A linear array laser and linear array near-infrared composite imaging fuze target recognition method

Abstract

The application discloses a line array laser and line array near-infrared composite imaging fuze target recognition method, and belongs to the technical field of optical fuze target recognition. The file structure of a simulation data sample is standardized, a composite imaging data set of tanks, trucks and cars under the interference of camouflage, dust raising and smoke is generated, a spatial transformation matrix and a calibration parameter matrix are established based on simulation parameters, and line array laser point cloud and line array near-infrared image pixel-level registration is realized; a voxel fusion network is established, near-infrared image pixel-level features are used to expand the voxel feature dimension at the input level, laser point cloud is fused at the point level, and the recognition accuracy of a camouflaged target is improved; a BEV feature fusion network is established, laser point cloud features and near-infrared image features are fused in the BEV space through adaptive weight coefficients and guided by near-infrared features, and target region feature expression is strengthened. The application is suitable for the technical field of optical fuze target recognition, and improves the target recognition accuracy under the interference of camouflage, dust raising and smoke.

Description

Technical Field

[0001] This invention relates to a method for identifying targets using a linear array laser and a linear array near-infrared composite imaging fuze, belonging to the field of optical fuze target identification technology. Background Technology

[0002] Optical fuses utilize changes in the light field to acquire target information. Laser imaging fuses represent a significant research direction for next-generation intelligent optical fuses, obtaining rich spatial geometric information about the target area through imaging. However, in real battlefield environments, target camouflage and aerosol interference severely impact imaging quality, increasing the difficulty of target identification. Near-infrared imaging, compared to visible light and mid-to-far-infrared imaging, possesses stronger resistance to aerosol interference such as dust and smoke. Therefore, fully leveraging the complementary advantages of laser and near-infrared technologies to enhance the detection performance and anti-interference capabilities of optical fuses is a crucial issue worthy of in-depth research in the field of fuses.

[0003] Linear laser detectors can acquire three-dimensional unstructured point clouds of targets, while linear near-infrared detectors can acquire two-dimensional structured radiation intensity maps of targets. Establishing effective fusion mechanisms and feature extraction methods is a prerequisite for target recognition. Currently, deep learning has been extensively studied in the field of target recognition based on laser point clouds and images. There are three main fusion mechanisms between laser point clouds and images: data-level fusion, feature-level fusion, and decision-level fusion. Data-level fusion expands the input channels through point-by-point or pixel-by-pixel spatial mapping; feature-level fusion cascades or stitches feature maps at the feature map level to enrich feature representation; and decision-level fusion integrates the output results of two branches to improve prediction accuracy. Current fusion recognition algorithms are mainly geared towards open-source datasets, typically laser point clouds and RGB images in clean scenes. However, extensive research has not yet been conducted on target recognition applications using linear laser and linear near-infrared composite optical fuses. Target recognition in real, complex scenes with camouflage, dust, and smoke interference remains to be explored.

[0004] Due to the limited availability of real data samples for linear laser and near-infrared imaging fuses, and the high cost of testing in complex scenarios, a large number of data samples for linear laser and near-infrared fuses under camouflage, dust, and smoke interference conditions were obtained through a simulation platform. Based on this, the study aims to determine the mapping relationship between linear laser point clouds and near-infrared images according to their specific data representation, propose effective fusion strategies, and establish a robust fusion target recognition network model. This is of great significance for improving the precision detection capabilities of optical fuses in complex scenarios. Summary of the Invention

[0005] To address the issue of low accuracy in single-mode laser imaging fuze recognition under complex environments, the main objective of this invention is to provide a target recognition method for linear array laser and linear array near-infrared composite imaging fuzes. Based on a simulation system, a dataset of linear array laser and linear array near-infrared composite imaging of armored vehicles and transport vehicles under camouflage, dust, and smoke interference is established. A multi-modal fusion strategy incorporating a voxel fusion module and a BEV feature fusion module is employed to establish a 3D target recognition network model for linear array laser and linear array near-infrared fusion under camouflage, dust, and smoke interference, thereby improving the accuracy of target recognition under such interference.

