A visual target detection method and system based on hypergraph computation

Abstract

The application relates to the field of computers and discloses a visual target detection method and system based on hypergraph calculation, which collects traffic data of different modes of a city through a multi-source sensor device; the traffic data of different modes is subjected to space-time alignment and denoising, and multi-scale features are extracted through a feature encoder; a hypergraph is constructed by taking traffic participants, road infrastructure and dynamic environment information as hypergraph nodes and taking the interaction relationship between the nodes as hyperedges; in a hypergraph feature space, a classifier is used to predict the target categories of the nodes; target tracking and prediction are carried out based on a hypergraph reasoning mechanism, historical trajectory data and real-time traffic flow information are fused, and the target position is predicted. The method can solve the problems of low detection precision and low calculation efficiency of existing visual target detection technologies in complex traffic environments and is suitable for application scenarios such as urban road monitoring, traffic flow analysis and abnormal target detection.

Description

Technical Field

[0001] This invention relates to the field of computer vision, specifically a visual target detection method and system based on hypergraph computing. Background Technology

[0002] With the rapid development of smart cities, efficient and accurate visual object detection technology plays a crucial role in traffic safety-related fields such as urban safety management, traffic monitoring, and autonomous driving. Currently, object detection technology mainly relies on deep learning models, such as convolutional neural networks (CNNs) and the Transformer architecture. However, these methods face many challenges when dealing with complex and dynamic scenes:

[0003] The fusion of multi-source heterogeneous data is challenging: Visual data in smart city environments comes from a wide range of sources, including surveillance cameras, drone imagery, LiDAR point clouds, and GPS trajectories. How to effectively fuse this heterogeneous data to improve target detection accuracy remains a key issue.

[0004] Target occlusion and environmental interference: In real traffic environments, targets such as pedestrians and vehicles are often occluded or interfered with by the background (such as rainy or foggy weather, changes in lighting), making it difficult for traditional detection algorithms to obtain accurate results.

[0005] Insufficient spatiotemporal correlation modeling: Existing target detection methods are mostly based on frame-level feature extraction, lacking modeling of long-term target motion patterns, making it difficult to capture the temporal changes of targets.

[0006] High computational complexity: Traditional detection algorithms typically rely on high-dimensional feature spaces for computation, resulting in high computational resource consumption and making them unsuitable for real-time applications. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a visual target detection method based on hypergraph computation, comprising the following steps:

[0008] Step 1: Collect traffic data of different modes in the city using multi-source sensor devices;

[0009] Step 2: Perform spatiotemporal alignment and denoising on traffic data of different modalities, and extract multi-scale features through a feature encoder;

[0010] Step 3: Construct a hypergraph by using traffic participants, road infrastructure, and dynamic environmental information as nodes and the interaction relationships between nodes as hyperedges.

[0011] Step 4: Using a hypergraph neural network combined with a spatiotemporal attention mechanism, perform feature learning on the constructed hypergraph to extract high-order correlation features between targets;

[0012] Step 5: Target detection and classification. In the hypergraph feature space, a dynamic hyperedge weight adjustment mechanism is introduced. Combined with a multi-scale feature fusion module, the features of each node are integrated, and a classifier is used to predict the target category of the node.

[0013] Step 6: Target tracking and prediction. Based on the hypergraph inference mechanism, historical trajectory data and real-time traffic flow information are integrated to predict the target location.

[0014] Furthermore, the aforementioned collection of urban traffic data of different modalities through multi-source sensor devices includes:

[0015] The different modalities of traffic data include camera video streams, LiDAR point clouds, radar detection, GPS trajectories, and time-series traffic flow data.

