Three-dimensional target identification method based on multi-view target identification network

Abstract

The invention relates to the technical field of target visual recognition, in particular to a three-dimensional target recognition method based on a multi-view target recognition network, and the method specifically comprises the following steps: data collection, data preprocessing, feature extraction, feature fusion, target classification and positioning, model training and real-time target recognition. Through comprehensive utilization of multi-view data, limitation of single-view data is reduced, recognition precision is improved, through a multi-view CNN sharing weight and an attention mechanism, features of a target object can be better captured, feature weights are dynamically adjusted through the attention mechanism, adaptability of a model to a complex scene is enhanced, and recognition accuracy is improved. According to the method, high recognition precision can still be kept even under the conditions of illumination change, shielding and view angle change, the parameter quantity of the model is reduced through the network design of the shared weight, the calculation complexity is reduced, the method is suitable for real-time application, the model can converge in a short time through staged training and selection of an optimizer, and the training efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of target visual recognition, and in particular to a three-dimensional target recognition method based on a multi-viewpoint target recognition network. Background Art

[0002] With the rapid development of technologies such as autonomous driving, robot navigation, virtual reality (VR) and augmented reality (AR), the application of 3D object recognition in these fields has become increasingly important. Efficient and accurate 3D object recognition methods can not only improve the performance of these technologies, but also expand their application scenarios.

[0003] Traditional 3D object recognition methods mainly rely on single-view image data. These methods are easily affected by factors such as viewing angle, lighting conditions, and object occlusion, resulting in low recognition accuracy and insufficient robustness. In recent years, the acquisition and use of multi-view data has become an important means to improve recognition accuracy. However, traditional multi-view data processing technology has the following shortcomings:

[0004] Viewpoint dependence: Traditional methods are highly dependent on viewpoints, and changes in viewpoints can easily lead to recognition failures;

[0005] Data noise: Single-view data is easily affected by lighting, shadows, and occlusions, which may introduce noise;

[0006] Computational complexity: The processing of multi-viewpoint data requires a lot of computing resources and has high complexity and computational cost;

[0007] Therefore, how to efficiently fuse data from multiple perspectives and perform accurate three-dimensional target recognition has become an urgent problem to be solved. Summary of the invention

[0008] The purpose of the present invention is to provide a three-dimensional target recognition method based on a multi-view target recognition network, which can efficiently fuse data from multiple perspectives and perform accurate three-dimensional target recognition to solve the technical problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solutions.

[0010] The three-dimensional target recognition method based on the multi-view target recognition network specifically includes the following steps:

[0011] Data acquisition, multi-view image acquisition, and recording of camera parameters;

[0012] Data preprocessing: standardization, data enhancement, and perspective alignment of the collected image data;

[0013] Feature extraction: using convolutional neural network technology to extract feature maps of each view;

[0014] Feature fusion, using the attention mechanism to perform weighted fusion of features from different perspectives;

[0015] Target classification and positioning, using fully connected networks for classification and regression processing;

[0016] Model training, using labeled datasets to train and optimize deep learning;

[0017] Real-time target recognition, real-time image acquisition and classification and positioning output.

[0018] Preferably, the multi-view image acquisition specifically includes: using multiple cameras to shoot the target object from different angles to obtain two-dimensional image data of multiple viewpoints;

[0019] Recording camera parameters specifically includes: recording the camera parameters of each viewing angle, including camera position, direction, and internal parameters.

[0020] Preferably, at least three cameras need to be arranged in the multi-view image acquisition step, and each camera is arranged on the same plane, and the angle between the lens axes of two adjacent cameras is not less than 85°.

[0021] Preferably, in the step of recording camera parameters, the internal parameters of the camera are the focal length and principal point of the camera, etc.

[0022] Preferably, the standardization process specifically includes: standardizing the collected multi-viewpoint images, adjusting the image size to a uniform size, and performing denoising;

[0023] The data enhancement process is as follows: improving the visual quality of the image through contrast enhancement technology to make the features of the target object more obvious;

[0024] The perspective alignment process is specifically to use camera parameters to align images from different perspectives to ensure the consistency of subsequent feature extraction.

[0025] Preferably, the feature extraction step specifically includes the following steps:

[0026] Build a deep learning model and construct a convolutional neural network as a feature extractor to extract common features of images from different perspectives;

[0027] Design convolutional layers, design multi-layer convolutional layers and pooling layers to extract multi-scale features. The output of the convolutional layer is nonlinearly transformed through the ReLU activation function;

[0028] Generate feature maps. Images from each perspective generate feature maps through CNN. These feature maps contain the spatial structure and texture information of the target object.

