Intelligent vehicle deployment and control method

Abstract

The invention provides an intelligent vehicle deployment and control method, and relates to the technical field of intelligent traffic management, and the method comprises the steps: obtaining the original data of each modal collected by each modal sensor, and carrying out the feature extraction of each modal original data based on a preset feature extraction rule, so as to determine an extraction result corresponding to each modal original data; performing dynamic multi-modal feature fusion based on each extraction result and a preset feature fusion rule to determine a comprehensive feature vector; performing similarity matching with a deployment and control target library based on the comprehensive feature vector, and dynamically adjusting deployment and control parameters according to a matching result; and performing deployment and control early warning on the vehicle according to the adjusted deployment and control parameters, and optimizing the feature extraction and feature fusion model based on user feedback data.

Description

Technical Field

[0001] This application belongs to the field of intelligent traffic management technology, specifically relating to an intelligent vehicle deployment method. Background Technology

[0002] In intelligent traffic management systems, existing vehicle deployment methods typically rely on single-modal features (such as license plate recognition) for vehicle identification and tracking. However, in complex dynamic scenarios, such as inclement weather, vehicle occlusion, multiple vehicles operating in parallel, varying lighting conditions, or driver camouflage, single-modal features are easily affected by interference, leading to low deployment accuracy, high false alarm rates, and poor system applicability. More specifically, traditional methods lack the ability to collaboratively fuse and adaptively adjust multimodal features, failing to effectively handle inconsistencies between features, temporal dynamics, and environmental interference, thus hindering accurate, real-time, and robust vehicle deployment. Furthermore, in large-scale real-time video stream processing, traditional methods, due to computational resource limitations and fixed strategies, cannot integrate multi-source heterogeneous data and dynamically optimize deployment strategies under low-latency requirements, further limiting system performance improvement. Summary of the Invention

[0003] This application provides a method for deploying intelligent vehicles to solve one of the aforementioned technical problems.

[0004] The technical solution adopted in this application is as follows: This application provides an intelligent vehicle deployment method, including: The original modal data collected by each modal sensor are acquired, and features are extracted from each of the original modal data based on preset feature extraction rules to determine the extraction results corresponding to each of the original modal data. Dynamic multimodal feature fusion is performed based on the extraction results and preset feature fusion rules to determine the comprehensive feature vector; Based on the comprehensive feature vector, a similarity match is performed with the target database, and the deployment parameters are dynamically adjusted according to the matching results; Based on the adjusted deployment parameters, vehicles are deployed and warned, and the feature extraction and feature fusion models are optimized based on user feedback data.

[0005] According to one embodiment of this application, the step of performing feature extraction on the original data of each modality based on preset feature extraction rules to determine the extraction result corresponding to each original data of each modality includes: Based on the first feature extraction model, license plate features, appearance features, and driver features are extracted from the original video data and / or original image data to determine the first extraction result; Based on the second feature extraction model, spatiotemporal features are extracted from the original video data and / or original image data to determine the second extraction result; The first feature extraction model and the second feature extraction model share the underlying convolutional features.

[0006] According to one embodiment of this application, before performing feature extraction on each of the original modal data based on a preset feature extraction rule to determine the extraction result corresponding to each of the original modal data, the method further includes: Adaptive filtering and image enhancement processing are performed on the original video data and / or original image data to determine the preprocessed data; The preprocessed data is resized and normalized to fit the input requirements of the feature extraction model.

[0007] According to one embodiment of this application, the step of performing dynamic multimodal feature fusion based on the extraction results and preset feature fusion rules to determine a comprehensive feature vector includes: A fusion model based on graph neural networks is constructed, with each extraction result as a node and the relationship between features as an edge, and the weights of each feature are dynamically learned. A context-aware mechanism is introduced to adaptively adjust feature weights based on real-time environmental data and historical deployment results to determine the comprehensive feature vector.

[0008] According to one embodiment of this application, the introduction of a context-aware mechanism to adaptively adjust feature weights based on real-time environmental data and historical deployment results includes: Based on real-time environmental data such as lighting conditions and weather information, as well as false alarm rate data from historical deployment results, the weights of license plate features, appearance features, driver features, and spatiotemporal features are dynamically adjusted.

[0009] According to one embodiment of this application, the step of performing similarity matching between the comprehensive feature vector and the deployment target database, and dynamically adjusting the deployment parameters based on the matching results, includes: A metric learning algorithm is used to calculate the similarity score between the comprehensive feature vector and the target features in the control target database, wherein the similarity score calculation combines cosine similarity and Mahalanobis distance; Based on the similarity score, a control warning is triggered, and the control parameters are dynamically adjusted through a reinforcement learning agent. The control parameters include the warning threshold, the control range, and the control duration.

[0010] According to one embodiment of this application, the optimization of the feature extraction and feature fusion model based on user feedback data includes: The system receives user confirmation or adjustment feedback on the deployment and early warning through a visual interactive interface. Based on the user feedback data, the feature extraction model and feature fusion model are incrementally learned and updated using an online learning engine.