[0016] Furthermore, the aforementioned spatiotemporal alignment and denoising of traffic data from different modalities, and the extraction of multi-scale features through a feature encoder, include:

[0017] Spatiotemporal alignment is based on sensor time synchronization mechanism, which synchronizes data of different modalities through timestamps; for image data, an adaptive median filtering method is used to remove noise; for point cloud data, a statistical filtering method is used to remove outliers; different feature encoders are used for different modalities, ResNet is used for visual data, PointNet++ is used for point cloud data, and a self-attention mechanism is combined for feature fusion.

[0018] Furthermore, the construction of the hypergraph by using traffic participants, road infrastructure, and dynamic environmental information as hypergraph nodes, and the interaction relationships between nodes as hyperedges, includes:

[0019] Building a hypergraph:

[0020]

[0021] in This represents a set of nodes, including vehicles, pedestrians, and roads. Represents a superedge set;

[0022] Superedge weight for:

[0023]

[0024] in Represents a node and The interaction strength between them is determined by the following formula:

[0025]

[0026] in, For nodes and The Euclidean distance between them and These represent the speeds of the corresponding targets. Indicates the similarity of the historical trajectories of nodes. , and For normalization parameters, , and These are weighting coefficients used to balance the contributions of spatial, velocity, and trajectory characteristics.

[0027] Furthermore, the aforementioned method of using a hypergraph neural network combined with a spatiotemporal attention mechanism to perform feature learning on the constructed hypergraph and extract high-order correlation features between targets includes:

[0028] The input layer of the hypergraph neural network uses high-dimensional feature encoding to map multimodal data onto the hypergraph structure, ensuring data distribution consistency. Hypergraph convolution is used to calculate feature updates, and its mathematical expression is as follows:

[0029]

[0030] in, It is the first Vertex features of the layer; and These are the degree matrices of the vertices and hyperedges, respectively; It is the incidence matrix of the hypergraph, representing the connection between vertices and hyperedges; It is the weight matrix of the hyperedge; It is a learnable parameter matrix.

[0031] Furthermore, the target detection and classification, in the hypergraph feature space, introduces a dynamic hyperedge weight adjustment mechanism, combined with a multi-scale feature fusion module, to integrate the features of each node, and uses a classifier to predict the target category of the node, including:

[0032] The dynamic hyperedge weighting method dynamically adjusts the connection weights of hyperedges based on the target category, motion pattern, and environment. The hyperedge weight calculation formula is as follows:

[0033]

[0034] in,

[0035]

[0036] , For speed, , For acceleration, , These are normalization parameters;

[0037] The features of each node are integrated, and a classifier is used to predict the category of the node. The category determination formula is as follows:

[0038]

[0039] in, Represents a node eigenvectors, This is the classifier weight matrix.

[0040] Furthermore, the target tracking and prediction, based on a hypergraph inference mechanism, integrates historical trajectory data and real-time traffic flow information to predict the target location, including:

[0041] Calculate the association probability between targets using the hypergraph attention mechanism:

[0042]

[0043] in, Contribution coefficient to the node The phase angle, For nodes The characteristic distribution; based on Bayesian inference, combined with historical trajectory data, the target location is predicted.

[0044] A visual target detection system based on hypergraph computation, applying the aforementioned visual target detection method based on hypergraph computation, includes a data acquisition unit, a data preprocessing unit, a hypergraph construction unit, a hypergraph feature extraction unit, a target detection and classification unit, a target tracking and prediction unit, a target tracking and prediction unit, and a data processing module;

[0045] The data acquisition unit is connected to the data preprocessing unit; the data preprocessing unit, the hypergraph construction unit, the hypergraph feature extraction unit, the target detection and classification unit, the target tracking and prediction unit, and the target tracking and prediction unit are respectively connected to the data processing module.

[0046] The data acquisition unit is used to collect multimodal traffic data, including camera video streams, lidar point clouds, radar detection, GPS trajectories, and time-series traffic flow data.

[0047] The data preprocessing unit is used to perform noise reduction, alignment, scale normalization, and multimodal feature extraction on the collected data.