[0029] Preferably, the feature fusion step specifically includes the following steps:

[0030] Design a feature fusion layer to integrate features from different perspectives through weighted fusion;

[0031] Design an attention mechanism and use a multi-head attention mechanism to adaptively weight feature maps from different perspectives, so that the model can pay more attention to important features;

[0032] The fusion result, the feature map of multi-view information, is used for subsequent target classification and positioning.

[0033] Preferably, target classification and positioning specifically include:

[0034] Classification layer: After feature fusion, the target is classified through the fully connected layer, and the softmax function is used as the output layer of the classifier to output the probability distribution of the target category;

[0035] Regression layer: While classifying, the three-dimensional position and posture of the target are determined through regression analysis. The output of the regression layer is the three-dimensional coordinates and rotation matrix of the target object;

[0036] Loss function: The cross entropy loss function is used for classification loss calculation, and the mean square error loss function is used for regression loss calculation, and the model performance is improved through joint optimization.

[0037] Preferably, the model training step specifically includes the following steps:

[0038] Dataset preparation: Use annotated 3D target datasets for model training. The datasets should contain images from different perspectives and under different lighting conditions.

[0039] Select a training strategy for training. Use a phased training strategy to first train the feature extraction network, then fix the feature extraction network parameters, and train the classification and regression networks.

[0040] Select the optimizer for optimization, use the Adam optimizer to optimize the model parameters, and adjust the learning rate to meet the needs of different training stages;

[0041] Training monitoring: monitor the model training process through the validation set to prevent overfitting and ensure the generalization ability of the model.

[0042] Preferably, real-time target recognition specifically includes:

[0043] Real-time data acquisition, in practical applications, multiple cameras are used to capture the image data of the target object in real time;

[0044] Real-time processing: using the trained multi-view target recognition network model to perform real-time image processing, extract features and perform fusion;

[0045] The recognition result output outputs the recognition category of the target object and its position and posture in three-dimensional space for subsequent application use.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0047] The three-dimensional target recognition method based on the multi-view target recognition network provided by the present invention reduces the limitations of single-view data and improves the recognition accuracy through the comprehensive utilization of multi-view data. Through the multi-view CNN and attention mechanism with shared weights, the model can better capture the characteristics of the target object. The feature weights are dynamically adjusted through the attention mechanism, which enhances the adaptability of the model to complex scenes. Even in the case of changes in illumination, occlusion and perspective, the model can still maintain a high recognition accuracy. The network design with shared weights reduces the number of model parameters and the computational complexity, making it suitable for real-time applications. In addition, through staged training and the selection of optimizers, the model can converge in a relatively short time, thereby improving the training efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of the steps of a three-dimensional target recognition method based on a multi-view target recognition network provided by the present invention;

[0049] Figure 2 A schematic diagram of the framework of a three-dimensional target recognition method based on a multi-view target recognition network provided by the present invention;

[0050] Figure 3 A schematic diagram of the camera scene arrangement position in the present invention;

[0051] Figure 4 It is a schematic diagram of the specific steps of feature extraction in the present invention;

[0052] Figure 5 It is a flowchart of the specific steps of feature fusion in the present invention;

[0053] Figure 6 The figure is a flowchart of the specific steps of model training in the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] See also Figure 1-Figure 6The present invention provides a three-dimensional target recognition method based on a multi-view target recognition network, which specifically includes the following steps:

[0056] Data acquisition, multi-view image acquisition, and recording of camera parameters;

[0057] Data preprocessing: standardization, data enhancement, and perspective alignment of the collected image data;

[0058] Feature extraction: using convolutional neural network technology to extract feature maps of each view;

[0059] Feature fusion, using the attention mechanism to perform weighted fusion of features from different perspectives;

[0060] Target classification and positioning, using fully connected networks for classification and regression processing;

[0061] Model training, using labeled datasets to train and optimize deep learning;

[0062] Real-time target recognition, real-time image acquisition and classification and positioning output.

