Water surface and ice surface amphibious unmanned ship target detection system under ice and snow environment

Abstract

The invention discloses a water surface and ice surface amphibious unmanned ship target detection system in an ice and snow environment, and relates to the field of intelligent unmanned ships. In order to solve the technical problem that in the prior art, due to the fact that an existing feature fusion method usually cannot effectively integrate high-level semantic information and low-level spatial information, the detection performance of a small target is insufficient, the technical scheme provided by the invention is that a water surface and ice surface amphibious unmanned ship target detection system under an ice and snow environment is provided. Comprising a module for adjusting a target detection threshold value to obtain multi-modal input data; a module for obtaining global features according to the multi-modal input data; a module for fusing global features; a test target detection module; if the detection result of the target for testing does not meet the preset condition, the module of the threshold value for target detection is readjusted; and the module is used for detecting the unmanned ship target after the detection result of the test target meets a preset condition. The method is suitable for target detection of the amphibious unmanned ship.

Description

Technical Field

[0001] It involves the field of intelligent unmanned smart ships, and specifically involves target detection of amphibious unmanned boats on water and ice in icy and snowy environments. Background Art

[0002] Performing unmanned boat missions in ice and snow environments is a very challenging technology. In recent years, neural networks and deep learning technologies have been widely used in the field of unmanned boat target detection and environmental perception, especially in target detection tasks in complex natural environments. For example, the YOLO (You Only Look Once) series of target detection algorithms have been widely used in fields such as autonomous driving and video surveillance due to their efficient single-stage detection capabilities, and have gradually been introduced into the visual detection of unmanned boats. In these applications, YOLO and its variant algorithms can quickly process image data, detect and identify targets in real time, and thus provide accurate environmental perception without human operation. However, the ice and snow environment poses new challenges to traditional target detection technology due to the extreme lighting conditions and the coexistence of complex water and ice surfaces.

[0003] In an ice and snow environment, the complex reflections of the water and ice surfaces as well as the ice and snow cover greatly reduce the detection accuracy, especially for the detection of small targets. For example, when the water surface produces strong reflections or the ice surface is covered with snow, the traditional detection method based on convolutional neural network (CNN) is prone to false detection or missed detection. In addition, the frequent changes in lighting conditions make the image data captured by the sensor easy to distort, further reducing the stability and reliability of the detection. Therefore, how to effectively deal with these environmental characteristics has become a bottleneck problem for existing target detection technology in polar unmanned boat operations.

[0004] At present, the research on ice and snow environments mainly focuses on the following aspects: (1) improving image preprocessing algorithms to reduce the impact of illumination changes and ice and snow reflections on detection; (2) using multimodal sensor fusion, such as combining optical photography and infrared imaging, to improve detection robustness by fusing data from different sources; (3) improving the capabilities of feature extraction networks, and continuously optimizing the backbone network and feature fusion modules in existing networks such as YOLOv5 and YOLOv7 to improve detection performance in complex backgrounds. However, although these improvements have enhanced the detection effect to a certain extent, there are still many limitations for the special water and ice conditions in ice and snow environments.

[0005] For example, traditional feature fusion methods are usually unable to effectively integrate high-level semantic information and low-level spatial information, resulting in insufficient detection performance for small targets. In addition, the adaptive ability of existing technologies in dealing with dynamic changes in ice and snow environments (such as light, temperature, etc.) is relatively poor, making it difficult to ensure the consistency and efficiency of the detection system throughout the unmanned boat mission. Therefore, the problem of missed detection of small targets in ice and snow environments, the balance between computational cost and detection accuracy, and the problem of model robustness under different environmental conditions are still key bottlenecks that the technical community needs to solve. Summary of the invention

[0006] In order to solve the technical problems existing in the prior art, the existing feature fusion methods are usually unable to effectively integrate high-level semantic information and low-level spatial information, resulting in insufficient detection performance for small targets, and it is difficult to ensure the consistency and efficiency of the detection system in the entire unmanned boat mission, the technical solution provided by the present invention is:

[0007] A method for detecting targets of amphibious unmanned boats on water and ice surfaces in an ice and snow environment, comprising:

[0008] The step of adjusting the target detection threshold according to the collected environmental data and obtaining multimodal input data;

[0009] The step of processing the multimodal input data respectively by a convolutional neural network and a Transformer structure to obtain global features;

[0010] A step of fusing high-level semantic information and low-level spatial information of the global feature;

[0011] Steps for performing target detection for testing;

[0012] If the test target detection result does not meet the preset condition, the step of re-adjusting the threshold value for target detection;

[0013] After the test target detection result meets the preset conditions, the step of detecting the unmanned boat target is carried out.