[0048] Hypergraph building unit is used to construct the hypergraph structure, taking traffic participants, road facilities and dynamic environmental information as hypergraph nodes, and using hyperedges to connect multi-source information;

[0049] The hypergraph feature extraction unit is used to extract feature information for target detection and tracking based on the hypergraph neural network combined with the spatiotemporal attention mechanism.

[0050] The target detection and classification unit, combined with a dynamic hyperedge weighting mechanism, achieves accurate target detection and classification;

[0051] The target tracking and prediction unit integrates historical trajectory data and real-time traffic information to improve target prediction accuracy;

[0052] The result optimization unit employs reinforcement learning and hypergraph reasoning mechanisms to improve the stability of detection and tracking.

[0053] The beneficial effects of this invention are: This invention proposes a visual target detection method based on hypergraph computing, which can perform high-order association modeling on complex traffic scenarios in smart cities, and can also simultaneously model the spatial, temporal and semantic associations between targets, thereby improving the accuracy and robustness of target detection, which is of great significance to the development of [the invention / technology]. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a visual object detection method based on hypergraph computation.

[0055] Figure 2 This is a schematic flowchart of an embodiment of a visual object detection method based on hypergraph computation;

[0056] Figure 3 This is a schematic diagram illustrating multi-source data acquisition.

[0057] Figure 4 This is a schematic diagram of data preprocessing;

[0058] Figure 5 Construct a schematic diagram for the hypergraph;

[0059] Figure 6 This is a schematic diagram of target detection and classification.

[0060] Figure 7 This is a schematic diagram illustrating the predicted congestion scenario for a certain road section during peak hours. Detailed Implementation

[0061] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0062] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0063] like Figure 1 As shown, a visual object detection method based on hypergraph computation includes the following steps:

[0064] Step 1: Collect traffic data of different modes in the city using multi-source sensor devices;

[0065] Step 2: Perform spatiotemporal alignment and denoising on traffic data of different modalities, and extract multi-scale features through a feature encoder;

[0066] Step 3: Construct a hypergraph by using traffic participants, road infrastructure, and dynamic environmental information as nodes and the interaction relationships between nodes as hyperedges.

[0067] Step 4: Using a hypergraph neural network combined with a spatiotemporal attention mechanism, perform feature learning on the constructed hypergraph to extract high-order correlation features between targets;

[0068] Step 5: Target detection and classification. In the hypergraph feature space, a dynamic hyperedge weight adjustment mechanism is introduced. Combined with a multi-scale feature fusion module, the features of each node are integrated, and a classifier is used to predict the target category of the node.

[0069] Step 6: Target tracking and prediction. Based on the hypergraph inference mechanism, historical trajectory data and real-time traffic flow information are integrated to predict the target location.

[0070] Furthermore, the aforementioned collection of urban traffic data of different modalities through multi-source sensor devices includes:

[0071] The different modalities of traffic data include camera video streams, LiDAR point clouds, radar detection, GPS trajectories, and time-series traffic flow data.

[0072] Furthermore, the aforementioned spatiotemporal alignment and denoising of traffic data from different modalities, and the extraction of multi-scale features through a feature encoder, include:

[0073] Spatiotemporal alignment is based on sensor time synchronization mechanism, which synchronizes data of different modalities through timestamps; for image data, an adaptive median filtering method is used to remove noise; for point cloud data, a statistical filtering method is used to remove outliers; different feature encoders are used for different modalities, ResNet is used for visual data, PointNet++ is used for point cloud data, and a self-attention mechanism is combined for feature fusion.

[0074] Furthermore, the construction of the hypergraph by using traffic participants, road infrastructure, and dynamic environmental information as hypergraph nodes, and the interaction relationships between nodes as hyperedges, includes:

[0075] Building a hypergraph:

[0076]

[0077] in This represents a set of nodes, including vehicles, pedestrians, and roads. Represents a superedge set;

[0078] Superedge weight for:

[0079]

[0080] in Represents a node and The interaction strength between them is determined by the following formula:

[0081]

[0082] in, For nodes and The Euclidean distance between them and These represent the speeds of the corresponding targets. Indicates the similarity of the historical trajectories of nodes. , and For normalization parameters, , and These are weighting coefficients used to balance the contributions of spatial, velocity, and trajectory characteristics.