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

[0007] This invention discloses a target recognition method for a linear array laser and linear array near-infrared composite imaging fuze. Based on a simulation system, it acquires linear array laser and linear array near-infrared imaging data samples, including three types of targets: tanks, trucks, and cars; and three types of interference environments: camouflage, dust, and smoke. A standardized data file structure is established, and a spatial transformation matrix and calibration parameter matrix are constructed to achieve pixel-level registration between the linear array laser point cloud and the linear array near-infrared image. A voxel fusion module is established to expand the feature dimension of the point cloud using pixel-level image features and convert it into voxel features. A dual-branch feature extraction backbone network is established to acquire multi-level feature maps for the laser branch and the near-infrared branch respectively. An LSS (Lift-Splat-Shoot) module is used to map image features into the BEV (Bird's Eye View) space and match it with the feature dimension of the point cloud. A BEV feature fusion module is established to fuse point cloud features and image features in the BEV space, outputting a composite feature map. A standard Anchor3DHead detection head is used to predict target classification, score, and 3D regression boxes to obtain the target recognition result, improving the accuracy of target recognition under such interference.

[0008] This invention discloses a method for target identification using a linear array laser and linear array near-infrared composite imaging fuze, comprising the following steps:

[0009] Step 1: Based on the simulation system, acquire linear laser and linear near-infrared imaging data samples, standardize the data file structure, establish the spatial transformation matrix and calibration parameter matrix, and realize pixel-level registration of linear laser point cloud and linear near-infrared image;

[0010] Based on the simulation-obtained linear array laser and linear array near-infrared imaging data samples of tanks, trucks, and automobiles, and following the processing flow of the KITTI-compatible mmdetection3d target detection framework library, the overall file system structure and path, data types, label formats, and parameter settings are defined to generate a dataset for training the target recognition network. According to the simulation parameters, the spatial transformation matrix is ​​determined through the geometric positional relationship between the laser origin and the near-infrared optical center. This spatial transformation matrix is ​​a combination of a three-dimensional rotation matrix and a translation vector, as shown in equation (1). Here, γ, ψ, and ε represent the rotation angles of the near-infrared optical center relative to the laser origin in the x, y, and z directions, respectively, and t... x t y t z These are the translation vectors in the x, y, and z directions, respectively.

[0011]

[0012] The calibration parameter matrix is ​​determined based on the linear array imaging system to achieve pixel-level registration between the linear array laser point cloud and the linear array near-infrared image, as shown in equation (2). Where F is the focal length of the near-infrared detector, h is the detection height, s is the pixel length, and f is the focal length of the near-infrared detector. c θ is the camera frame rate, m is the number of pixels in the image width direction, n is the number of pixels in the image height direction, θ is the attack angle, and v is the image height direction. r This represents the relative velocity.

[0013]

[0014] Step 2: Construct a modular network model for environments with camouflage, dust, and smoke interference, establish a fusion strategy between linear array laser point cloud and linear array near-infrared image, and build a target recognition network based on deep learning;

[0015] Dust and smoke interference increase noise in laser imaging, making it difficult to extract the geometric features of the target, while near-infrared images are relatively less affected. Camouflage interference affects near-infrared radiation intensity and the elevation information of the laser point cloud, but has little impact on the geometric features under the BEV perspective. Therefore, laser point cloud and near-infrared image fusion modules are used at both the data level and feature level, and a deep learning-based target recognition network is established. The target recognition network includes a voxel fusion network, a feature extraction backbone network, a BEV mapping module, a BEV feature fusion module, and a 3D detection head. The construction process of each part is as follows:

[0016] (1) Voxel Fusion Network. The laser point cloud and near-infrared image are input into the voxel fusion network. The laser point cloud is converted into a dynamic voxelized mesh, and the geometric features of the voxels, the geometric offset between the voxels and the mesh, and the feature offset are obtained. The near-infrared image is preprocessed to obtain pixel-level features. Through the spatial transformation matrix and calibration parameter matrix in step one, the laser point cloud is mapped point-by-point to the near-infrared image at the input level, and pixel-level features are obtained by interpolation. These features are concatenated with the geometric offset and feature offset to expand the voxel feature channels and output voxel vectors to the feature extraction backbone network.

[0017] (2) Feature extraction backbone network. A dual-branch backbone network is adopted. The laser branch adopts the SECOND+FPN backbone architecture. The voxel vectors output by the voxel fusion network are input into the laser branch to obtain multi-level 3D features. The near-infrared branch adopts the cascaded Swin Transformer+FPN backbone architecture to extract multi-level image features. The laser point cloud branch and the near-infrared image branch are executed independently to obtain multi-level feature maps of the point cloud stream and the image stream, and output them to the BEV mapping module.