[0014] Furthermore, a preferred embodiment is provided, wherein the environmental data includes light, temperature and humidity data.

[0015] Furthermore, a preferred implementation is provided, which fuses the high-level semantic information and the low-level spatial information of the global features through shallow auxiliary fusion and high-level auxiliary fusion methods.

[0016] Furthermore, a preferred implementation is provided for performing test target detection through a YOLOv8 network.

[0017] Furthermore, a preferred implementation is provided to adjust the threshold of target detection based on an adaptive threshold adjustment mechanism, based on real-time data of light intensity, temperature and humidity, and by calculating environmental noise levels and signal characteristics through statistical analysis.

[0018] Furthermore, a preferred implementation is provided, in which the Transformer structure extracts global features by introducing a multi-head self-attention mechanism.

[0019] Based on the same inventive concept, the present invention also provides a target detection system for amphibious unmanned boats on water and ice in an ice and snow environment, comprising:

[0020] According to the collected environmental data, the threshold of target detection is adjusted and a module for multimodal input data is obtained;

[0021] A module for obtaining global features by processing the multimodal input data respectively through a convolutional neural network and a Transformer structure;

[0022] A module for fusing high-level semantic information and low-level spatial information of the global features;

[0023] Module for target detection for testing;

[0024] If the test target detection result does not meet the preset conditions, a module for re-adjusting the threshold value of the target detection;

[0025] A module that detects unmanned boat targets after the test target detection results meet the preset conditions.

[0026] Based on the same inventive concept, the present invention also provides a computer storage medium for storing a computer program. When the computer program is read by a computer, the computer executes the method described.

[0027] Based on the same inventive concept, the present invention also provides a computer, including a processor and a storage medium. When the processor reads a computer program stored in the storage medium, the computer executes the method described.

[0028] Based on the same inventive concept, the present invention also provides a computer program product, which is a computer program. When the computer program is executed, the method described above is implemented.

[0029] Compared with the prior art, the technical solution provided by the present invention is beneficial in that:

[0030] The CSP_PTB structure is used to improve the neural network backbone, combining the advantages of convolutional neural network (CNN) and Transformer, significantly improving the ability of feature extraction. In ice and snow environments, small target features are easily masked due to the reflection of ice and water surfaces, and it is difficult for traditional pure CNN structures to effectively extract global features. CSP_PTB allocates some channels to Transformer for processing, making global feature extraction more efficient and reducing computational costs, thereby ensuring that the feature information of small targets can be stably captured under extreme environmental conditions. Compared with the existing CNN or Transformer single structures, this hybrid structure improves the recognition rate of targets in diverse environments.

[0031] The introduction of the BIMAFPN module effectively integrates high-level semantic information and low-level spatial information, solving the bottleneck problem of information fusion in existing target detection networks. Traditional feature fusion modules are difficult to take into account the effective integration of information at different levels, resulting in poor detection performance, especially in small target scenarios. Through the secondary innovation of BIMAFPN, the effective integration of information at different levels is achieved by combining the shallow auxiliary fusion (SAF) module and the advanced auxiliary fusion (AAF) module. The retention of shallow information enhances the detection capability of local small targets, while the integration of high-level semantic information improves the understanding of complex backgrounds. Compared with the common FPN and BIFPN structures, the detection accuracy has been significantly improved.