[0083] The method of learning features from a constructed hypergraph using a hypergraph neural network combined with a spatiotemporal attention mechanism to extract high-order correlation features between targets includes:

[0084] The input layer of the hypergraph neural network uses high-dimensional feature encoding to map multimodal data onto the hypergraph structure, ensuring data distribution consistency. Hypergraph convolution is used to calculate feature updates, and its mathematical expression is as follows:

[0085]

[0086] in, It is the first Vertex features of the layer; and These are the degree matrices of the vertices and hyperedges, respectively; It is the incidence matrix of the hypergraph, representing the connection between vertices and hyperedges; It is the weight matrix of the hyperedge; It is a learnable parameter matrix.

[0087] The target detection and classification, in the hypergraph feature space, introduces a dynamic hyperedge weight adjustment mechanism, combines a multi-scale feature fusion module to integrate the features of each node, and uses a classifier to predict the target category of the node, including:

[0088] The dynamic hyperedge weighting method dynamically adjusts the connection weights of hyperedges based on the target category, motion pattern, and environment. The hyperedge weight calculation formula is as follows:

[0089]

[0090] in,

[0091]

[0092] , For speed, , For acceleration, , These are normalization parameters;

[0093] The features of each node are integrated, and a classifier is used to predict the category of the node. The category determination formula is as follows:

[0094]

[0095] in, Represents a node eigenvectors, This is the classifier weight matrix.

[0096] The target tracking and prediction, based on a hypergraph inference mechanism, integrates historical trajectory data and real-time traffic flow information to predict the target location, including:

[0097] Calculate the association probability between targets using the hypergraph attention mechanism:

[0098]

[0099] in, Contribution coefficient to the node The phase angle, For nodes The characteristic distribution; based on Bayesian inference, combined with historical trajectory data, the target location is predicted.

[0100] A visual target detection system based on hypergraph computation, applying the aforementioned visual target detection method based on hypergraph computation, includes a data acquisition unit, a data preprocessing unit, a hypergraph construction unit, a hypergraph feature extraction unit, a target detection and classification unit, a target tracking and prediction unit, a target tracking and prediction unit, and a data processing module;

[0101] The data acquisition unit is connected to the data preprocessing unit; the data preprocessing unit, the hypergraph construction unit, the hypergraph feature extraction unit, the target detection and classification unit, the target tracking and prediction unit, and the target tracking and prediction unit are respectively connected to the data processing module.

[0102] The data acquisition unit is used to collect multimodal traffic data, including camera video streams, lidar point clouds, radar detection, GPS trajectories, and time-series traffic flow data.

[0103] The data preprocessing unit is used to perform noise reduction, alignment, scale normalization, and multimodal feature extraction on the collected data.

[0104] Hypergraph building unit is used to construct the hypergraph structure, taking traffic participants, road facilities and dynamic environmental information as hypergraph nodes, and using hyperedges to connect multi-source information;

[0105] The hypergraph feature extraction unit is used to extract feature information for target detection and tracking based on the hypergraph neural network combined with the spatiotemporal attention mechanism.

[0106] The target detection and classification unit, combined with a dynamic hyperedge weighting mechanism, achieves accurate target detection and classification;

[0107] The target tracking and prediction unit integrates historical trajectory data and real-time traffic information to improve target prediction accuracy;

[0108] The result optimization unit employs reinforcement learning and hypergraph reasoning mechanisms to improve the stability of detection and tracking.

