Traffic condition detection method based on multi-channel target detection and graph neural network

Through the multi-channel object detection and graph neural network method, combined with YOLOv8, Faster R-CNN and SSD network, the target semantic relationship in traffic condition detection is captured, and the problems of limited area and noise information influence in existing traffic condition detection are solved. More efficient and accurate traffic condition detection is achieved.

CN118865281BActive Publication Date: 2024-12-03THE THIRD ENG CO LTD OF CHINA RAILWAY SEVENTH GRP +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411330542.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-24
Publication Date
2024-12-03
Estimated Expiration
2044-09-24

AI Technical Summary

Technical Problem

The existing traffic condition detection methods have problems such as limited camera shooting areas, complex vehicle flow and a lot of noise information, which affect the accuracy and efficiency of the target detection algorithm.

Method used

The multi-channel object detection and graph neural network methods are adopted, including multi-channel object detection module, target fusion learning module and target separate learning module. The target is detected through YOLOv8, Faster R-CNN and SSD networks, and the target is captured between targets in combination with graph neural networks. semantic relationships of , generate hidden vectors for category prediction.

Benefits of technology

Improve the accuracy and efficiency of target detection, capture more comprehensive target information, and improve the excellence of prediction tags and confidence.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118865281B_ABST
    Figure CN118865281B_ABST
Patent Text Reader

Abstract

This application relates to a traffic condition detection method based on multi-channel target detection and graph neural network, which uses YOLOv8, Faster R-CNN, and SSD in the field of target detection to identify targets in the graph, and then uses target individual learning and target fusion learning to capture the relationship between targets, strengthen the model's judgment of target categories, and finally calculate the dot product of the two target hidden vectors obtained, and use the FC layer to calculate the probability distribution of each target category. The method of this application captures more comprehensive targets, and the labels and confidence of the predicted targets are better.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular, to a traffic condition detection method based on multi-channel target detection and graph neural network. Background Art

[0002] In the current rapidly developing intelligent transportation system, road condition detection has become an important task in the construction of intelligent transportation. With the continuous acceleration of urbanization and the gradual increase in car users, traffic problems such as traffic congestion and vehicle management are becoming increasingly prominent. Traffic congestion not only brings troubles to people's lives and travel, but also seriously affects the overall development of the city and the quality of life of residents. Therefore, accurate and efficient detection of road traffic conditions is of great significance for further optimizing traffic design, improving urban traffic planning, and promoting the allocation of traffic resources.

[0003] The traditional method of traffic condition detection is to capture road traffic with a camera and then view it with a monitor. This method requires a lot of manpower and material resources to judge road conditions. It is not only time-consuming and labor-intensive, but also unable to efficiently process massive amounts of road traffic data. With the rapid development of machine learning and deep learning technologies, traffic condition detection methods based on deep learning have gradually attracted attention. Among them, the target detection algorithm can efficiently detect vehicles and pedestrians in the picture and count the number of each category of traffic roads in real time. Compared with traditional manual methods, the deep learning method based on target detection has higher efficiency and accuracy, and can better meet the needs of road condition detection. However, there are still some challenges and problems in the current road traffic condition detection technology. For example, the camera shooting area is limited, the vehicle flow on the road is complex, and there may be a lot of noise and irrelevant information in the captured pictures, which affects the target detection algorithm's accurate discrimination of the target category. Summary of the invention

[0004] In order to overcome at least one shortcoming in the prior art, the present application provides a traffic condition detection method based on multi-channel target detection and graph neural network.