[0032] The adaptive threshold adjustment mechanism enables the system to respond to the dynamic changes of the ice and snow environment in real time, enhancing the robustness of the detection system. In the ice and snow environment, due to the frequent changes in factors such as light intensity, temperature, and ice and snow coverage, the traditional fixed threshold detection method cannot cope with these complex environmental conditions, often resulting in increased false detection and missed detection rates. By introducing a dynamic adjustment mechanism for the adaptive threshold, the present invention can optimize the detection parameters in real time according to environmental data, thereby effectively reducing false detection and missed detection. Compared with the existing fixed parameter detection method, this mechanism significantly improves the sensitivity and reliability of the detection system under various extreme conditions.

[0033] The fusion of multimodal sensors further enhances the detection performance of targets. The system uses a combination of high-resolution optical cameras, infrared thermal imagers, and lidar to capture diverse features in different environments. Compared with a single sensor system, this solution can obtain information from multiple dimensions, compensating for the defects of single visual information under extreme conditions. For example, in low light conditions, infrared thermal imaging can effectively detect subtle changes in the ice surface, while lidar can accurately measure the distance and contour of objects. This fusion of multimodal information greatly improves the accuracy of target detection in complex environments, making the system more adaptable to various environmental conditions.

[0034] By introducing the generative adversarial network (GAN) for data enhancement, the model's adaptability to environmental changes is improved. In ice and snow environments, environmental conditions such as lighting and ice surface conditions are constantly changing, and conventional models find it difficult to maintain high performance when data distribution changes. The use of generative adversarial networks to simulate diverse environmental changes, such as lighting changes and ice cracks, greatly enriches the training data, enabling the model to better cope with various extreme scenarios. Compared with models that do not use data enhancement, this method significantly enhances the stability and robustness of the target detection system in different extreme environments, ensuring the reliability of task execution.

[0035] It is suitable for use in amphibious unmanned boat target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic diagram of a method for detecting amphibious unmanned boat targets on water and ice surfaces in an ice and snow environment;

[0037] Figure 2 It is a schematic diagram of the structure of YOLOv8;

[0038] Figure 3 It is a schematic diagram of the structure of CSP_PTB;

[0039] Figure 4 Schematic diagram of the structure of BIMAFPN. DETAILED DESCRIPTION

[0040] In order to make the advantages and benefits of the technical solution provided by the present invention more clearly reflected, the technical solution provided by the present invention is now further described in detail in conjunction with the accompanying drawings, specifically:

[0041] Embodiment 1: This embodiment provides a method for detecting targets on water and ice amphibious unmanned boats in an ice and snow environment, including:

[0042] The step of adjusting the target detection threshold according to the collected environmental data and obtaining multimodal input data;

[0043] The step of processing the multimodal input data respectively by a convolutional neural network and a Transformer structure to obtain global features;

[0044] A step of fusing high-level semantic information and low-level spatial information of the global feature;

[0045] Steps for performing target detection for testing;

[0046] If the test target detection result does not meet the preset condition, the step of re-adjusting the threshold value for target detection;

[0047] After the test target detection result meets the preset conditions, the step of detecting the unmanned boat target is carried out.

[0048] Specifically,

[0049] The technical solution provided in this embodiment designs a CSP_PTB structure, combines convolutional neural networks and Transformer to improve the efficiency and accuracy of feature extraction, and proposes a BIMAFPN module to replace the original feature fusion part, thereby achieving efficient multi-layer feature fusion. In addition, an adaptive threshold adjustment mechanism is adopted, which can automatically adjust the detection sensitivity according to real-time environmental conditions to ensure detection performance under extreme environmental conditions. These improvements effectively cope with the environmental conditions of strong reflection of polar water and ice surfaces, frequent ice and snow coverage, and significant changes in light and temperature, and significantly improve the accuracy and efficiency of target detection.

[0050] Step 1: Environmental data acquisition and adaptive threshold adjustment

[0051] According to the environmental data obtained in real time, the threshold of target detection is dynamically adjusted to ensure the best target detection effect in different ice and snow environments. In this step, the system obtains environmental data of polar water and ice surfaces in real time through multiple sensors (high-resolution optical cameras, infrared thermal imagers, lidar), including information such as light intensity, temperature, humidity and ice surface status. Based on these real-time data, the system performs environmental assessment and calculates statistical indicators (such as mean and variance) to determine the noise level and signal characteristics in the current environment. Through the adaptive threshold adjustment mechanism, the system can adjust the sensitivity of target detection according to the dynamic changes of external conditions. For example, when the light intensity decreases, the system automatically lowers the detection threshold to improve the recognition rate of small targets. The adjusted threshold will be used as the input parameter for the next step of target detection to cope with sudden changes in environmental conditions and ensure the sensitivity and reliability of detection.