[0011] A second aspect of this application provides an intelligent vehicle deployment device, comprising: The feature extraction module is used to acquire the original modal data collected by each modal sensor, and to extract features from each of the original modal data based on preset feature extraction rules, so as to determine the extraction results corresponding to each of the original modal data. The feature fusion module is used to perform dynamic multimodal feature fusion based on the extraction results and preset feature fusion rules to determine the comprehensive feature vector; The matching and adjustment module is used to perform similarity matching between the comprehensive feature vector and the control target library, and dynamically adjust the control parameters according to the matching results; The deployment execution module is used to issue deployment warnings to vehicles based on the adjusted deployment parameters, and to optimize the feature extraction and feature fusion models based on user feedback data.

[0012] A third aspect of this application provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps described in the method.

[0013] A fourth aspect of this application provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method as described.

[0014] Due to the adoption of the above technical solution, the beneficial effects achieved by this application are as follows: This application overcomes the limitations of single-modal features by acquiring multimodal raw data collected by various modal sensors and extracting license plate features, appearance features, driver features, and spatiotemporal features based on preset feature extraction rules, thereby enhancing feature coverage and environmental adaptability. Especially in complex scenarios (such as occlusion or changes in lighting), multimodal feature extraction ensures the capture of key information, thus significantly improving deployment accuracy.

[0015] Dynamic multimodal feature fusion is performed based on the extracted results and preset feature fusion rules. Graph neural networks and other models are used to dynamically learn feature weights, and a context-aware mechanism (such as real-time environmental data) is introduced to adaptively adjust feature importance. This enables the system to flexibly respond to different environmental conditions (such as nighttime or severe weather), reduce the impact of feature inconsistencies, and improve fusion performance and real-time response capabilities.

[0016] By performing similarity matching between the comprehensive feature vector and the target database (combining metric learning methods such as cosine similarity and Mahalanobis distance), and dynamically adjusting deployment parameters (such as warning threshold, deployment range, and deployment duration) based on the matching results, the system can accurately identify target vehicles and reduce false alarms. Simultaneously, a reinforcement learning-based parameter adjustment strategy further optimizes deployment decisions and improves system efficiency.

[0017] The system provides vehicle control and early warning based on the adjusted deployment parameters, and optimizes the feature extraction and feature fusion models online based on user feedback data (such as confirmation or adjustment records). This closed-loop adaptive learning mechanism ensures that the system can continuously learn and adapt to new scenarios and threats after deployment, reducing the need for manual intervention and improving long-term performance and user experience. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating an intelligent vehicle deployment method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0019] Figure label: 810, Processor; 820, Communication interface; 830, Memory; 840, Communication bus. Detailed Implementation

[0020] To more clearly illustrate the overall concept of this application, a detailed explanation is provided below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below. It should be noted that, unless otherwise specified, the embodiments of this application and the features thereof can be combined with each other.

[0022] In this application, unless otherwise expressly specified and limited, the "above" or "below" of the second feature can mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples.

[0023] Example 1 like Figure 1 As shown, a method for deploying intelligent vehicles includes: The original modal data collected by each modal sensor are acquired, and features are extracted from each of the original modal data based on preset feature extraction rules to determine the extraction results corresponding to each of the original modal data.

[0024] As described above, this step is the initial data processing stage of the intelligent vehicle deployment system. It aims to acquire raw data from multiple sensor sources and transform this raw data into structured feature representations using preset feature extraction rules, providing a foundation for subsequent multimodal fusion and deployment decisions. Specifically, "acquiring raw data from each modality of the sensors" involves real-time acquisition of heterogeneous data streams from distributed sensor networks (such as cameras, LiDAR, environmental sensors, etc.). This data may include videos, images, environmental parameters (such as lighting, weather), etc., ensuring the diversity and comprehensiveness of the data sources. "Extracting features from the raw data of each modality based on preset feature extraction rules" means processing the data of each modality independently according to predefined rules or models (such as deep learning architectures) to extract key features, such as extracting vehicle appearance features from video streams, license plate information from images, or contextual features from environmental data. "Determining the extraction results corresponding to the raw data of each modality" means converting the extracted features into standardized vectors or representations, such as feature maps or embedding vectors, to facilitate subsequent fusion and matching. The core objective of this step is to overcome the limitations of a single data source through multimodal feature extraction, thereby enhancing the robustness and accuracy of the system in complex scenarios.

[0025] For example, in practical applications of intelligent traffic management systems, the system might be deployed on highways or urban intersections, using cameras to capture real-time video streams, LiDAR to acquire 3D point cloud data of vehicles, and environmental sensors to record light intensity and weather conditions. In the feature extraction stage, pre-defined rules might include using convolutional neural networks (CNNs) to extract appearance features such as vehicle color, model, and texture from video frames, employing an improved optical character recognition (OCR) model to process the license plate area to extract license plate character features, and combining this with long short-term memory networks (LSTMs) to analyze vehicle trajectories to extract spatiotemporal features. For environmental data, rules might involve simple statistical processing (such as mean or variance calculations) to generate contextual feature vectors. Ultimately, the extraction results for each modality are output as independent feature vectors, such as appearance feature vectors, license plate feature vectors, and spatiotemporal feature vectors. These vectors serve as inputs to subsequent fusion modules, ensuring the system can comprehensively integrate multi-source information for precise deployment.