[0005] In a first aspect, a traffic condition detection method based on multi-channel target detection and graph neural network is provided, comprising:

[0006] Obtain a model training data set; the samples in the model training data set are traffic condition pictures;

[0007] Construct a traffic condition detection model; the traffic condition detection model includes a multi-channel target detection module, a target fusion learning module, a target individual learning module and a category prediction module;

[0008] The multi-channel target detection module is used to perform multi-channel target detection on traffic condition images to obtain multiple preliminary target detection results; the target fusion learning module is used to fuse multiple preliminary target detection results, and construct a graph structure based on the fusion results, and use the graph neural network to obtain the first hidden vector according to the graph structure; the target separate learning module is used to construct a graph structure according to each preliminary target detection result, and use the graph neural network to obtain multiple second hidden vectors according to the graph structure, and splice the multiple second hidden vectors to obtain the spliced ​​hidden vector; the category prediction module is used to output the category probability distribution of the target according to the first hidden vector and the spliced ​​hidden vector;

[0009] The traffic condition detection model is trained based on the model training data set to obtain a trained traffic condition detection model;

[0010] The traffic condition picture to be detected is input into the trained traffic condition detection model to obtain the traffic condition detection result.

[0011] In one embodiment, the multi-channel target detection module includes a YOLOv8 network, a Faster R-CNN network and an SSD network; traffic condition pictures are respectively input into the YOLOv8 network, the Faster R-CNN network and the SSD network to obtain multiple preliminary target detection results.

[0012] In one embodiment, the target fusion learning module includes a repeated region removal module, a first ResNet module, a first cosine similarity calculation module, a first graph structure construction module and a first graph neural network;

[0013] The duplicate region removal module is used to remove duplicate targets from multiple preliminary target detection results to obtain processed target detection results; the processed target detection results include multiple targets;

[0014] The first ResNet module is used to convert each target in the processed target detection result into a feature vector;

[0015] The first cosine similarity calculation module is used to calculate the cosine similarity between any two feature vectors;

[0016] The first graph structure building module is used to build a graph structure, the nodes of the graph structure are multiple targets, and the two nodes constituting the edge of the graph structure are two targets corresponding to the cosine similarity that meets the set conditions;

[0017] The first graph neural network is used to capture the semantic relationship between objects in the graph structure and generate a first hidden vector.

[0018] In one embodiment, the target separate learning module includes a second ResNet module, a second cosine similarity calculation module, a second graph structure construction module, a second graph neural network and a splicing module;

[0019] The second ResNet module is used to convert each target in each preliminary target detection result into a feature vector; each preliminary target detection result includes multiple targets;

[0020] The second cosine similarity calculation module is used to calculate the cosine similarity between any two feature vectors for each preliminary target detection result;

[0021] The second graph structure construction module is used to construct a graph structure for each preliminary target detection result; the nodes of the graph structure are multiple targets, and the two nodes constituting the edge of the graph structure are two targets corresponding to the cosine similarity that meets the set conditions;

[0022] The second graph neural network is used to capture the semantic relationship between the objects in each graph structure and generate multiple second hidden vectors.

[0023] The concatenation module is used to concatenate multiple second hidden vectors to obtain a concatenated hidden vector.

[0024] In one embodiment, the category prediction module includes a dot product module and a fully connected layer; the dot product module is used to implement the dot product operation of the first hidden vector and the concatenated hidden vector; the result of the dot product operation passes through the fully connected layer to output the category probability distribution of the target.

[0025] In one embodiment, the total loss function used in the model training process is:

[0026]

[0027] in, is the total loss function, , , , They are , , and The corresponding weight parameter, is the loss corresponding to the YOLOv8 network, is the loss corresponding to the Faster R-CNN network, is the loss corresponding to the SSD network, is the label loss.

[0028] In a second aspect, a traffic condition detection device based on multi-channel target detection and graph neural network is provided, comprising:

[0029] The training data set acquisition module is used to acquire the model training data set; the samples in the model training data set are traffic condition pictures;

[0030] A model building module is used to build a traffic condition detection model; the traffic condition detection model includes a multi-channel target detection module, a target fusion learning module, a target individual learning module and a category prediction module;