[0052] Step 2: Multimodal sensor data fusion and feature extraction

[0053] Combined with the data from multimodal sensors, the improved neural network backbone is used to extract environmental features. CSP_PTB (Cross-Phase Partial Transformer Block) is a hybrid structure that combines convolutional neural network (CNN) and Transformer. First, the input feature map is divided into two parts, one of which is processed by CNN to extract local spatial features, and the other is processed by Transformer to extract global semantic information. By partially allocating the input channels, the system effectively enhances the feature extraction capability without significantly increasing the computational complexity. The multimodal features processed by CSP_PTB will be used as the input for subsequent feature fusion to improve the accuracy and robustness of target detection.

[0054] Step 3: Innovative application of multi-layer feature fusion and BIM AFPN module

[0055] The BIMAFPN module is used to fuse multi-layer features to improve the system's target detection capabilities at different levels. The BIMAFPN module is a secondary innovation based on BIFPN (Bidirectional Feature Pyramid Network), adding a shallow auxiliary fusion (SAF) module and an advanced auxiliary fusion (AAF) module. The SAF module is used to fuse the information output by the neural network trunk and neck to ensure that the optimal level of shallow features is retained during the fusion process to improve the detection capability of small targets. The AAF module further incorporates more diverse gradient information in the output layer to enhance the expression capability of deep semantics. After the multi-layer feature fusion of the BIMAFPN module, high-level semantic information and low-level spatial information can be effectively integrated, thereby enhancing the system's adaptability and detection accuracy to the changeable ice and snow environment. The fused features will be passed to the detection layer for final target recognition.

[0056] Step 4: Target detection and model optimization based on YOLOv8 network

[0057] Target detection is performed based on the YOLOv8 framework, and the detection results are used to determine whether repeated optimization is needed. After the multi-layer feature fusion is completed, the fused features are input into the YOLOv8 target detection framework. YOLOv8 is a single-stage detection model adopted by the present invention. The feature map in CSPDarknet53 is fused with the feature map in 2-Stage FPN through the C2f module to further improve the detection performance. The system first performs target detection based on the fused features and calculates the classification loss and regression loss of each type of target. To ensure the detection effect, the system sets the threshold of detection accuracy to 0.5. If the detection accuracy is lower than the threshold, the system will automatically adjust the adaptive threshold and detection parameters in the previous steps, and repeat the above process to optimize the detection results. This iteration continues until the detection result reaches the predetermined accuracy requirement, thereby ensuring the efficiency and accuracy of target detection.

[0058] Step 5: Model training and data enhancement

[0059] Generative adversarial networks (GANs) are used for data enhancement to improve the model's adaptability to environmental changes. In order to enhance the model's adaptability to ice and snow environments, generative adversarial networks (GANs) are used to enhance the training data during the model training phase. GANs are used to simulate changes in different environmental conditions, such as changes in light intensity and cracks in the ice, to generate diverse training data, thereby enriching the model's training set. The purpose of this is to improve the model's robustness to extreme environmental conditions, so that it can maintain efficient detection performance even when light, temperature, humidity, and other conditions change significantly. The model after data enhancement shows higher detection accuracy and stability in different ice and snow environments, ensuring that target detection tasks can be effectively completed in a variety of scenarios.

[0060] Implementation method 2: This implementation method further limits the method for detecting amphibious unmanned boat targets on water and ice surfaces in an ice and snow environment provided in implementation method 1, and the environmental data includes light, temperature and humidity data.

[0061] Implementation method three: This implementation method further limits the method for detecting amphibious unmanned boat targets on water and ice surfaces in an ice and snow environment provided in implementation method one. The high-level semantic information and low-level spatial information of the global features are fused through shallow auxiliary fusion and advanced auxiliary fusion methods.