[0063] The recognition method provided in the present application reduces the limitations of single-view data and improves recognition accuracy through the comprehensive utilization of multi-view data. Through the multi-view CNN and attention mechanism with shared weights, the model can better capture the features of the target object. The feature weights are dynamically adjusted through the attention mechanism, which enhances the model's adaptability to complex scenes. Even in the case of changes in lighting, occlusion, and perspective, the model can still maintain a high recognition accuracy. The network design with shared weights reduces the number of model parameters and the computational complexity, making it suitable for real-time applications. In addition, through staged training and the selection of optimizers, the model can converge in a shorter time, thereby improving training efficiency.

[0064] Multi-view image acquisition is specifically: using multiple cameras to shoot the target object from different angles to obtain multi-view two-dimensional image data.

[0065] Specifically, at least three cameras need to be arranged in the multi-view image acquisition step, and each camera is arranged on the same plane, and the angle between the lens axes of two adjacent cameras is not less than 85°. The schematic diagram of the camera scene arrangement position is as follows: Figure 3 As shown in the figure, the accuracy and robustness of 3D target recognition can be effectively improved. This arrangement can capture the multi-angle information of the target to the maximum extent, reduce the information loss caused by the perspective occlusion, and avoid the parallax between cameras being too small, so that the system can better extract the target features and enhance the learning effect of the deep learning model, thereby improving the target recognition performance and the ability to adapt to complex environments.

[0066] Recording camera parameters specifically means recording the camera parameters of each viewpoint, including the camera position, direction, and internal parameters. The camera internal parameters are specifically the focal length and principal point of the camera, etc., so as to facilitate subsequent 3D reconstruction and viewpoint alignment. The camera parameters can be accurately calibrated using a calibration plate or a known object, which can maximize the accuracy of 3D reconstruction and viewpoint alignment. Accurate camera parameters can ensure accurate alignment of image data, reduce errors during 3D reconstruction, thereby improving the accuracy and robustness of 3D target recognition, ensuring the consistency and reliability of the model at different viewpoints, and enhancing the integrity of the system.

[0067] The standardization process is as follows: standardize the collected multi-viewpoint images to 224x224 pixels, adjust the image size to a uniform size, and perform denoising. The denoising process can use Gaussian filtering, bilateral filtering and other methods, which can significantly improve image quality and data consistency. This process not only helps to reduce noise interference and retain image details, but also enables the deep learning model to process input data more efficiently, thereby improving the accuracy and stability of three-dimensional target recognition.

[0068] The data enhancement processing is specifically as follows: improving the visual quality of the image through contrast enhancement technology to make the features of the target object more obvious. Among them, the contrast enhancement technology can use histogram equalization technology or CLAHE technology, which significantly increases the contrast and detail clarity of the image, which not only helps to improve the recognizability of the target, but also enriches the diversity of training data and enhances the learning ability of the deep learning model, thereby improving the accuracy and reliability of three-dimensional target recognition.

[0069] The perspective alignment process is as follows: use camera parameters to align images from different perspectives to ensure the consistency of subsequent feature extraction. The alignment method can adopt feature point-based matching technology, such as SIFT, SURF, etc., or directly use camera parameters for geometric transformation, so that images from different perspectives are accurately aligned in space, thereby reducing errors in feature extraction, ensuring that feature information from different perspectives can be effectively integrated, and improving the accuracy and stability of overall recognition.

[0070] As the instruction manual Figure 4 As shown, the feature extraction step specifically includes the following steps:

[0071] Constructing a deep learning model: Constructing a convolutional neural network as a feature extractor to extract common features of images from each perspective. Specifically, the present invention adopts a multi-view CNN architecture based on shared weights, aiming to extract common features of images from each perspective. The shared weight design reduces the number of model parameters, reduces storage and computing costs, and improves training efficiency, making the model more efficient and stable when processing multi-view data, thereby improving the overall recognition performance and generalization ability;

[0072] Design convolutional layers: Design multiple convolutional layers and pooling layers to extract multi-scale features. The output of the convolutional layer is nonlinearly transformed through the ReLU activation function. Specifically, three convolutional layers are designed, each containing 32, 64, and 128 convolution kernels. The convolution kernel size is 3x3, the stride is 1, and the padding is 1. The pooling layer uses maximum pooling, the pooling window size is 2x2, and the stride is 2, which can effectively reduce the size of the feature map, retain the main features, and accelerate the calculation. This architecture design improves the model's ability to resolve complex images and enhances its recognition accuracy and generalization ability;

[0073] Generate feature maps: CNN generates feature maps for images of each perspective. These feature maps contain the spatial structure and texture information of the target object. The generated feature maps can be represented as high-dimensional vectors for subsequent feature fusion and classification. This process effectively captures the target features from multiple perspectives and provides a rich source of information for subsequent feature fusion and classification. These high-dimensional feature vectors enable the model to understand and describe the target object more comprehensively, significantly improving the efficiency and accuracy of feature fusion, thereby enhancing the performance and robustness of classification tasks.