[0031] The multi-channel target detection module is used to perform multi-channel target detection on traffic condition images to obtain multiple preliminary target detection results; the target fusion learning module is used to fuse multiple preliminary target detection results, and construct a graph structure based on the fusion results, and use the graph neural network to obtain the first hidden vector according to the graph structure; the target separate learning module is used to construct a graph structure according to each preliminary target detection result, and use the graph neural network to obtain multiple second hidden vectors according to the graph structure, and splice the multiple second hidden vectors to obtain the spliced ​​hidden vector; the category prediction module is used to output the category probability distribution of the target according to the first hidden vector and the spliced ​​hidden vector;

[0032] A model training module is used to train the traffic condition detection model based on the model training data set to obtain a trained traffic condition detection model;

[0033] The detection module is used to input the traffic condition picture to be detected into the trained traffic condition detection model to obtain the traffic condition detection result.

[0034] In a third aspect, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the above-mentioned traffic condition detection method based on multi-channel target detection and graph neural network is implemented.

[0035] Compared with the prior art, the present application has the following beneficial effects: the traffic condition detection method based on multi-channel target detection and graph neural network of the present application adopts YOLOv8, Faster R-CNN, and SSD in the field of target detection to identify the targets in the graph, and then adopts target individual learning and target fusion learning to capture the relationship between targets, strengthen the model's judgment on target categories, and finally calculates the dot product of the two target hidden vectors obtained, and uses the FC layer to calculate the probability distribution of each target category. The method of the present application captures more comprehensive targets, and the labels and confidence of the predicted targets are better. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The present application may be better understood by referring to the following description given in conjunction with the accompanying drawings, which together with the following detailed description are included in this specification and form a part of this specification. In the drawings:

[0037] Figure 1 A flowchart of a traffic condition detection method based on multi-channel target detection and graph neural network according to an embodiment of the present application is shown;

[0038] Figure 2 The structural block diagram of the traffic condition detection model is shown;

[0039] Figure 3 A structural block diagram of a traffic condition detection device based on multi-channel target detection and graph neural network according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0040] The exemplary embodiments of the present application will be described below in conjunction with the accompanying drawings. For the sake of clarity and conciseness, not all features of the actual embodiments are described in the specification. However, it should be understood that many implementation-specific decisions can be made in the process of developing any such actual embodiments in order to achieve the specific goals of the developer, and these decisions may vary from embodiment to embodiment.

[0041] It is also necessary to explain here that, in order to avoid obscuring the present application due to unnecessary details, only the device structure closely related to the scheme according to the present application is shown in the drawings, while other details that are not closely related to the present application are omitted.

[0042] It should be understood that the present application is not limited to the described implementation forms due to the following description with reference to the accompanying drawings. In this article, where feasible, the embodiments can be combined with each other, features between different embodiments can be replaced or borrowed, and one or more features can be omitted in one embodiment.

[0043] The present application embodiment provides a traffic condition detection method based on multi-channel target detection and graph neural network. Figure 1 A flow chart of a traffic condition detection method based on multi-channel target detection and graph neural network according to an embodiment of the present application is shown, see Figure 1 , methods include:

[0044] Step S1, obtaining a model training data set; the samples in the model training data set are traffic condition pictures.

[0045] Here, the model training data set may be a COCO data set, the target in the sample is a car, and the label in the sample is a car category. The model training data set may also include a bird's-eye view of road traffic captured by a camera.

[0046] Step S2, constructing a traffic condition detection model; Figure 2 The structural block diagram of the traffic condition detection model is shown in Figure 2,The traffic condition detection model includes a multi-channel target detection module, a target fusion learning module, a target individual learning module and a category prediction module;

[0047] The multi-channel target detection module is used to perform multi-channel target detection on traffic condition images to obtain multiple preliminary target detection results; the target fusion learning module is used to fuse multiple preliminary target detection results, and construct a graph structure based on the fusion results, and use the graph neural network to obtain the first hidden vector according to the graph structure; the target separate learning module is used to construct a graph structure according to each preliminary target detection result, and use the graph neural network to obtain multiple second hidden vectors according to the graph structure, and splice the multiple second hidden vectors to obtain the spliced ​​hidden vector; the category prediction module is used to output the category probability distribution of the target according to the first hidden vector and the spliced ​​hidden vector;

[0048] Step S3, training the traffic condition detection model based on the model training data set to obtain a trained traffic condition detection model.