[0062] Implementation method 4: This implementation method further limits the method for detecting amphibious unmanned boat targets on water and ice surfaces in an ice and snow environment provided in implementation method 1, and performs test target detection through the YOLOv8 network.

[0063] Implementation method five. This implementation method further limits the method for detecting amphibious unmanned boat targets on water and ice surfaces in an ice and snow environment provided in implementation method one. The threshold for target detection is adjusted based on an adaptive threshold adjustment mechanism, based on real-time data of light intensity, temperature and humidity, and is achieved by statistically analyzing and calculating environmental noise levels and signal characteristics.

[0064] Implementation method six: This implementation method further limits the method for detecting amphibious unmanned boat targets on water and ice surfaces in an ice and snow environment provided in implementation method one. The Transformer structure extracts global features by introducing a multi-head self-attention mechanism.

[0065] Embodiment 7: This embodiment provides a target detection system for amphibious unmanned boats on water and ice in an ice and snow environment, including:

[0066] According to the collected environmental data, the threshold of target detection is adjusted and a module for multimodal input data is obtained;

[0067] A module for obtaining global features by processing the multimodal input data respectively through a convolutional neural network and a Transformer structure;

[0068] A module for fusing high-level semantic information and low-level spatial information of the global features;

[0069] Module for target detection for testing;

[0070] If the test target detection result does not meet the preset conditions, a module for re-adjusting the threshold value of the target detection;

[0071] A module that detects unmanned boat targets after the test target detection results meet the preset conditions.

[0072] Embodiment 8: This embodiment provides a computer storage medium for storing a computer program. When the computer program is read by a computer, the computer executes the method provided in embodiment 1.

[0073] Embodiment 9: This embodiment provides a computer, including a processor and a storage medium. When the processor reads a computer program stored in the storage medium, the computer executes the method provided in embodiment 1.

[0074] Embodiment 10: This embodiment provides a computer program product, which is a computer program. When the computer program is executed, the method provided in embodiment 1 is implemented.

[0075] Implementation Method XI: Combination Figure 1-4This embodiment is described in detail by using specific examples to further describe the technical solution provided above. Specifically:

[0076] A method for detecting amphibious unmanned boats on water and ice in an ice and snow environment is based on the YOLOv8 network framework. By improving the neural network backbone and network framework, it can accurately detect targets on water and ice in an ice and snow environment. The sensitivity and accuracy of target detection are automatically optimized according to changes in environmental conditions, improving the accuracy of amphibious unmanned boat target detection on water and ice in polar scenes.

[0077] The purpose of this implementation is to provide a target detection system for amphibious unmanned boats on water and ice in an ice and snow environment. The C2f module is used in YOLOv8. The function of the C2f module is to connect and fuse the feature map in CSPDarknet53 with the feature map in 2-Stage FPN. The improved CSP_PTB hybrid structure divides the input feature map into two parts, which are processed by CNN and Transformer respectively. The computing efficiency and feature extraction capability are optimized by partially allocating the input channels. The advantages of convolutional neural network (CNN) and Transformer are combined to reduce the computing cost while ensuring efficient feature extraction. This implementation is based on the coexistence of polar water and ice surfaces and the complex environment. In order to meet the needs of polar unmanned boat operations, based on the YOLOv8 detection model, CSP_PTB is used to improve the neural network backbone, and BIMAFPN is used to replace the neck part of the original framework to improve the network framework, so as to improve the detection accuracy and efficiency of small targets in polar water and ice environments, and ensure efficient detection in complex environments. The specific implementation method is as follows.

[0078] 1. Multi-layer feature extraction: Use sensors to collect image data under different environmental conditions and extract corresponding features (such as color, shape, temperature, etc.). Fuse data from different sensors and improve overall detection performance through multimodal learning technology. Use techniques such as generative adversarial networks (GAN) to enhance training data and simulate different environmental changes (such as lighting changes, ice cracks, etc.) to further improve the robustness of the model. Combined with deep learning algorithms, the sensor data is processed through convolutional neural networks (CNNs) to extract target features from multiple levels and improve the detection rate of small targets.