[0026] It should be noted that, in specific implementation scenarios, the above solutions can be expanded to include the integration of more modal sensors in terms of data acquisition. For example, infrared cameras can be used for nighttime monitoring, or microphone arrays can be used for sound feature acquisition to further enhance data diversity. Secondly, adaptive learning mechanisms can be introduced for feature extraction rules. For example, the parameters of the feature extraction model can be dynamically adjusted through online learning to cope with changes in data distribution or new scenarios (such as extreme weather). In addition, the rules can be combined with a multi-task learning framework to allow feature extraction from different modalities to share the underlying network structure, reducing computational redundancy and improving feature consistency. Furthermore, the preprocessing stage can also be expanded. For example, generative adversarial networks (GANs) can be used for data augmentation or denoising to improve the accuracy of feature extraction on low-quality data. Finally, the representation of the extraction results can be extended to a hierarchical or graph structure. For example, features can be organized into multi-scale vectors or relationship graphs to better capture the semantic relationships between features.

[0027] Dynamic multimodal feature fusion is performed based on the extraction results and preset feature fusion rules to determine the comprehensive feature vector.

[0028] As described above, this step is a core processing step in the intelligent vehicle deployment system, aiming to deeply fuse feature extraction results from different modalities to generate a comprehensive vector representation that fully characterizes vehicle features. "Based on the extraction results" refers to utilizing multiple independent feature vectors obtained in the preceding steps, including license plate features, appearance features, driver features, and spatiotemporal features. "Preset feature fusion rules" define the specific mechanisms and methods for effectively integrating these heterogeneous features, with the core being dynamic and adaptive feature weight allocation. "Perform dynamic multimodal feature fusion" emphasizes that the fusion process is not a simple feature concatenation or fixed weight superposition, but rather a dynamic adjustment of the fusion strategy based on real-time scene context and feature quality. Finally, "determining the comprehensive feature vector" outputs a unified feature representation rich in multimodal information, providing high-quality input for subsequent similarity matching. This process effectively solves the complementarity and redundancy issues between multimodal features by establishing semantic associations and environmental awareness mechanisms, significantly improving the robustness and discriminative ability of feature representation.

[0029] For example, in a specific traffic monitoring scenario, when the system simultaneously acquires a vehicle's visual features (such as license plate recognition results and vehicle color), spatiotemporal features (such as driving trajectory), and environmental context features (such as current lighting conditions), the preset feature fusion rule might adopt a fusion architecture based on graph neural networks: treating each modality feature as a node in a graph, and dynamically learning the correlation strength between nodes through an attention mechanism. For instance, under low-light conditions at night, the system automatically reduces the dependence weight on color features while increasing the attention to license plate recognition and trajectory features; when a vehicle is partially occluded, the system enhances the contribution of spatiotemporal features, compensating for the lack of visual features by analyzing the vehicle's motion patterns. In practice, the system might first perform dimensional alignment and standardization on each feature, then propagate feature information through multi-layer graph convolution operations, and finally generate a unified comprehensive feature vector through pooling operations. This vector not only contains the original information of each modality but, more importantly, captures the semantic correlation between modalities, enabling it to maintain high representational ability even when the quality of some modal features is poor.

[0030] It should be noted that, in specific implementation scenarios, the above solutions can be expanded to include more complex neural network structures, such as hierarchical graph attention networks, to achieve feature interactions at different granular levels. Regarding dynamism, reinforcement learning frameworks can be combined to automatically optimize the fusion strategy based on historical deployment results, forming a closed-loop learning mechanism of fusion-evaluation-optimization. In terms of context awareness, richer environmental factors, such as traffic flow density and weather severity, can be further integrated to make weight adjustments more refined and accurate. Furthermore, an uncertainty estimation mechanism can be introduced into the fusion process, assigning a confidence score to each feature to fully consider the credibility of each feature during fusion. Cross-scenario transfer learning schemes can also be designed to enable the fusion model to quickly adapt to new deployment environments.

[0031] Based on the comprehensive feature vector, a similarity match is performed with the target database, and the deployment parameters are dynamically adjusted according to the matching results.

[0032] As described above, this step is the decision-making and optimization stage in the intelligent vehicle deployment system, realizing closed-loop control from feature matching to strategy adjustment. "Similarity matching based on the comprehensive feature vector and the deployment target library" refers to comparing the comprehensive feature vector generated in the previous steps with the pre-established deployment target feature library, and evaluating the degree of matching with each target using a specific similarity metric. "Dynamically adjusting deployment parameters based on matching results" emphasizes that the system does not use fixed thresholds or rules, but rather adaptively optimizes key parameters in the deployment strategy based on real-time matching results and historical performance data. This process, by establishing a feedback mechanism between matching quality and parameter adjustment, solves the problem of decreased matching accuracy due to scene changes in traditional methods, achieving continuous optimization of the deployment strategy and scene adaptation.