[0049] Step S4, inputting the traffic condition picture to be detected into the trained traffic condition detection model to obtain the traffic condition detection result.

[0050] In one embodiment, see Figure 2 The multi-channel target detection module includes the YOLOv8 (You Only LookOnce version 8) network, the Faster R-CNN (Faster Region-based Convolutional Neural Network) network and the SSD (Single Shot MultiBox Detector) network; the traffic condition pictures are input into the YOLOv8 network, the Faster R-CNN network and the SSD network respectively to obtain multiple preliminary target detection results.

[0051] Each preliminary target detection result includes multiple targets. Each target includes a predicted label, location information and a detection area. The detection area refers to the area where the target is located, generally represented by a rectangular box. The location information is the coordinates of the center point of the detection area.

[0052] In this embodiment, a three-channel target detection structure of YOLOv8, Faster R-CNN and SSD is designed. Since the accuracy and performance of each target detection algorithm are different, the detected areas will also be different. The multi-channel structure can integrate the detection results of each target detection algorithm, and further jointly train and optimize to improve the generalization ability of the model.

[0053] In one embodiment, see Figure 2The target fusion learning module includes a repeated region removal module, a first ResNet module (Residual Network), a first cosine similarity calculation module, a first graph structure construction module and a first graph neural network (GGNN, Gated Graph Neural Network); the functions of each module are introduced in detail below.

[0054] The duplicate region removal module is used to remove duplicate targets from multiple preliminary target detection results to obtain processed target detection results; the processed target detection results include multiple targets; here, duplicate targets can be removed according to the position information of each target.

[0055] The first ResNet module is used to convert each target in the processed target detection result into a feature vector.

[0056] The first cosine similarity calculation module is used to calculate the cosine similarity between any two feature vectors.

[0057] The first graph structure construction module is used to construct a graph structure, the nodes of the graph structure are multiple targets, and the two nodes constituting the edge of the graph structure are two targets corresponding to the cosine similarity that meets the set conditions; here, the set condition can be, for example, that the cosine similarity is greater than a set threshold, and the set threshold can be, for example, 0.8, which is selected according to actual conditions without specific limitation.

[0058] The first graph neural network is used to capture the semantic relationship between objects in the graph structure and generate a first hidden vector.

[0059] This embodiment aims at the differences in the detection targets of different models, and fuses the results of three target detections to construct a graph structure. First, the areas with repeated detections are deleted according to the detected position coordinates, so as to achieve the complementary detection targets of the three detection algorithms, and finally obtain all the combined detection targets. Then ResNet is used to convert the detection targets into feature vectors for representation, and the cosine similarity calculation is used to connect the areas with high similarity to construct a graph structure. Graph neural networks can effectively capture the semantic relationship between detection targets, so that the model can more accurately identify the label category corresponding to the target. The multi-channel target detection algorithms YOLOv8, Faster R-CNN and SSD convert images into feature vectors for learning, ignoring the relationship modeling between targets in the image, while the GGNN model can capture the relationship between targets well.

[0060] In one embodiment, see Figure 2 The target independent learning module includes the second ResNet module, the second cosine similarity calculation module, the second graph structure construction module, the second graph neural network (GGNN) and the splicing module. The functions of each module are introduced in detail below.

[0061] The second ResNet module is used to convert each target in each preliminary target detection result into a feature vector; each preliminary target detection result includes multiple targets;

[0062] The second cosine similarity calculation module is used to calculate the cosine similarity between any two feature vectors for each preliminary target detection result.