[0018] (3) BEV mapping module. A dual-branch mapping network is adopted. The laser branch converts the multi-level 3D features extracted from the feature extraction backbone network into the BEV space through convolutional layers. The near-infrared branch uses the LSS module to perform 3D depth estimation on the multi-level image features output from the feature extraction backbone network and projects them into the BEV space. Convolutional layers are used to unify the image feature map size and the point cloud feature map size, match the feature dimensions, and output the dual-branch BEV features to the BEV feature fusion module.

[0019] (4) BEV Feature Fusion Module. The dual-branch BEV features in the BEV mapping module are adaptively output as near-infrared image feature weights through multi-layer convolutional layers and residual connection structures. This guides the fusion of laser point cloud BEV features and near-infrared image BEV features in the target area, deepening the feature expression of the target area, obtaining BEV spatial depth fusion features, and outputting them to the 3D detection head.

[0020] (5) 3D Detection Head. A series of multi-sized 3D bounding boxes generated in the BEV space are used. During training, positive and negative samples are obtained by calculating the 3D intersection-union ratio (IU) between the generated boxes and the ground truth boxes. The BEV fusion features in the BEV feature fusion module output predicted values ​​through a linear layer, including detection boxes, classification, and scores. These are iterated using a loss function, which is divided into four parts: classification loss, detection box regression loss, detection box orientation loss, and image depth estimation loss. During prediction, a detection threshold is set to filter the predicted scores. The multi-sized 3D bounding boxes are subjected to non-maximum suppression to obtain the final recognition result.

[0021] Step 3: Train the target recognition network constructed in Step 2, establish network performance evaluation indicators, evaluate the training results, and obtain the target recognition model.

[0022] The dataset generated in step one is divided into a training set and a test set. The target recognition network constructed in step two is trained on the training set. Training is completed when the total loss value is less than α, where α ∈ (0, 0.15). The 3D target detection performance evaluation metric mAP (mean Average Precision) is used. For three target classes (tanks, trucks, and cars), under three interference conditions (camouflage, dust, and smoke), the precision value under the β coverage threshold is calculated using three metric types: 3D projection onto 2D detection boxes, detection boxes from the BEV perspective, and 3D detection boxes, where β ∈ (0, 1). The model is tested using the test set, and the precision and recall are statistically obtained. Based on this, the overall average precision is calculated to evaluate the target recognition network performance. An overall average precision greater than μ proves the model is effective, where μ ∈ (80%, 100%). Models with an overall average precision less than μ are discarded, and hyperparameters are readjusted for retraining. The effective model is exported in a standard format for subsequent engineering applications.

[0023] It also includes step four, which involves deploying the target recognition model from step three onto a linear laser and linear near-infrared composite imaging fuze to identify tank, truck, and car targets using the acquired laser point cloud and near-infrared images. This improves the recognition accuracy in camouflaged, dusty, and smoke-filled environments, enabling precise detonation control.

[0024] Beneficial effects:

[0025] 1. The present invention discloses a target recognition method for a linear array laser and linear array near-infrared composite imaging fuze. By standardizing the file structure of simulation data samples, a composite imaging dataset of typical ground targets such as tanks, trucks and cars under camouflage, dust and smoke interference environments is generated. Based on simulation parameters, a spatial transformation matrix and a calibration parameter matrix are established to achieve pixel-level registration of linear array laser point cloud and linear array near-infrared image.

[0026] 2. The present invention discloses a target recognition method for a linear array laser and linear array near-infrared composite imaging fuze. By establishing a voxel fusion network, the voxel feature dimension is expanded at the input level using pixel-level features of near-infrared images, and point-level fusion is performed with the laser point cloud to improve the recognition accuracy of camouflaged targets.

[0027] 3. The present invention discloses a target recognition method for a linear array laser and linear array near-infrared composite imaging fuze. By establishing a BEV feature fusion network, the laser point cloud features and near-infrared image features are fused with adaptive weight coefficients in the BEV space, thereby enhancing the feature expression of the target area and improving the target recognition accuracy under dust and smoke interference. Attached Figure Description

[0028] Figure 1 This is a flowchart of a target recognition method for a linear array laser and linear array near-infrared composite imaging fuze disclosed in this invention;

[0029] Figure 2 This is the pixel-level registration result of the tank target in step one of this embodiment.

