the echo signal of the laser linear array under motion conditions was acquired, and local sparse point cloud target data of the missile-borne / aircraft-borne laser imaging fuse was generated under spatiotemporal two-dimensional conditions. Based on the established point cloud acquisition simulation system, target point cloud data under different action positions and rendezvous attitudes were simulated and acquired by setting different rendezvous conditions, generating a local sparse point cloud target dataset.

4. The method for identifying local sparse point cloud targets using a missile-borne / airborne laser imaging fuze as described in claim 3, characterized in that: The implementation method for step three is as follows: A cylinder pixel partitioning method is used to process point cloud data. The point cloud data space is uniformly divided into a fixed-length, infinite-height cylinder pixel grid. Features are extracted and fused from points within each cylinder pixel, transforming the disordered and massive 3D point cloud features into ordered and lightweight 2D feature pseudo-images. This method retains more target information while avoiding computationally intensive 3D convolution operations, ensuring high real-time performance. A multi-scale balanced convolutional feature extraction layer, composed of a downsampling module, an upsampling module, and a connection module, extracts and stitches features of different sizes from the 2D feature pseudo-images, generating high-information-content, high-information-density feature pseudo-images, ensuring high recognition accuracy. A single-shot multi-box detector (SSD) method based on anchor box generation is used to perform target recognition and regression on the feature pseudo-images.

5. The method for identifying locally sparse point cloud targets using a missile-borne / airborne laser imaging fuze as described in claim 4, characterized in that: The implementation method for step four is as follows: Based on the established local sparse point cloud target dataset and annotation information, the established deep learning neural network is trained; data augmentation methods are used to expand the data volume, improve the reliability of training, and prevent overfitting; transfer learning methods are used to train the recognition of local sparse point cloud target data on the basis of relatively mature neural network weight parameters, thereby improving the recognition accuracy and training efficiency.

Images