[0079] 2. Adaptive threshold setting: Adaptive threshold setting is a method of dynamically adjusting the detection threshold to automatically optimize the sensitivity and accuracy of target detection according to changes in environmental conditions. Especially in ice and snow environments, changes in light, temperature and weather conditions have a significant impact on sensor performance. The sensor monitors the environmental conditions in real time, including temperature, humidity, light intensity and ice surface conditions. The collected data is used to evaluate the current detection conditions. Statistical analysis is performed using historical data and real-time data to calculate the mean and variance under the current environmental conditions. Based on these statistical information, typical noise levels and signal characteristics are identified. During the target detection process, the system monitors changes in the external environment in real time, identifies potential interference factors (such as snowflakes, wind and waves, etc.), and adjusts the detection strategy to dynamically adjust the threshold. When the environmental conditions suddenly change, the system can respond quickly and automatically switch to an adaptive detection mode. For example, if the ambient light intensity is low, the detection threshold is lowered to improve the recognition rate of small targets. The system continuously evaluates the detection results during the detection process and further adjusts the threshold based on the feedback of the detection success rate. If the target detection rate is low, the system will automatically lower the threshold, otherwise it will increase the threshold to maintain the detection effect.

[0080] Through adaptive threshold setting and environmental adaptation mechanism, this system can effectively cope with changing conditions in ice and snow environments, improve the detection accuracy and efficiency of small targets, and ensure good performance in complex ice and water environments. This method not only improves the flexibility of target detection, but also provides reliable data support for subsequent applications.

[0081] 3. Such as Figure 3 As shown in the flowchart, the input feature map is divided into two parts, which are processed by CNN and Transformer respectively. The Transformer structure has a powerful global feature extraction capability in computer vision tasks. However, due to the high computational complexity of the Transformer structure, directly applying it to all channels will result in significant computational overhead. In order to reduce the computational cost while ensuring efficient feature extraction, a hybrid structure is designed to divide the input feature map into two parts, which are processed by CNN and Transformer respectively. The module combines the convolutional neural network (CNN) and the Transformer mechanism to enhance the feature extraction capability. A module named CSP_PTB (Cross Stage Partial-PartiallyTransformer Block) is proposed, which aims to combine the advantages of CNN and Transformer to optimize the computational efficiency and feature extraction capability by partially allocating the input channels.

[0082] Many studies have shown that CNN has a small receptive field, which means it can only extract local features, but Transformer's MHSA can extract global features and take advantage of both. In order to introduce the Transformer structure to extract global features without significantly increasing the computational complexity, the Partially Transformer Block is proposed, which only uses TransformerBlock for some channels. MHSA_CGLU contains Mutil-Head-Self-Attention and ConvolutionalGLU, where Mutil-Head-Self-Attention is responsible for extracting global features and ConvolutionalGLU is used to enhance the expression of nonlinear features. Compared with traditional FFN, ConvolutionalGLU has stronger performance. The number of channels used for Transformer can be adjusted according to different model sizes and specific operating conditions.

[0083] 4. Such as Figure 4 As shown in the structure diagram, in order to effectively and adaptively integrate high-level semantic information and low-level spatial information at the same time, the idea of ​​BIFPN is used to make a secondary innovation on MAFPN to get BIMAFPN. The shallow auxiliary fusion (SAF) module combines the outputs of the backbone and the neck to retain the optimal level of shallow information to facilitate subsequent learning. At the same time, the advanced auxiliary fusion (AAF) module embedded in the neck transmits more diverse gradient information to the output layer.

[0084] 5. Model training and detection: Use transfer learning or incremental learning methods, use newly collected data as training sets, and regularly update the target detection model. By continuously optimizing model parameters, it can maintain high detection accuracy under different environmental conditions. Finally, perform detection through Detect, calculate classification loss and regression loss, and set the detection accuracy of each category of the detection result as 0.5 as the threshold to determine whether the task is completed. If it is less than the threshold, repeat the above steps to finally complete the target detection.

[0085] Figure 2-4 In the article, deep learning terminology is explained as follows:

[0086] 1.conv (Convolution): refers to the convolution layer, which is used to extract features from images in convolutional neural networks.

[0087] 2.C2f: refers to the feature fusion module, which is usually used to fuse features from multiple convolutional layers between different network layers to improve the model's detection ability for multi-scale features.