[0033] For example, in practical applications, after the system generates a comprehensive feature vector of a suspect vehicle, it calculates the similarity between this vector and the features of all registered vehicles in the target database. For instance, the system might use a hybrid metric combining cosine similarity and Mahalanobis distance, considering both the consistency of feature direction and the characteristics of feature distribution, thus obtaining a more accurate matching score. If the matching score exceeds the currently set warning threshold, a control warning is triggered. Simultaneously, the system continuously monitors the matching effect: if it detects a significant increase in the false alarm rate for a certain type of vehicle during specific time periods (such as peak traffic hours) or in specific areas (such as around schools), the system will automatically adjust relevant parameters through a reinforcement learning agent, such as appropriately increasing the warning threshold for that type of vehicle or narrowing the control range in that area. Furthermore, for frequently occurring specific types of targets, the system may automatically extend their control duration to ensure the effectiveness of continuous monitoring. This dynamic adjustment mechanism allows the system to adapt to constantly changing traffic environments and control needs.

[0034] It should be noted that, in specific implementation scenarios, the above solutions can be further enhanced by introducing multi-level matching strategies and establishing a tiered target library. Different similarity thresholds can be applied to targets with different risk levels. Furthermore, an online target library update function can be integrated to support the real-time addition of new targets or the updating of existing target features during the matching process. Regarding parameter tuning, a multi-objective optimization framework based on deep reinforcement learning can be established, considering multiple performance indicators such as accuracy, response time, and resource consumption. A distributed parameter coordination mechanism based on federated learning can also be introduced, enabling multiple systems deployed in different regions to share parameter tuning experience. In addition, historical traceability and visualization functions for parameter tuning can be designed to facilitate managers' understanding of the system's decision-making logic and enable necessary manual intervention.

[0035] Based on the adjusted deployment parameters, vehicles are deployed and warned, and the feature extraction and feature fusion models are optimized based on user feedback data.

[0036] As described above, this step is the execution and self-optimization stage in the intelligent vehicle deployment system, realizing closed-loop adaptive learning from deployment decision-making to model iteration. Specifically, "providing deployment warnings to vehicles based on adjusted deployment parameters" means that the system uses dynamically optimized parameters (such as warning thresholds, deployment range, and duration) to monitor and trigger alarms on target vehicles in real time, ensuring that deployment actions match the needs of the current scenario. "Optimizing feature extraction and feature fusion models based on user feedback data" emphasizes that the system drives online updates of feature extraction and fusion models by collecting user interaction data such as confirmation of warning results, adjustments, or additions to deployment targets, thereby continuously improving the system's accuracy and adaptability. This process, by directly integrating user feedback into the model training loop, solves the performance degradation problem caused by fixed models in traditional systems, achieving the co-evolution of deployment strategies and user experience.

[0037] For example, in a highway monitoring scenario, the system issues a warning to a suspected cloned vehicle based on dynamically adjusted parameters (such as a higher warning threshold set after reducing the weight of color features at night). Traffic management personnel view the warning information through a web-based visual interface, confirming the vehicle as the real target, while marking another falsely reported vehicle as an "irrelevant target." This user interaction data is captured in real time and fed back to the system optimization engine: the system may employ incremental learning techniques to enhance the current feature extraction model's ability to handle blurred license plates using confirmation records, while adjusting feature fusion rules based on false alarm data to reduce the fusion weight of appearance features in low-light environments. For instance, if users frequently adjust the deployment results under specific weather conditions, the system will automatically fine-tune the environmental perception module in the fusion model, making the model focus more on trajectory features rather than easily affected visual features. This optimization not only improves the deployment accuracy in the current scenario but also allows the system to gradually adapt to new threat patterns, such as new vehicle camouflage techniques, through accumulated feedback data.

[0038] It should be noted that, in specific implementation scenarios, a multi-source feedback integration mechanism can be introduced on top of the above solutions. This not only integrates direct user interaction but also incorporates third-party data sources (such as traffic violation records) as indirect feedback to enhance the data foundation for model optimization. A distributed optimization framework based on federated learning can be established, enabling multiple systems deployed in different regions to share model update experience while protecting local data privacy. A hierarchical feedback processing flow can also be designed, employing differentiated model update strategies for different types of feedback data (such as confirmations, false alarms, and new targets), for example, implementing more aggressive parameter adjustments for high-frequency false alarm types. Furthermore, the optimization process can be extended with meta-learning capabilities, enabling the system to learn from historical feedback how to adapt to new scenarios more quickly and reduce reliance on large amounts of new data; or a feedback quality assessment mechanism can be introduced to automatically identify and weight high-credibility user feedback, avoiding interference from noisy data on model optimization.

[0039] According to one embodiment of this application, the step of performing feature extraction on the original data of each modality based on preset feature extraction rules to determine the extraction result corresponding to each original data of each modality includes: Based on the first feature extraction model, license plate features, appearance features, and driver features are extracted from the original video data and / or original image data to determine the first extraction result; Based on the second feature extraction model, spatiotemporal features are extracted from the original video data and / or original image data to determine the second extraction result; The first feature extraction model and the second feature extraction model share the underlying convolutional features.