[0109] Specifically, such as Figure 2 As shown, this embodiment provides a visual target detection method based on hypergraph computing, which can be used in scenarios such as intelligent traffic light control, abnormal target detection, and traffic flow analysis, and includes the following steps:

[0110] Step 1: Multi-source data acquisition. In a smart city environment, multi-source sensors, such as surveillance cameras installed at intersections and main roads, and LiDAR deployed on drones and fixed base stations, are used to collect real-time road images and videos, LiDAR point clouds, GPS trajectories, and time-series traffic flow data. Figure 3 As shown;

[0111] Step 2: Data preprocessing. Spatiotemporal alignment, denoising, and scale normalization are performed on data of different modalities. Multi-scale features are extracted using a feature encoder. Threshold segmentation, sliding window detection, and perspective transformation are applied to image and video data to extract clear target contours, such as... Figure 4As shown, the lidar point cloud data is filtered and resampled to unify the spatial scale; the time synchronization module is used to align the data from each data source according to the timestamp, and multi-scale features are extracted by the feature encoder; in this embodiment, video data is acquired at a rate of 30 frames per second, and a 128-dimensional feature vector is extracted for each video frame using a pre-trained convolutional encoder; at the same time, the point cloud data is filtered by Voxel Grid to make the point cloud data resolution uniform;

[0112] Step 3: Construct a hypergraph representation, using traffic participants (such as vehicles, pedestrians, and non-motorized vehicles), road facilities, and environmental information as hypergraph nodes, and establishing hyperedges based on spatiotemporal and semantic associations. The specific method is as follows:

[0113] Define hypergraph ,in It is a set of nodes, containing the feature vectors of each target;

[0114] Super Edge The weights of targets that are spatially close or have similar motion patterns are used to connect targets. The calculation formula is:

[0115]

[0116] in Represents a node and The interaction strength between them is calculated based on features such as distance, speed, and historical trajectory, using the following formula:

[0117]

[0118] in, For nodes and The Euclidean distance between them and These represent the speeds of the corresponding targets. Indicates the similarity of the historical trajectories of nodes. , and For normalization parameters, , and These are weighting coefficients used to balance the contributions of spatial, velocity, and trajectory characteristics.

[0119] like Figure 5 As shown, the system constructs dynamic hyperedges based on the distance, speed differences, and historical trajectory information between vehicles and other participants, ensuring that the hypergraph can reflect the spatiotemporal interaction relationships between targets;

[0120] Step 4: Hypergraph Feature Extraction. The constructed hypergraph is learned using a Hypergraph Neural Network (HGNN) to extract high-order correlation features between targets. The hypergraph convolution calculation formula is as follows:

[0121]

[0122] in, It is the first Vertex features of the layer; and These are the degree matrices of the vertices and hyperedges, respectively; It is the incidence matrix of the hypergraph, representing the connection between vertices and hyperedges; It is the weight matrix of the hyperedge; It is a learnable parameter matrix.

[0123] In this embodiment, the initial feature The multimodal features extracted in step 2 are subjected to two layers of hypergraph convolution to obtain an enhanced feature representation. .

[0124] Step 5: Object detection and classification. Using the multi-scale feature fusion module, the features of each node are integrated, and a classifier is used to predict the node's category. The category determination formula is as follows:

[0125]

[0126] in, Represents a node eigenvectors, This is the classifier weight matrix.

[0127] In this embodiment, the system sets predefined categories for vehicles, pedestrians, and non-motorized vehicles, and uses a classifier to accurately determine the targets, such as... Figure 6 As shown, the category information for each target bounding box has been correctly labeled;

[0128] Step 6: Target tracking and prediction. Calculate the association probability between targets using a hypergraph attention mechanism.

[0129]

[0130] in, Contribution coefficient to the node The phase angle, For nodes The characteristic distribution; based on Bayesian inference, combined with historical trajectory data, the target location at future moments is predicted.