[0088] 3.CSP_PTB (Cross Stage Partial-Partially Transformer Block): It is a module that combines convolutional neural network (CNN) and Transformer to enhance the efficiency of feature extraction. By using Transformer in some channels, it can take into account the extraction of local and global features, and optimize the computational efficiency and feature expression ability.

[0089] 4. BIMAFPN (Bi-directional Improved Multi-scale Feature Pyramid Network): It is an improved module based on BIFPN (Bidirectional Feature Pyramid Network) and is used to efficiently fuse features from different levels, especially high-level semantics and low-level spatial information. Its main function is to improve the accuracy and robustness of the model when detecting small objects.

[0090] 5. MHSA (Multi-Head Self-Attention): Multi-head self-attention mechanism is a commonly used technology in Transformer, which is used to extract global features and enhance the model's understanding of complex backgrounds.

[0091] 6.Convolutional GLU (Convolutional Gated Linear Unit): It is an improved activation function structure used to enhance the expression ability of nonlinear features. Compared with the traditional fully connected layer structure, it has better feature expression performance.

[0092] 7.AAF (Advanced Auxiliary Fusion): Advanced auxiliary fusion module, which is used to integrate more diverse gradient information into the output layer of the deep learning model to enhance the expression ability of deep features, especially to effectively meet the needs of small target detection in complex ice and snow environments.

[0093] 8.SAF (Shallow Auxiliary Fusion): Shallow auxiliary fusion module, which is used to combine the trunk and neck outputs of the neural network to retain shallow feature information, thereby improving the detection capability of small targets and ensuring that shallow features are not lost during the fusion process.

[0094] 9. Detect: Usually refers to the target detection module. In single-stage detection models such as YOLO, the Detect module is used to calculate the final classification loss and regression loss to complete the tasks of target positioning and category judgment.

[0095] 10. SPPF (Spatial Pyramid Pooling-Fast): A spatial pyramid pooling method that is used to reduce the size of feature maps while maintaining feature spatial information, helping the model to better extract multi-scale feature information. SPPF is designed to enhance the detection performance of objects of different scales without significantly increasing computational complexity, and is particularly suitable for scenarios where the size of the object varies greatly.

[0096] 11. norm (Normalization): refers to the normalization operation. Commonly used normalization methods in deep learning include batch normalization (BN) and layer normalization (LN). The main purpose of normalization is to accelerate the convergence speed of the model and reduce the problem of gradient disappearance or explosion.

[0097] 12. Transformer-ch (Transformer Channel): refers to the channel part of the model that processes features when using Transformer. In some improved network structures, the feature channel will be partially or completely processed by Transformer to extract global context information, thereby enhancing the recognition ability of complex targets.

[0098] 13.MHSA-CGLU (Multi-Head Self-Attention with Convolutional Gated Linear Unit): A module that combines the multi-head self-attention mechanism (MHSA) and the convolutional gated linear unit (Convolutional GLU). MHSA-CGLU extracts global features through the self-attention mechanism and uses the convolutional gated unit to enhance the nonlinear expression ability of the features to balance the global information extraction and computational complexity.

[0099] 14. Bottleneck: A neural network module used to reduce feature dimensions without losing important information. It is usually composed of "1x1 convolution-3x3 convolution-1x1 convolution" to reduce the amount of calculation and retain the representation ability of the model.

[0100] 15.Droppath: Droppath is a regularization method similar to Dropout, but it randomly discards the entire neural network path instead of a single neuron. Droppath is used to increase the generalization ability of the model and reduce overfitting.

[0101] 16. Add: The Add operation is a common tensor operation, which is usually used for feature fusion between network layers. For example, in the residual network, it is used to add the output of the backbone network to the residual connection to help the information propagate effectively.

[0102] 17. Concat (Concatenation): Concat is an operation that connects two or more tensors in a specific dimension, which is often used for feature fusion. Concat can enhance the representation ability of the model.

[0103] 18.BatchNorm2d (2D Batch Normalization): BatchNorm2d is a 2D version of batch normalization, mainly used in convolutional layers. It normalizes the mean and variance of each small batch of data to alleviate the problems of gradient vanishing and gradient exploding.