[0040] As described above, the first feature extraction model processes the raw video data and / or raw image data to extract three types of key features. License plate features are extracted using an improved optical character recognition model combined with an attention mechanism, specifically addressing license plate recognition problems under conditions of blur, occlusion, or uneven lighting. Appearance features are extracted using a convolutional neural network, including visual attributes such as vehicle body color, vehicle type, and surface texture, and adversarial training is employed to enhance its robustness to environmental changes. Driver features are extracted using a face recognition network to extract the driver's facial features and establish a correlation with vehicle information.

[0041] The second feature extraction model processes the same raw video data and / or raw image data to specifically extract spatiotemporal features. These features are analyzed using a temporal analysis model to examine the vehicle's motion trajectory and behavior patterns in consecutive frames, capturing dynamic information such as the vehicle's speed, acceleration, and direction of motion.

[0042] The first and second feature extraction models employ a shared underlying convolutional feature architecture. This means that both models share the same convolutional layers to extract basic visual features during the initial processing stage, avoiding redundant computation, ensuring consistency in the underlying representation of different modalities, and effectively reducing the system's computational burden and memory usage. This sharing mechanism ensures that the extracted features have a unified semantic foundation, providing a solid basis for subsequent feature fusion.

[0043] According to one embodiment of this application, before performing feature extraction on each of the original modal data based on a preset feature extraction rule to determine the extraction result corresponding to each of the original modal data, the method further includes: Adaptive filtering and image enhancement processing are performed on the original video data and / or original image data to determine the preprocessed data; The preprocessed data is resized and normalized to fit the input requirements of the feature extraction model.

[0044] As described above, adaptive filtering and image enhancement processing are performed on the raw video data and / or raw image data. Adaptive filtering dynamically selects filtering parameters based on real-time environmental conditions, effectively suppressing noise interference. Image enhancement processing employs techniques including a denoising model based on generative adversarial networks, dynamically adjusting the processing intensity for different environments to improve image clarity and feature discernibility, thereby obtaining preprocessed data.

[0045] The preprocessed data undergoes resizing and normalization. Resizing scales the image data to a fixed size required by the feature extraction model, ensuring consistent input data specifications. Normalization standardizes pixel values ​​to a specific range, eliminating dimensional differences and ensuring the data distribution conforms to the input requirements of the feature extraction model, providing a standardized data foundation for subsequent feature extraction.

[0046] These two preprocessing steps effectively improve the quality and standardization of the raw data, providing an important guarantee for the accuracy and stability of subsequent feature extraction.

[0047] According to one embodiment of this application, the step of performing dynamic multimodal feature fusion based on the extraction results and preset feature fusion rules to determine a comprehensive feature vector includes: A fusion model based on graph neural networks is constructed, with each extraction result as a node and the relationship between features as an edge, and the weights of each feature are dynamically learned. A context-aware mechanism is introduced to adaptively adjust feature weights based on real-time environmental data and historical deployment results to determine the comprehensive feature vector.

[0048] As described above, a fusion model based on graph neural networks is constructed, where the extracted results of license plate features, appearance features, driver features, and spatiotemporal features are treated as nodes in a graph structure, and the relationships between features are treated as edges, forming a feature relationship graph. Through the message passing mechanism of graph neural networks, the importance weights of each feature node are dynamically learned, capturing the deep relationships between features.

[0049] A context-aware mechanism is introduced to adaptively adjust the weight allocation of each feature in the graph neural network based on real-time environmental data, including lighting conditions, weather conditions, and false alarm and accuracy data from historical deployment results. For example, the weight of color features is reduced under low-light conditions, while the weight of trajectory features is increased under severe weather conditions, enabling the feature fusion process to adapt to environmental changes.

[0050] By employing multi-layer propagation and aggregation operations within a graph neural network, combined with dynamically adjusted feature weights, features from various modalities are deeply fused to generate a unified comprehensive feature vector. This comprehensive feature vector retains the feature information of each modality while incorporating the correlations between features and environmental contextual information, providing a high-quality feature representation for subsequent similarity matching.

[0051] This fusion approach ensures the effectiveness and robustness of feature representations under different environmental conditions, improving the system's adaptability to complex scenarios.

[0052] According to one embodiment of this application, the introduction of a context-aware mechanism to adaptively adjust feature weights based on real-time environmental data and historical deployment results includes: Based on real-time environmental data such as lighting conditions and weather information, as well as false alarm rate data from historical deployment results, the weights of license plate features, appearance features, driver features, and spatiotemporal features are dynamically adjusted.

[0053] As described above, a fusion model based on graph neural networks is constructed, where the extracted results of license plate features, appearance features, driver features, and spatiotemporal features are treated as nodes in a graph structure, and the relationships between features are treated as edges, forming a feature relationship graph. Through the message passing mechanism of graph neural networks, the importance weights of each feature node are dynamically learned, capturing the deep relationships between features.