[0063] The second graph structure construction module is used to construct a graph structure for each preliminary target detection result; the nodes of the graph structure are multiple targets, and the two nodes that constitute the edge of the graph structure are two targets corresponding to the cosine similarity that meets the set conditions; here, the set condition can be, for example, that the cosine similarity is greater than a set threshold, and the set threshold can be, for example, 0.8, which is selected according to actual conditions without specific limitation.

[0064] The second graph neural network is used to capture the semantic relationship between the objects in the graph structure for each graph structure and generate multiple second hidden vectors. Here, multiple second graph neural networks can be set to generate second hidden vectors for each graph structure respectively.

[0065] The concatenation module is used to concatenate multiple second hidden vectors to obtain a concatenated hidden vector.

[0066] In this embodiment, the target separate learning module can build semantic connections for each target detection result and learn the target categories in the image more comprehensively. First, the ResNet model is used to calculate the feature vector of the target, and then the cosine similarity between the feature vectors is calculated to build graph structures for the three detection structures. Then, the GGNN model is used to learn the three graph structures and learn the semantic relationship between targets under each target detection algorithm.

[0067] In one embodiment, the category prediction module includes a dot product module and a fully connected layer; the dot product module is used to implement the dot product operation of the first hidden vector and the concatenated hidden vector; the result of the dot product operation passes through the fully connected layer (FC, Fully Connected Layer) to output the category probability distribution of the target.

[0068] In one embodiment, the total loss function used in the model training process is:

[0069]

[0070] in, is the total loss function, , , , They are , , and The corresponding weight parameter, is the loss corresponding to the YOLOv8 network, is the loss corresponding to the Faster R-CNN network, is the loss corresponding to the SSD network, is the label loss.

[0071] Specifically, the label loss :

[0072]

[0073] in, is the category number, C is the number of categories, For the The true labels of the categories, For the predicted The probability distribution of the categories.

[0074] , and Both include three parts of loss, namely positioning loss, confidence loss and classification loss. Among them, positioning loss is used to measure the gap between the predicted bounding box and the true bounding box. Confidence loss is used to measure the confidence of whether the predicted bounding box contains the object. Classification loss is used to measure the gap between the predicted target category and the true category.

[0075] In order to further verify the effectiveness of the method of this application, the detection results of the method of this application and the existing methods are compared and analyzed. Accuracy, precision, recall and macro-F1 are calculated as evaluation indicators. Table 1 shows the comparison results.

[0076] Table 1

[0077]

[0078] From the experimental results, it can be seen that the performance of the method of the present application on traffic images and video data is significantly improved compared to the baseline model. In the image traffic target detection results, the method of the present application captures targets significantly more comprehensively, and the labels and confidence of the predicted targets are better. The present application adopts a multi-channel structure, and the results of the detection by integrating YOLOv8, Faster R-CNN and SSD methods can more comprehensively detect targets in the image. In addition, the present application designs a target fusion learning module and a target separate learning module, and uses the GGNN model to learn the graph results constructed by the target detection results, capturing the deep semantic relationship between targets, thereby optimizing the model's performance in predicting the target category.

[0079] Using the same inventive concept as the traffic condition detection method based on multi-channel target detection and graph neural network, this embodiment also provides a corresponding traffic condition detection device based on multi-channel target detection and graph neural network. Figure 3 The structure block diagram of the traffic condition detection device based on multi-channel target detection and graph neural network according to an embodiment of the present application is shown, including:

[0080] The training data set acquisition module 31 is used to acquire a model training data set; the samples in the model training data set are traffic condition pictures;

[0081] A model building module 32 is used to build a traffic condition detection model; the traffic condition detection model includes a multi-channel target detection module, a target fusion learning module, a target individual learning module and a category prediction module;

[0082] The multi-channel target detection module is used to perform multi-channel target detection on the traffic condition image to obtain multiple preliminary target detection results; the target fusion learning module is used to fuse multiple preliminary target detection results, and build a graph structure based on the fusion results, and use the graph neural network to obtain the first hidden vector according to the graph structure; the target separate learning module is used to build a graph structure according to each preliminary target detection result, and use the graph neural network to obtain multiple second hidden vectors according to the graph structure, and splice the multiple second hidden vectors to obtain the spliced ​​hidden vector; the category prediction module is used to output the category probability distribution of the target according to the first hidden vector and the spliced ​​hidden vector;

[0083] A model training module 33 is used to train the traffic condition detection model based on the model training data set to obtain a trained traffic condition detection model;

[0084] The detection module 34 is used to input the traffic condition picture to be detected into the trained traffic condition detection model to obtain the traffic condition detection result.