[0104] 19. SiLU (Sigmoid Linear Unit, also called Swish): An activation function, defined as `f(x) = x*sigmoid(x)`, which combines the effects of linear and nonlinearity to improve the learning ability and nonlinear expression ability of the model.

[0105] 20.MaxPool2d (two-dimensional maximum pooling): MaxPool2d is a two-dimensional version of the maximum pooling operation, which is mainly used to reduce the size of feature maps while retaining important features. It is often used for downsampling in convolutional neural networks.

[0106] 21.UpSample: UpSample is an upsampling operation used to increase the size of the feature map and is often used to generate an output with the same spatial resolution as the original input.

[0107] 22.Split: Split is an operation that splits a tensor along a specified dimension, usually used in multi-input, parallel processing or multi-branch structures.

[0108] 23.Bbox.Loss (Bounding Box Loss): used to measure the gap between the bounding box predicted by the model and the actual target bounding box. Common methods include L1 loss, Smooth L1 loss, IoU loss, etc.

[0109] 24.Cls.Loss (Classification Loss): used to measure the gap between the target category predicted by the model and the true category. Commonly used classification losses include Cross-Entropy Loss and FocalLoss.

[0110] The technical solution provided by the present invention is further described in detail above through several specific implementation modes in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the several specific implementation modes described above are not intended to be used as limitations on the present invention. Any reasonable modification and improvement of the present invention, combination of implementation modes and equivalent substitution within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting amphibious unmanned boat targets on water and ice surfaces in an ice and snow environment, characterized in that: include: The step of adjusting the target detection threshold according to the collected environmental data and obtaining multimodal input data; The step of processing the multimodal input data respectively by a convolutional neural network and a Transformer structure to obtain global features; A step of fusing high-level semantic information and low-level spatial information of the global feature; Steps for performing target detection for testing; If the test target detection result does not meet the preset condition, the step of re-adjusting the threshold value for target detection; After the test target detection result meets the preset conditions, the step of detecting the unmanned boat target is carried out.

2. The method for detecting amphibious unmanned boat targets on water and ice surfaces in an ice and snow environment according to claim 1 is characterized in that: The environmental data includes light, temperature and humidity data.

3. The method for detecting amphibious unmanned boat targets on water and ice surfaces in an ice and snow environment according to claim 1 is characterized in that: The high-level semantic information and the low-level spatial information of the global features are fused through shallow-level auxiliary fusion and high-level auxiliary fusion methods.

4. The method for detecting amphibious unmanned boat targets on water and ice surfaces in an ice and snow environment according to claim 1 is characterized in that: Through the YOLOv8 network, target detection is performed for testing.

5. The method for detecting amphibious unmanned boat targets on water and ice surfaces in an ice and snow environment according to claim 1 is characterized in that: The threshold for target detection is adjusted based on an adaptive threshold adjustment mechanism, which is achieved by calculating the ambient noise level and signal characteristics through statistical analysis based on real-time data of light intensity, temperature and humidity.

6. The method for detecting amphibious unmanned boat targets on water and ice surfaces in an ice and snow environment according to claim 1 is characterized in that: The Transformer structure extracts global features by introducing a multi-head self-attention mechanism.

7. A target detection system for amphibious unmanned boats on water and ice in an ice and snow environment, characterized in that: include: According to the collected environmental data, the threshold of target detection is adjusted and a module for multimodal input data is obtained; A module for obtaining global features by processing the multimodal input data respectively through a convolutional neural network and a Transformer structure; A module for fusing high-level semantic information and low-level spatial information of the global features; Module for target detection for testing; If the test target detection result does not meet the preset conditions, a module for re-adjusting the threshold value of the target detection; A module that detects unmanned boat targets after the test target detection results meet the preset conditions.

8. A computer storage medium for storing a computer program, characterized in that: When the computer program is read by a computer, the computer executes the method of claim 1 .

9. A computer, comprising a processor and a storage medium, characterized in that: When the processor reads the computer program stored in the storage medium, the computer executes the method of claim 1 .

10. A computer program product, being a computer program, characterized in that When the computer program is executed, the method of claim 1 is implemented.

Images