[0030] Figure (a) shows the camouflage interference situation, Figure (b) shows the dust interference situation, and Figure (c) shows the smoke interference situation.

[0031] Figure 3 This is the pixel-level registration result of the truck target in step one of this embodiment.

[0032] Figure (a) shows the camouflage interference situation, Figure (b) shows the dust interference situation, and Figure (c) shows the smoke interference situation.

[0033] Figure 4 This is the pixel-level registration result of the vehicle target in step one of this embodiment.

[0034] Figure (a) shows the camouflage interference situation, Figure (b) shows the dust interference situation, and Figure (c) shows the smoke interference situation.

[0035] Figure 5 This is a network architecture diagram of the voxel fusion module in step two of this embodiment;

[0036] Figure 6 This is the backbone network architecture diagram for step two, point cloud feature extraction, in this embodiment.

[0037] Figure 7 This is the backbone network architecture diagram for image feature extraction in step two of this embodiment;

[0038] Figure 8 This is the network architecture diagram of the BEV feature fusion module in step two of this embodiment;

[0039] Figure 9 This is the training loss change curve in step three of this embodiment;

[0040] Figure 10 To visualize the tank target identification results in step three of this embodiment,

[0041] Figure (a) shows the camouflage interference situation, Figure (b) shows the dust interference situation, and Figure (c) shows the smoke interference situation.

[0042] Figure 11 To visualize the truck target recognition results in step three of this embodiment,

[0043] Figure (a) shows the camouflage interference situation, Figure (b) shows the dust interference situation, and Figure (c) shows the smoke interference situation.

[0044] Figure 12 To visualize the vehicle target recognition results in step three of this embodiment,

[0045] Figure (a) shows the camouflage interference situation, Figure (b) shows the dust interference situation, and Figure (c) shows the smoke interference situation. Detailed Implementation

[0046] 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.

[0047] The target recognition method of linear array laser and linear array near-infrared composite imaging fuze disclosed in this invention is applicable to ground target detection application scenarios of dynamic platforms with no roll flight attitude, such as loitering munitions and UAVs. This embodiment uses a loitering munition as an example for illustration.

[0048] This embodiment discloses a target recognition method for a linear array laser and linear array near-infrared composite imaging fuze, such as... Figure 1 As shown, it includes the following steps:

[0049] Step 1: Based on the simulation system, acquire linear laser and linear near-infrared imaging data samples, standardize the data file structure, establish the spatial transformation matrix and calibration parameter matrix, and realize pixel-level registration of linear laser point cloud and linear near-infrared image.

[0050] Based on simulation data samples of linear array laser and linear array near-infrared imaging, a standardized dataset compatible with the KITTI style and the mmdetection3d target detection framework was established. The file system structure and path were determined, with point cloud data in .bin format, image data in .png format, and label data in .txt format. Coordinate system parameters were defined, including laser coordinate system, camera coordinate system, and pixel coordinate system, with the laser coordinate system set as the global world coordinate system. Pixel-level registration of laser point clouds and near-infrared images was achieved based on the spatial transformation matrix and parameter calibration matrix. The tank target registration visualization is shown below. Figure 2 As shown, the truck target registration visualization is as follows: Figure 3 As shown, the vehicle target registration visualization is as follows: Figure 4 As shown in the figure. A total of 1440 sets of laser point cloud and near-infrared image data samples were ultimately generated for subsequent target recognition algorithm training and testing.

[0051] Step 2: Construct the overall architecture of the target recognition network based on deep learning, establish a fusion strategy between linear laser point cloud and linear near-infrared image, and build a modular network model for camouflage, dust and smoke interference environments.

[0052] The target recognition network for linear laser and linear near-infrared composite imaging includes a voxel fusion network, a feature extraction backbone network, a BEV mapping module, a BEV feature fusion module, and a 3D detection head. The specific construction process of each part is as follows:

[0053] (1) Voxel Fusion Network. Voxel fusion networks are used to expand the point cloud feature dimension from pixel-level image features and convert them into voxel features. They are built upon a preprocessing network based on VoxelNet, such as... Figure 5 As shown. The input point cloud has a dimension of 3. The voxel fusion network converts the input point cloud into non-empty voxels, using a dynamic voxelization mesh. The point cloud within a voxel undergoes max pooling and point-by-point mapping to obtain voxel features. Then, the feature offset and geometric spatial offset between the original point cloud and the voxel mesh are calculated point-by-point. The image size is 1×960×1024 pixels. Image pixel-level features are obtained through preprocessing such as normalization and padding. These features are then mapped point-by-point onto the image pixel feature map using the spatial transformation matrix and calibration parameter matrix from step one. Image information is obtained through bilinear interpolation. Finally, the original point cloud features, feature offset features, geometric spatial offset features, and image pixel features are concatenated. The resulting fused information is used as the input to the point cloud branch and provided to the subsequent feature extraction backbone network.