[0131] In this embodiment, for a traffic congestion scenario during peak hours on a certain road segment, the system continuously tracks vehicle trajectories and predicts the driving path within the next 5 seconds, such as... Figure 7 As shown, this provides a basis for intelligent control of traffic lights;

[0132] Step 7: Result optimization. The hypergraph structure and hyperedge weights are dynamically adjusted through reinforcement learning algorithms to reduce errors caused by environmental interference. The detection and tracking results are fused across multiple time periods to improve the stability and accuracy of the overall system.

[0133] This example also provides a visual target detection system based on hypergraph computing, including: a data acquisition unit, a data preprocessing unit, a hypergraph construction unit, a hypergraph feature extraction unit, a target detection and classification unit, a target tracking and prediction unit, and a target tracking and prediction unit;

[0134] The data acquisition unit is used to collect multimodal traffic data, including camera video streams, LiDAR point clouds, radar detection, GPS trajectories, and time-series traffic flow data.

[0135] The data preprocessing unit performs noise reduction, alignment, scale normalization, and multimodal feature extraction on the collected data.

[0136] Hypergraph building unit is used to construct the hypergraph structure, taking traffic participants, road facilities and dynamic environmental information as hypergraph nodes, and using hyperedges to connect multi-source information;

[0137] The hypergraph feature extraction unit extracts feature information for target detection and tracking based on a hypergraph neural network combined with a spatiotemporal attention mechanism.

[0138] The target detection and classification unit, combined with a dynamic hyperedge weighting mechanism, achieves accurate target detection and classification;

[0139] The target tracking and prediction unit integrates historical trajectory data and real-time traffic information to improve target prediction accuracy;

[0140] The result optimization unit employs reinforcement learning and hypergraph reasoning mechanisms to improve the stability of detection and tracking.

[0141] This embodiment provides a visual target detection system based on hypergraph computation, which can be used to implement a visual target detection method based on hypergraph computation.

Claims

1. A visual target detection method based on hypergraph computation, characterized in that, Includes the following steps: Step 1: Collect traffic data of different modes in the city using multi-source sensor devices; Step 2: Perform spatiotemporal alignment and denoising on traffic data of different modalities, and extract multi-scale features through a feature encoder; Step 3: Construct a hypergraph by using traffic participants, road infrastructure, and dynamic environmental information as nodes and the interaction relationships between nodes as hyperedges. Step 4: Using a hypergraph neural network combined with a spatiotemporal attention mechanism, perform feature learning on the constructed hypergraph to extract high-order correlation features between targets; Step 5: Target detection and classification. In the hypergraph feature space, a dynamic hyperedge weight adjustment mechanism is first introduced. Combined with the multi-scale feature fusion module, the features of each node are integrated, and the classifier is used to predict the target category of the node. Step 6: Target tracking and prediction. Based on the hypergraph inference mechanism, historical trajectory data and real-time traffic flow information are integrated to predict the target location. The target tracking and prediction, based on a hypergraph inference mechanism, integrates historical trajectory data and real-time traffic flow information to predict the target location, including: Calculate the association probability between targets using the hypergraph attention mechanism: in, Contribution coefficient to the node The phase angle, For nodes The characteristic distribution; based on Bayesian inference, combined with historical trajectory data, the target location is predicted.

2. The visual target detection method based on hypergraph computation according to claim 1, characterized in that, The aforementioned method of collecting urban traffic data of different modalities using multi-source sensor devices includes: The different modalities of traffic data include camera video streams, LiDAR point clouds, radar detection, GPS trajectories, and time-series traffic flow data.

3. The visual target detection method based on hypergraph computation according to claim 1, characterized in that, The aforementioned spatiotemporal alignment and denoising of traffic data from different modalities, and extraction of multi-scale features through a feature encoder, includes: Spatiotemporal alignment is based on sensor time synchronization mechanism, which synchronizes data of different modalities through timestamps; for image data, an adaptive median filtering method is used to remove noise; for point cloud data, a statistical filtering method is used to remove outliers; different feature encoders are used for different modalities, ResNet is used for visual data, PointNet++ is used for point cloud data, and a self-attention mechanism is combined for feature fusion.