[0054] A context-aware mechanism is introduced to adaptively adjust the weight allocation of each feature in the graph neural network based on real-time environmental data, including lighting conditions, weather conditions, and false alarm and accuracy data from historical deployment results. For example, the weight of color features is reduced under low-light conditions, while the weight of trajectory features is increased under severe weather conditions, enabling the feature fusion process to adapt to environmental changes.

[0055] By employing multi-layer propagation and aggregation operations within a graph neural network, combined with dynamically adjusted feature weights, features from various modalities are deeply fused to generate a unified comprehensive feature vector. This comprehensive feature vector retains the feature information of each modality while incorporating the correlations between features and environmental contextual information, providing a high-quality feature representation for subsequent similarity matching.

[0056] This fusion approach ensures the effectiveness and robustness of feature representations under different environmental conditions, improving the system's adaptability to complex scenarios.

[0057] According to one embodiment of this application, the step of performing similarity matching between the comprehensive feature vector and the deployment target database, and dynamically adjusting the deployment parameters based on the matching results, includes: A metric learning algorithm is used to calculate the similarity score between the comprehensive feature vector and the target features in the control target database, wherein the similarity score calculation combines cosine similarity and Mahalanobis distance; Based on the similarity score, a control warning is triggered, and the control parameters are dynamically adjusted through a reinforcement learning agent. The control parameters include the warning threshold, the control range, and the control duration.

[0058] As described above, a metric learning algorithm is used to calculate the similarity score between the comprehensive feature vector and the target features in the target database. This calculation process combines two metrics: cosine similarity and Mahalanobis distance. Cosine similarity measures the directional consistency of feature vectors, while Mahalanobis distance addresses imbalanced distributions in high-dimensional feature spaces. This combination of metrics improves the accuracy and robustness of the similarity assessment.

[0059] The system triggers a control alert based on the calculated similarity score. When the similarity score exceeds the currently set alert threshold, the system automatically generates the corresponding control alert information.

[0060] The agent dynamically adjusts deployment parameters through reinforcement learning. Based on real-time matching results, historical false alarm data, and system performance indicators, the agent continuously optimizes three key deployment parameters: the warning threshold is dynamically adjusted according to the false alarm rate to ensure that the false alarm level is controlled while maintaining a high recall rate; the deployment range is flexibly adjusted according to the frequency of target occurrence and regional risk level; and the deployment duration is adaptively set according to the target threat level and historical behavior patterns.

[0061] This dynamic parameter adjustment mechanism enables the system to adapt to the deployment requirements of different scenarios, effectively reducing the false alarm rate and improving the overall deployment efficiency while ensuring system sensitivity.

[0062] According to one embodiment of this application, the optimization of the feature extraction and feature fusion model based on user feedback data includes: The system receives user confirmation or adjustment feedback on the deployment and early warning through a visual interactive interface. Based on the user feedback data, the feature extraction model and feature fusion model are incrementally learned and updated using an online learning engine.

[0063] As described above, the system receives user confirmation or adjustment feedback on deployment warnings through a visual interactive interface. This interface displays the deployment warning information generated by the system in real time, and users can confirm or adjust the warning targets, including marking false alarm targets, correcting target attributes, or adding new deployment targets. All user operation data is captured in real time and formed into a feedback dataset.

[0064] Based on the user feedback data, an online learning engine is used to incrementally learn and update the feature extraction model and the feature fusion model. The online learning engine employs incremental learning techniques, using user feedback data to fine-tune the model parameters, avoiding complete retraining. For the feature extraction model, the parameters of the feature extraction network are adjusted based on user-confirmed correct samples and false positive samples; for the feature fusion model, the feature weight allocation strategy is optimized based on user feedback. This incremental update method enables the model to continuously adapt to new scenario requirements and threat patterns while maintaining its original knowledge base, achieving the system's self-evolution capability.

[0065] A second aspect of this application provides an intelligent vehicle deployment device, comprising: The feature extraction module is used to acquire the original modal data collected by each modal sensor, and to extract features from each of the original modal data based on preset feature extraction rules, so as to determine the extraction results corresponding to each of the original modal data. The feature fusion module is used to perform dynamic multimodal feature fusion based on the extraction results and preset feature fusion rules to determine the comprehensive feature vector; The matching and adjustment module is used to perform similarity matching between the comprehensive feature vector and the control target library, and dynamically adjust the control parameters according to the matching results; The deployment execution module is used to issue deployment warnings to vehicles based on the adjusted deployment parameters, and to optimize the feature extraction and feature fusion models based on user feedback data.

[0066] A second aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any of the embodiments of the first aspect above.

[0067] Figure 2 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 2 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 may call logical instructions in the memory 830 to execute the method in any of the embodiments of the first aspect described above, the method including: The original modal data collected by each modal sensor are acquired, and features are extracted from each of the original modal data based on preset feature extraction rules to determine the extraction results corresponding to each of the original modal data. Dynamic multimodal feature fusion is performed based on the extraction results and preset feature fusion rules to determine the comprehensive feature vector; Based on the comprehensive feature vector, a similarity match is performed with the target database, and the deployment parameters are dynamically adjusted according to the matching results; Based on the adjusted deployment parameters, vehicles are deployed and warned, and the feature extraction and feature fusion models are optimized based on user feedback data.