[0074] As the instruction manual Figure 5 As shown, the feature fusion step specifically includes the following steps:

[0075] Design feature fusion layer: Design feature fusion layer to integrate features from different perspectives by weighted fusion, dynamically adjust the weights of features from each perspective, and improve the ability to understand the target. The input of the feature fusion layer is feature maps from different perspectives, and the output is a feature map that integrates multi-perspective information. This design enables the model to comprehensively utilize multi-perspective information, effectively capture the full range of features of the target object, and improve the comprehensiveness and accuracy of feature extraction. The dynamic weight adjustment mechanism further optimizes the feature fusion process, ensuring that the key information of each perspective is fully utilized, thereby enhancing the recognition performance and robustness of the model;

[0076] Design attention mechanism: A multi-head attention mechanism is used to adaptively weight feature maps from different perspectives, so that the model can pay more attention to important features. Specifically, three attention heads are designed, and the attention weight of each head is calculated by the softmax function. The weighted sum of the weight and the feature map is used as the final fusion result, so that each feature map can dynamically adjust its importance during the fusion process. This mechanism effectively strengthens the expression of key information and improves the representativeness of feature maps, thereby enhancing the model's recognition ability and classification accuracy for complex patterns, so that the final fusion result can better reflect the true characteristics of the target object;

[0077] Fusion results: The feature map that integrates multi-view information is used for subsequent target classification and positioning. The fused feature map retains the important information of each view while reducing the influence of noise and irrelevant features. This optimized feature map not only enhances the representativeness of the features, enabling the model to more accurately identify and locate target objects, but also improves the robustness and generalization ability of the overall system, thereby showing better performance in complex environments.

[0078] Target classification and positioning specifically include:

[0079] Classification layer: After feature fusion, target classification is performed through the fully connected layer, and the softmax function is used as the output layer of the classifier to output the probability distribution of the target category. The input of the classification layer is the fused feature map, and the output is the probability distribution of the target category, which effectively realizes the efficient use and accurate classification of the fused features. This method not only improves the accuracy and robustness of classification, but also ensures that the model can effectively handle complex and changeable feature information, thereby showing better performance in various application scenarios. The final target category probability distribution provides a reliable basis for subsequent decision-making and significantly enhances the system's recognition and judgment capabilities;

[0080] Regression layer: While classifying, the three-dimensional position and posture of the target are determined through regression analysis. The output of the regression layer is the three-dimensional coordinates and rotation matrix of the target object. The input of the regression layer is the fused feature map, and the output is the three-dimensional coordinates (x, y, z) and rotation matrix (R) of the target. This design helps to achieve accurate positioning and posture estimation of the target, and improves the application ability of the model in complex environments, such as robot navigation, augmented reality and other tasks. At the same time, combined with classification information, the regression layer provides important support for comprehensive target recognition and three-dimensional space understanding, and enhances the efficiency and accuracy of the system in multi-task processing;

[0081] Loss function: The cross entropy loss function is used to calculate the classification loss, and the mean square error loss function is used to calculate the regression loss, and the model performance is improved through joint optimization. Specifically, the cross entropy loss function is used for classification loss, and the mean square error loss function is used for regression loss. The total loss function is obtained by weighted sum of the two. This joint optimization strategy significantly improves the performance of the model. By optimizing the classification and regression tasks separately, the model can more effectively learn the category information and spatial position of the target, thereby achieving more accurate recognition and positioning. In addition, this weighted method allows the relative importance of classification and regression to be adjusted according to specific task requirements, so as to optimize the overall performance in different scenarios and improve the stability and accuracy of the system in complex task environments.

[0082] As the instruction manual Figure 6 As shown, the model training step specifically includes the following steps:

[0083] Dataset preparation: Use annotated 3D target datasets for model training. The datasets should contain images from different perspectives and under different lighting conditions. The datasets can be obtained from public datasets (such as KITTI, PASCAL3D+) or built from self-collected data. The model can be exposed to more diverse scenes and targets, thereby learning more comprehensive and accurate feature representations. This method not only improves the model's ability to identify and locate targets in various real environments, but also helps reduce the risk of overfitting, making the model more stable and reliable in unseen scenes and conditions, ultimately improving the practicality and application scope of the overall system.