[0054] (2) Feature Extraction Backbone Network. A dual-branch backbone network architecture is adopted, with the laser point cloud branch and the near-infrared image branch executing independently. The laser point cloud branch is designed based on the SECOND processing flow, using dynamic SECOND+FPN to acquire multi-layer 3D features, such as... Figure 6 As shown. By continuously performing convolution, normalization, activation, and downsampling operations on the voxel vectors output from the voxel fusion module, multi-level 3D feature maps are obtained. These multi-level feature maps are then output to a fixed size through deconvolution upsampling in the FPN layer. The near-infrared image branch employs a cascaded Swing Transformer feature encoder architecture, as shown... Figure 7 As shown, the network acquires multi-scale feature maps through four stages. Each stage uses a cascaded Swing Transformer module, which cascades window attention and sliding window attention respectively. The resulting multi-level features are integrated through a feature pyramid structure. The feature maps are then upsampled, averaged pooled, and dimensionality transformed. Finally, the multi-level features are concatenated and convolved to obtain the image branch feature vectors and output the BEV mapping module.

[0055] (3) BEV mapping module. The laser branch transforms multi-layer 3D features into BEV space through multi-layer convolution. The image branch first generates a view frustum space, performs depth estimation on the image features in the view frustum pixel by pixel, calculates the depth probability density by predicting the depth distribution, and maps it to 3D space using the spatial transformation matrix in step one to generate a 3D feature map and project it into the BEV mesh to obtain the image BEV space features, which are aligned with the point cloud BEV feature space for subsequent fusion.

[0056] (4) BEV Feature Fusion Module. Its structure is as follows: Figure 8 As shown, the BEV spatial features of the image obtained from the BEV mapping module are bilinearly interpolated to obtain a unified BEV feature space size. The point cloud features are dimension-matched using a 1×1 convolution operation. Then, the image features and point cloud features are subjected to two residual connection operations to gradually enhance the fusion features of the target region. Learnable weight parameters are added to adjust the proportion of image features. Finally, the fusion feature map is output through the standard structure of convolution, BatchNorm, and ReLU, which will be fed into the 3D detection head network for prediction.

[0057] (5) 3D Detection Head. The standard Anchor3DHead detection head is used to generate a series of multi-size 3D detection boxes in the BEV space. The 3D intersection-union ratio (IU) of the generated boxes and the ground boxes is calculated. The IU value is greater than 0.6 and the value is less than 0.4. The BEV fusion features are output as predicted values ​​through a linear layer and iteratively trained using a loss function. The loss function includes four aspects, which are obtained by summing the classification loss, regression loss, orientation loss and depth loss respectively. The classification loss adopts the FocalLoss loss function, which can better improve the negative effect of uneven distribution of positive and negative samples, as shown in Equation (3). The regression loss adopts the SmoothL1 Loss loss function, which uses the difference between the generated box and the ground box as input, as shown in Equation (4). The orientation prediction is used to output the target orientation, which is divided into forward and backward. The Cross Entropy Loss binary classification loss function is used, as shown in Equation (5). The depth loss supervises the image depth prediction through the point cloud depth value and adopts the Binary Cross Entropy Loss loss function, as shown in Equation (6).

[0058]

[0059]

[0060]

[0061]

[0062] Step 3: Train the target recognition network constructed in Step 2, establish network performance evaluation indicators, and evaluate the training results.

[0063] The dataset generated in step one was divided into 800 training sets and 800 test sets. The object recognition network constructed in step two was trained on a server with the following hardware parameters: CPU: i9-13900HX, 32 Cores; GPU: GTX3090, 24GB. The training configuration hyperparameters are shown in Table 1. The training reached convergence when the loss value was less than 0.1 after 35,000 total iterations. Figure 9 As shown.