[0068] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0069] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer being able to perform the methods provided by the above methods, the method comprising: The original modal data collected by each modal sensor are acquired, and features are extracted from each of the original modal data based on preset feature extraction rules to determine the extraction results corresponding to each of the original modal data. Dynamic multimodal feature fusion is performed based on the extraction results and preset feature fusion rules to determine the comprehensive feature vector; Based on the comprehensive feature vector, a similarity match is performed with the target database, and the deployment parameters are dynamically adjusted according to the matching results; Based on the adjusted deployment parameters, vehicles are deployed and warned, and the feature extraction and feature fusion models are optimized based on user feedback data.

[0070] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the methods provided by the above methods, the method comprising: The original modal data collected by each modal sensor are acquired, and features are extracted from each of the original modal data based on preset feature extraction rules to determine the extraction results corresponding to each of the original modal data. Dynamic multimodal feature fusion is performed based on the extraction results and preset feature fusion rules to determine the comprehensive feature vector; Based on the comprehensive feature vector, a similarity match is performed with the target database, and the deployment parameters are dynamically adjusted according to the matching results; Based on the adjusted deployment parameters, vehicles are deployed and warned, and the feature extraction and feature fusion models are optimized based on user feedback data.

[0071] Example 2 1. Data Acquisition and Preprocessing The system acquires multimodal raw data from a distributed sensor network, including real-time video streams and image data captured by cameras, as well as contextual data such as light intensity and weather information collected by environmental sensors. The preprocessing module performs adaptive filtering and image enhancement on the raw video and image data: adaptive filtering dynamically selects filtering parameters (such as Gaussian filtering or median filtering) based on real-time environmental noise to suppress noise; image enhancement employs a denoising model based on a generative adversarial network (GAN) to dynamically adjust the processing intensity to improve image clarity. After preprocessing, the data undergoes resizing (e.g., scaling to 224×224 pixels) and normalization (standardizing pixel values ​​to the [0,1] range) to adapt to the input requirements of the feature extraction model.

[0072] 2. Multimodal Feature Extraction Based on preset feature extraction rules, the system uses a multi-task deep learning model to extract multimodal features in parallel. Specifically: The first feature extraction model processes preprocessed video and image data to extract license plate features, appearance features, and driver features. License plate features are extracted using an improved OCR model (incorporating an attention mechanism) to handle blurry or occluded conditions. Appearance features are extracted using a CNN to capture vehicle color, model, and texture features, with adversarial training used to enhance robustness. Driver features are extracted using a face recognition network (such as ArcFace) to capture facial features and associate them with the vehicle. The output is the first extraction result, including license plate feature vector, appearance feature vector, and driver feature vector.

[0073] The second feature extraction model processes the same video and image data to extract spatiotemporal features. It analyzes vehicle trajectories using LSTM or Transformer models to capture temporal dynamic behaviors such as speed, acceleration, and direction of motion. The output is the second extraction result, i.e., the spatiotemporal feature vector.

[0074] The first and second feature extraction models share the underlying convolutional features, that is, they use the same convolutional layers to extract basic visual features, in order to reduce computational redundancy and improve feature consistency.

[0075] 3. Dynamic multimodal feature fusion Based on the extraction results, the system constructs a graph neural network (GNN) fusion model for dynamic feature fusion. License plate features, appearance features, driver features, and spatiotemporal features are used as graph nodes, and the relationships between features are used as edges. The weights of each feature are dynamically learned through graph convolution operations. A context-aware mechanism is introduced into the fusion process: feature weights are adaptively adjusted based on real-time environmental data (such as lighting conditions and weather information) and historical deployment results (such as false alarm rate data). For example, under low-light conditions, the weight of color features is reduced, while the weight of trajectory and license plate features is increased; in areas with high historical false alarm rates, the weight of features prone to false alarms is reduced. The fused comprehensive feature vector is calculated using the following formula:

[0076] Where z is the comprehensive feature vector. Let be the i-th feature vector, be the weights for dynamic learning, and N be the number of features. Weights Generated using the attention mechanism of a graph neural network, taking into account environmental context and historical data.

[0077] 4. Similarity matching and dynamic parameter adjustment The comprehensive feature vector is matched with the target database for similarity. A metric learning algorithm is used to calculate the similarity score, combining cosine similarity and Mahalanobis distance. Cosine similarity: measures the directional consistency of feature vectors, and the formula is:

[0078] Where t is a feature vector in the target library.

[0079] Mahalanobis distance: used to handle imbalances in high-dimensional features; the formula is:

[0080] Where Σ is the covariance matrix.

[0081] Similarity scores are combined as follows α and β are weighting coefficients. Control alerts are triggered based on similarity scores (e.g., when the score exceeds a threshold). Simultaneously, control parameters are dynamically adjusted via a reinforcement learning agent (e.g., DQN): the alert threshold is adjusted based on the real-time false alarm rate; the control range is adjusted based on the target's frequency of appearance and regional risk; and the control duration is adjusted based on the target's threat level. The reinforcement learning agent uses a reward function. , where γ is a coefficient, and the parameters are optimized through a policy network.