4. The visual target detection method based on hypergraph computation according to claim 1, characterized in that, The method of constructing a hypergraph by using traffic participants, road infrastructure, and dynamic environmental information as hypergraph nodes, and the interaction relationships between nodes as hyperedges, includes: Building a hypergraph: in This represents a set of nodes, including vehicles, pedestrians, and roads. Represents a superedge set; Superedge weight for: in Represents a node and The interaction strength between them is determined by the following formula: in, For nodes and The Euclidean distance between them and These represent the speeds of the corresponding targets. Indicates the similarity of the historical trajectories of nodes. and For normalization parameters, and These are weighting coefficients used to balance the contributions of spatial, velocity, and trajectory characteristics.

5. The visual target detection method based on hypergraph computation according to claim 1, characterized in that, The method of learning features from a constructed hypergraph using a hypergraph neural network combined with a spatiotemporal attention mechanism to extract high-order correlation features between targets includes: The input layer of the hypergraph neural network uses high-dimensional feature encoding to map multimodal data onto the hypergraph structure, ensuring data distribution consistency. Hypergraph convolution is used to calculate feature updates, and its mathematical expression is as follows: in, It is the first Vertex features of the layer; and These are the degree matrices of the vertices and hyperedges, respectively; It is the incidence matrix of the hypergraph, representing the connection between vertices and hyperedges; It is the weight matrix of the hyperedge; It is a learnable parameter matrix.

6. The visual target detection method based on hypergraph computation according to claim 1, characterized in that, The target detection and classification, in the hypergraph feature space, introduces a dynamic hyperedge weight adjustment mechanism, combines a multi-scale feature fusion module to integrate the features of each node, and uses a classifier to predict the target category of the node, including: The dynamic hyperedge weighting method dynamically adjusts the connection weights of hyperedges based on the target category, motion pattern, and environment. The hyperedge weight calculation formula is as follows: in, For speed, For acceleration, These are normalization parameters; The features of each node are integrated, and a classifier is used to predict the category of the node. The category determination formula is as follows: in, Represents a node eigenvectors, This is the classifier weight matrix.

7. A visual target detection system based on hypergraph computing, characterized in that, A visual target detection method based on hypergraph computing according to any one of claims 1-6 includes a data acquisition unit, a data preprocessing unit, a hypergraph construction unit, a hypergraph feature extraction unit, a target detection and classification unit, a target tracking and prediction unit, a target tracking and prediction unit, and a data processing module; The data acquisition unit is connected to the data preprocessing unit; the data preprocessing unit, the hypergraph construction unit, the hypergraph feature extraction unit, the target detection and classification unit, the target tracking and prediction unit, and the target tracking and prediction unit are respectively connected to the data processing module. The data acquisition unit is used to collect multimodal traffic data, including camera video streams, lidar point clouds, radar detection, GPS trajectories, and time-series traffic flow data. The data preprocessing unit is used to perform noise reduction, alignment, scale normalization, and multimodal feature extraction on the collected data. Hypergraph building unit is used to construct the hypergraph structure, taking traffic participants, road facilities and dynamic environmental information as hypergraph nodes, and using hyperedges to connect multi-source information; The hypergraph feature extraction unit is used to extract feature information for target detection and tracking based on the hypergraph neural network combined with the spatiotemporal attention mechanism. The target detection and classification unit, combined with a dynamic hyperedge weighting mechanism, achieves accurate target detection and classification; The target tracking and prediction unit integrates historical trajectory data and real-time traffic information to improve target prediction accuracy; The result optimization unit employs reinforcement learning and hypergraph reasoning mechanisms to improve the stability of detection and tracking.

Images