[0084] Select a training strategy for training: adopt a phased training strategy, first train the feature extraction network, then fix the feature extraction network parameters, and train the classification and regression networks. Through phased training, the convergence speed and accuracy of the model can be improved. Use a phased training strategy, first train the feature extraction network, then fix its parameters, and then train the classification and regression networks. This strategy can significantly improve the convergence speed and accuracy of the model. Through phased optimization, the model can gradually focus on learning tasks at different levels, avoiding the convergence difficulties caused by too many parameters in the initial stage. In addition, after fixing the feature extraction network parameters, the classification and regression networks can more intensively learn how to effectively use the extracted features, thereby improving the performance and stability of the overall model. This strategy not only speeds up the training process, but also ensures that the model can achieve optimal performance at each stage, and ultimately achieves efficient and accurate target recognition and positioning;

[0085] Select optimizer for optimization: Use Adam optimizer to optimize model parameters and adjust the learning rate to meet the needs of different training stages. The initial learning rate is set to 0.001, and the learning rate decays to the original 0.1 after every 10 epochs. This can reduce the shock in the late training period and make the model more stably approach the local optimal solution. This flexible learning rate scheduling strategy combined with the adaptive learning ability of Adam optimizer helps to balance the convergence speed and training stability, thereby improving the overall performance and training effect of the model.

[0086] Training monitoring: Monitor the model training process through the validation set to prevent overfitting and ensure the generalization ability of the model. After each training epoch, use the validation set to evaluate the performance of the model. If the loss of the validation set continues to increase, stop training in advance to avoid overfitting. By regularly using the validation set to monitor the model training process, overfitting can be effectively prevented and the model can be ensured to have good generalization ability. Use the validation set to evaluate the model performance after each epoch to timely capture the performance changes of the model on unseen data. If the loss of the validation set continues to increase, it indicates that the model may begin to overfit the training data. Stopping training in advance can avoid a decrease in model performance and a waste of resources. This monitoring strategy helps maintain the stability and accuracy of the model in practical applications and ensures that it still performs well in new data environments.

[0087] Real-time target recognition specifically includes:

[0088] Real-time data acquisition: In practical applications, multiple cameras are used to capture the image data of the target object in real time. The camera can be connected to the computing device using a USB interface or a network interface to achieve high-speed data transmission, significantly improving the real-time and flexibility of the system. This configuration not only ensures the timely acquisition and processing of data, but also can cope with complex and changing environments and adapt to different application scenarios. High-speed data transmission ensures that the computing device can quickly receive and process a large amount of image information, thereby improving the efficiency and accuracy of target detection and recognition. In addition, the deployment of a multi-camera system increases the redundancy and reliability of the system. Even if some cameras fail, the system can still maintain normal operation, further enhancing the stability and maintainability of the system. Overall, this method not only optimizes the process of real-time data processing, but also improves the performance and response speed of the entire system, ensuring efficient and reliable target tracking and recognition capabilities;

[0089] Real-time processing: Use the trained multi-view target recognition network model to perform real-time image processing, extract features and fuse them. Real-time processing can be performed on the GPU to ensure that the processing speed meets the real-time requirements. This real-time processing method not only greatly improves the processing efficiency and meets the needs of real-time applications, but also enhances the accuracy and robustness of recognition through the fusion of multi-view data. The parallel computing capability of the GPU enables complex image processing tasks to be completed quickly, thereby realizing the real-time tracking and recognition of targets in dynamic scenes, providing strong technical support for various real-time applications;

[0090] Recognition result output: Output the recognition category of the target object and its position and posture in three-dimensional space for subsequent application use. Outputting the recognition category of the target object and its position and posture in three-dimensional space can provide key decision-making information and operation basis for subsequent applications. The recognition results can be output through a display, network interface or API interface, which is convenient for integration with other systems and timely sharing of information. It not only improves the coordination ability between systems, but also supports a variety of application scenarios, such as automated control, monitoring systems and augmented reality. Through standardized output methods, it can simplify docking with third-party systems, accelerate the application of technology, and improve overall work efficiency and accuracy.

[0091] In summary, the method provided by the present invention can be extended to a variety of application fields, such as autonomous driving, robot navigation, virtual reality and augmented reality, and by adjusting the network structure and training data, the model can adapt to different application scenarios and target objects.