[0064] Table 1. Training hyperparameters of the target recognition network

[0065]

[0066] The network performance evaluation metric adopted is the 3D target detection performance evaluation metric mAP. For tanks, trucks, and cars under three interference conditions (camouflage, dust, and smoke), three metric types were used: 3D bounding boxes projected onto 2D boxes, bounding boxes from the BEV's perspective, and 3D bounding boxes. AP values ​​were calculated at coverage thresholds of 0.7 and 0.5. The test set was used for testing, and the model performance was statistically analyzed. The relevant parameters are as follows:

[0067] TP: The predicted result is positive, and the actual sample is positive.

[0068] FP: The predicted result is positive, but the actual sample is negative.

[0069] FN: The predicted result is negative, but the actual sample is positive.

[0070] Precision and recall are calculated as shown in Equation (7):

[0071]

[0072] First, the detection score (TP) of each frame is calculated. After sorting, 40 recall points are uniformly sampled. Then, the TP, FP, and FN of the test set are calculated again at the 40 recall levels. Finally, the area under the precision and recall curves is calculated to obtain the AP value. The overall precision is the average AP of the three categories. The network performance evaluation results are shown in Table 2, and the tank target recognition results are visualized as follows: Figure 10 As shown, the truck target recognition results are visualized as follows: Figure 11 As shown, the vehicle target recognition results are visualized as follows: Figure 12As shown in the figure. The results show that tanks, trucks, and cars can be effectively identified under camouflage, dust, and smoke interference environments. The accuracy of tank target identification is 97.29%, truck target identification is 96.29%, and car target identification is 95.32%, with an overall average accuracy of 96.30%. The accuracy of over 95% proves the effectiveness of the model, and the model is exported to a standard format.

[0073] Table 2 Network performance evaluation results

[0074]

[0075] In summary, the target identification method for a linear array laser and linear array near-infrared composite imaging fuze disclosed in this embodiment can effectively identify typical armored vehicles and transport vehicles such as tanks, trucks, and automobiles under camouflage, dust, and smoke interference environments, and has high identification accuracy.

[0076] 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 linear array laser and linear array near-infrared composite imaging fuze, characterized in that: Includes the following steps: Step 1: Based on the simulation system, acquire linear laser and linear near-infrared imaging data samples, standardize the data file structure, establish the spatial transformation matrix and calibration parameter matrix, and realize pixel-level registration of linear laser point cloud and linear near-infrared image; Step 2: Construct a modular network model for environments with camouflage, dust, and smoke interference, establish a fusion strategy between linear array laser point cloud and linear array near-infrared image, and build a target recognition network based on deep learning; Therefore, laser point cloud and near-infrared image fusion modules are used at both the data level and feature level, respectively, and a deep learning-based target recognition network is established. This target recognition network includes a voxel fusion network, a feature extraction backbone network, a BEV mapping module, a BEV feature fusion module, and a 3D detection head. The construction process includes: (1) Voxel fusion network: The laser point cloud and near-infrared image are input into the voxel fusion network, the laser point cloud is converted into a dynamic voxelized mesh, and the geometric features of the voxels, the geometric offset between the voxels and the mesh and the feature offset are obtained; the near-infrared image is preprocessed to obtain pixel-level features, and the laser point cloud is mapped to the near-infrared image point by point at the input level through the spatial transformation matrix and calibration parameter matrix in step one and interpolation is used to obtain pixel-level features, which are concatenated with the geometric offset and feature offset to expand the voxel feature channels and output the voxel vector to the feature extraction backbone network; (2) Feature extraction backbone network: A dual-branch backbone network is adopted. The laser branch adopts the SECOND+FPN backbone architecture. The voxel vector output by the voxel fusion network is input into the laser branch to obtain multi-level 3D features. The near-infrared branch adopts the cascaded SwinTransformer+FPN backbone architecture to extract multi-level image features. The laser point cloud branch and the near-infrared image branch are executed independently to obtain multi-level feature maps of the point cloud stream and the image stream, and output them to the BEV mapping module. (3) BEV mapping module; a dual-branch mapping network is adopted. The laser branch converts the multi-level 3D features extracted by the feature extraction backbone network into the BEV space through the convolutional layer. The near-infrared branch uses the LSS module to perform 3D depth estimation on the multi-level image features output by the feature extraction backbone network and projects them into the BEV space. The convolutional layer unifies the size of the image feature map with the size of the point cloud feature map, matches the feature dimensions, and outputs the dual-branch BEV features to the BEV feature fusion module. (4) BEV feature fusion module: The dual-branch BEV features in the BEV mapping module are adaptively output to the near-infrared image feature weights through multi-layer convolutional layers and residual connection structures, and the laser point cloud BEV features and near-infrared image BEV features are fused in the target area to deepen the feature expression of the target area, obtain the spatial depth fusion features of BEV, and output to the 3D detection head. Step 3: Train the target recognition network constructed in Step 2, establish network performance evaluation indicators, evaluate the training results, and obtain the target recognition model.