[0082] 5. Optimization of deployment, early warning, and feedback Based on the adjusted deployment parameters, the system provides real-time vehicle deployment warnings and displays the results through a web-based visualization interface. Users can confirm, adjust, or add new deployment targets, and user interaction data is captured in real time. Based on user feedback data, the system uses an online learning engine to incrementally learn and update the feature extraction and feature fusion models: it uses stochastic gradient descent (SGD) or similar algorithms to fine-tune model parameters, and the loss function includes cross-entropy loss and contrastive learning loss to ensure the model adapts to new scenarios. For example, for false alarms confirmed by the user, the system reduces the weight of relevant features; for newly added targets, the system updates the feature library. The optimization process runs in a "matching-feedback-update" loop to achieve continuous self-improvement.

[0083] For any parts not mentioned in this application, existing technologies may be used or referenced.

[0084] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0085] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A method for deploying intelligent vehicles, characterized in that, include: The original modal data collected by each modal sensor are acquired, and features are extracted from each of the original modal data based on preset feature extraction rules to determine the extraction results corresponding to each of the original modal data. Dynamic multimodal feature fusion is performed based on the extraction results and preset feature fusion rules to determine the comprehensive feature vector; Based on the comprehensive feature vector, similarity matching is performed with the control target database, and the control parameters are dynamically adjusted according to the matching results; Based on the adjusted deployment parameters, vehicles are deployed and warned, and the feature extraction and feature fusion models are optimized based on user feedback data.

2. The intelligent vehicle deployment method according to claim 1, characterized in that, The step of extracting features from the original data of each modality based on preset feature extraction rules to determine the extraction results corresponding to each original data of each modality includes: Based on the first feature extraction model, license plate features, appearance features, and driver features are extracted from the original video data and / or original image data to determine the first extraction result; Based on the second feature extraction model, spatiotemporal features are extracted from the original video data and / or original image data to determine the second extraction result; The first feature extraction model and the second feature extraction model share the underlying convolutional features.

3. The intelligent vehicle deployment method according to claim 2, characterized in that, Before performing feature extraction on the original data of each modality based on preset feature extraction rules to determine the extraction results corresponding to each original data of each modality, the method further includes: Adaptive filtering and image enhancement processing are performed on the original video data and / or original image data to determine the preprocessed data; The preprocessed data is resized and normalized to fit the input requirements of the feature extraction model.

4. The intelligent vehicle deployment method according to claim 1, characterized in that, The dynamic multimodal feature fusion based on the extraction results and preset feature fusion rules to determine the comprehensive feature vector includes: A fusion model based on graph neural networks is constructed, with each extraction result as a node and the relationship between features as an edge, and the weights of each feature are dynamically learned. A context-aware mechanism is introduced to adaptively adjust feature weights based on real-time environmental data and historical deployment results to determine the comprehensive feature vector.

5. The intelligent vehicle deployment method according to claim 4, characterized in that, The introduction of a context-aware mechanism, which adaptively adjusts feature weights based on real-time environmental data and historical deployment results, includes: Based on real-time environmental data such as lighting conditions and weather information, as well as false alarm rate data from historical deployment results, the weights of license plate features, appearance features, driver features, and spatiotemporal features are dynamically adjusted.

6. The intelligent vehicle deployment method according to claim 1, characterized in that, The step of performing similarity matching between the comprehensive feature vector and the target database, and dynamically adjusting the deployment parameters based on the matching results, includes: A metric learning algorithm is used to calculate the similarity score between the comprehensive feature vector and the target features in the control target database, wherein the similarity score calculation combines cosine similarity and Mahalanobis distance; Based on the similarity score, a control warning is triggered, and the control parameters are dynamically adjusted through a reinforcement learning agent. The control parameters include the warning threshold, the control range, and the control duration.

7. The intelligent vehicle deployment method according to claim 1, characterized in that, The optimization of the feature extraction and feature fusion model based on user feedback data includes: The system receives user confirmation or adjustment feedback on the deployment and early warning through a visual interactive interface. Based on the user feedback data, the feature extraction model and feature fusion model are incrementally learned and updated using an online learning engine.

8. An intelligent vehicle deployment and control device, characterized in that, include: The feature extraction module is used to acquire the original modal data collected by each modal sensor, and to extract features from each of the original modal data based on preset feature extraction rules, so as to determine the extraction results corresponding to each of the original modal data. The feature fusion module is used to perform dynamic multimodal feature fusion based on the extraction results and preset feature fusion rules to determine the comprehensive feature vector; The matching and adjustment module is used to perform similarity matching between the comprehensive feature vector and the control target library, and dynamically adjust the control parameters according to the matching results; The deployment execution module is used to issue deployment warnings to vehicles based on the adjusted deployment parameters, and to optimize the feature extraction and feature fusion models based on user feedback data.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-7.

Images