[0092] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A three-dimensional target recognition method based on a multi-view target recognition network, characterized in that: The specific steps include: Data acquisition, multi-view image acquisition, and recording of camera parameters; Data preprocessing: standardization, data enhancement, and perspective alignment of the collected image data; Feature extraction: using convolutional neural network technology to extract feature maps of each view; Feature fusion, using the attention mechanism to perform weighted fusion of features from different perspectives; Target classification and positioning, using fully connected networks for classification and regression processing; Model training, using labeled datasets to train and optimize deep learning; Real-time target recognition, real-time image acquisition and classification and positioning output.

2. The three-dimensional target recognition method based on a multi-view target recognition network according to claim 1, characterized in that: The multi-view image acquisition specifically includes: using multiple cameras to shoot the target object from different angles to obtain two-dimensional image data of multiple viewpoints; The recording of camera parameters specifically includes: recording the camera parameters of each viewing angle, including the camera position, direction and internal parameters.

3. The three-dimensional target recognition method based on a multi-view target recognition network according to claim 2, characterized in that: In the multi-view image acquisition step, at least three cameras need to be arranged, and each camera is arranged on the same plane, and the angle between the lens axes of two adjacent cameras is not less than 85°.

4. The three-dimensional target recognition method based on a multi-view target recognition network according to claim 2, characterized in that: In the step of recording camera parameters, the internal parameters of the camera include the focal length and principal point of the camera.

5. The three-dimensional target recognition method based on a multi-view target recognition network according to claim 1, characterized in that: The standardization process specifically includes: standardizing the collected multi-viewpoint images, adjusting the image size to a uniform size, and performing denoising; The data enhancement processing specifically includes: improving the visual quality of the image by contrast enhancement technology to make the features of the target object more obvious; The perspective alignment process specifically includes: using camera parameters to align images of different perspectives to ensure consistency of subsequent feature extraction.

6. The three-dimensional target recognition method based on a multi-view target recognition network according to claim 1, characterized in that: The feature extraction step specifically includes the following steps: Build a deep learning model and construct a convolutional neural network as a feature extractor to extract common features of images from different perspectives; Design convolutional layers, design multi-layer convolutional layers and pooling layers to extract multi-scale features. The output of the convolutional layer is nonlinearly transformed through the ReLU activation function; Generate feature maps. Images from each perspective generate feature maps through CNN. These feature maps contain the spatial structure and texture information of the target object.

7. The three-dimensional target recognition method based on a multi-view target recognition network according to claim 1, characterized in that: The feature fusion step specifically includes the following steps: Design a feature fusion layer to integrate features from different perspectives through weighted fusion; Design an attention mechanism and use a multi-head attention mechanism to adaptively weight feature maps from different perspectives, so that the model can pay more attention to important features; The fusion result, the feature map of multi-view information, is used for subsequent target classification and positioning.

8. The three-dimensional target recognition method based on a multi-view target recognition network according to claim 1, characterized in that: The target classification and positioning specifically include: Classification layer: After feature fusion, the target is classified through the fully connected layer, and the softmax function is used as the output layer of the classifier to output the probability distribution of the target category; Regression layer: While classifying, the three-dimensional position and posture of the target are determined through regression analysis. The output of the regression layer is the three-dimensional coordinates and rotation matrix of the target object; Loss function: The cross entropy loss function is used for classification loss calculation, and the mean square error loss function is used for regression loss calculation, and the model performance is improved through joint optimization.

9. The three-dimensional target recognition method based on a multi-view target recognition network according to claim 1, characterized in that: The model training step specifically includes the following steps: Dataset preparation: Use annotated 3D target datasets for model training. The datasets should contain images from different perspectives and under different lighting conditions. Select a training strategy for training. Use a phased training strategy to first train the feature extraction network, then fix the feature extraction network parameters, and train the classification and regression networks. Select the optimizer for optimization, use the Adam optimizer to optimize the model parameters, and adjust the learning rate to meet the needs of different training stages; Training monitoring: monitor the model training process through the validation set to prevent overfitting and ensure the generalization ability of the model.

10. The three-dimensional target recognition method based on a multi-view target recognition network according to claim 1, characterized in that: The real-time target recognition specifically includes: Real-time data acquisition, in practical applications, multiple cameras are used to capture the image data of the target object in real time; Real-time processing: using the trained multi-view target recognition network model to perform real-time image processing, extract features and perform fusion; The recognition result output outputs the recognition category of the target object and its position and posture in three-dimensional space for subsequent application use.

Images