2. The target identification method for a linear array laser and linear array near-infrared composite imaging fuze as described in claim 1, characterized in that: It also includes step four, which involves deploying the target recognition model from step three onto a linear laser and linear near-infrared composite imaging fuze to identify tank, truck, and car targets using the acquired laser point cloud and near-infrared images. This improves the recognition accuracy in camouflaged, dusty, and smoke-filled environments, enabling precise detonation control.

3. The target identification method for a linear array laser and linear array near-infrared composite imaging fuze as described in claim 1, characterized in that: The implementation method for step one is as follows: Based on the simulation data samples of linear array laser and linear array near-infrared imaging of tanks, trucks and cars, and following the processing flow of the mmdetection3d target detection framework library compatible with KITTI style, the overall file system structure and path, data types, label formats and parameter settings are defined to generate a dataset for training the target recognition network. Based on the simulation parameters, the spatial transformation matrix is ​​determined by the geometric positional relationship between the laser origin and the near-infrared pixel optical center. The spatial transformation matrix is ​​a combination of a three-dimensional rotation matrix and a translation vector, as shown in equation (1); where γ, ψ, and ε are the rotation angles of the pixel optical center relative to the laser origin in the x, y, and z directions, respectively, and t x t y t z These are the translation vectors in the x, y, and z directions, respectively; (1) The calibration parameter matrix is ​​determined according to the linear array imaging system to achieve pixel-level registration between the linear array laser point cloud and the linear array near-infrared image, as shown in Equation (2); where F is the focal length of the near-infrared detector, h is the detection height, s is the pixel length, and f is the focal length of the near-infrared detector. c θ is the camera frame rate, m is the number of pixels in the image width direction, n is the number of pixels in the image height direction, θ is the attack angle, and v is the image height direction. r It is the relative velocity; (2)。 4. The target identification method for a linear array laser and linear array near-infrared composite imaging fuze as described in claim 1, characterized in that: Building a deep learning-based target recognition network also includes: (5) 3D detection head; a series of multi-size 3D boxes generated in BEV space are used. During training, positive and negative samples are obtained by calculating the 3D intersection-union ratio of the generated boxes and the real boxes. The BEV fusion features in the BEV feature fusion module output the predicted value through a linear layer, which includes the detection box, classification and score. The loss function is iterated and the loss calculation is divided into four parts: classification loss, regression loss of the detection box, orientation loss of the detection box and image depth estimation loss. During the prediction process, the detection threshold is set to filter the predicted score. The multi-size 3D boxes are suppressed by non-maximum to obtain the final recognition result.

5. The target identification method for a linear array laser and linear array near-infrared composite imaging fuze as described in claim 1, characterized in that: The implementation method for step three is as follows: The dataset generated in step one is divided into a training set and a test set. The target recognition network constructed in step two is trained on the training set. Training is completed when the total loss value is less than α, where α ∈ (0, 0.15). The 3D target detection performance evaluation metric mAP, or mean Average Precision, is used. For three types of targets—tanks, trucks, and cars—under three interference conditions—camouflage, dust, and smoke—the accuracy values ​​under the β coverage threshold are calculated using three metric types: 3D projection onto 2D detection boxes, detection boxes from the BEV perspective, and 3D detection boxes, where β ∈ (0, 1). The test set is used for testing, and the model's precision and recall are statistically obtained. Based on this, the overall average precision is calculated to evaluate the target recognition network performance. An overall average precision greater than μ proves the model is effective, where μ ∈ (80%, 100%). Models not greater than μ are discarded, and hyperparameters are readjusted for retraining. The effective model is exported in a standard format for easy subsequent engineering applications.

Images