[0085] The traffic condition detection device based on multi-channel target detection and graph neural network of this embodiment has the same inventive concept as the traffic condition detection method based on multi-channel target detection and graph neural network mentioned above. Therefore, the specific implementation method of the device can be seen in the implementation example part of the traffic condition detection method based on multi-channel target detection and graph neural network mentioned above, and its technical effect corresponds to the technical effect of the above method, which will not be repeated here.

[0086] An embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned traffic condition detection method based on multi-channel target detection and graph neural network.

[0087] In summary, this application has the following technical effects:

[0088] This application uses YOLOv8, Faster R-CNN, and SSD in the field of target detection to identify targets in the image, and then uses target individual learning and target fusion learning to capture the relationship between targets, strengthen the model's judgment of target categories, and finally calculate the dot product of the two target hidden vectors obtained, and use the FC layer to calculate the probability distribution of each target category. The method of this application captures more comprehensive targets, and the labels and confidence of the predicted targets are better.

[0089] The above are only various implementations of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A traffic condition detection method based on multi-channel target detection and graph neural network, characterized in that: include: Get the model training dataset; The samples in the model training data set are traffic condition pictures; Constructing a traffic condition detection model; the traffic condition detection model includes a multi-channel target detection module, a target fusion learning module, a target individual learning module and a category prediction module; The multi-channel target detection module is used to perform multi-channel target detection on the traffic condition picture to obtain multiple preliminary target detection results; the target fusion learning module is used to fuse the multiple preliminary target detection results, and construct a graph structure based on the fusion results, and use a graph neural network to obtain a first hidden vector according to the graph structure; The target individual learning module is used to construct a graph structure according to each of the preliminary target detection results, and use a graph neural network to obtain a plurality of second hidden vectors according to the graph structure, and concatenate the plurality of second hidden vectors to obtain a concatenated hidden vector; the category prediction module is used to output a category probability distribution of the target according to the first hidden vector and the concatenated hidden vector; Training the traffic condition detection model based on the model training data set to obtain a trained traffic condition detection model; Inputting the traffic condition picture to be detected into the trained traffic condition detection model to obtain the traffic condition detection result; The multi-channel target detection module includes a YOLOv8 network, a Faster R-CNN network and an SSD network; the traffic condition picture is respectively input into the YOLOv8 network, the Faster R-CNN network and the SSD network to obtain a plurality of preliminary target detection results; The target fusion learning module includes a repeated region removal module, a first ResNet module, a first cosine similarity calculation module, a first graph structure construction module and a first graph neural network; The repeated region removal module is used to remove repeated targets in the multiple preliminary target detection results to obtain processed target detection results; the processed target detection results include multiple targets; The first ResNet module is used to convert each target in the processed target detection result into a feature vector; The first cosine similarity calculation module is used to calculate the cosine similarity between any two feature vectors; The first graph structure construction module is used to construct a graph structure, the nodes of the graph structure are the multiple targets, and the two nodes constituting the edge of the graph structure are the two targets corresponding to the cosine similarity that meets the set conditions; The first graph neural network is used to capture the semantic relationship between the objects in the graph structure and generate a first hidden vector; The target separate learning module includes a second ResNet module, a second cosine similarity calculation module, a second graph structure construction module, a second graph neural network and a splicing module; The second ResNet module is used to convert each target in each preliminary target detection result into a feature vector; each of the preliminary target detection results includes multiple targets; The second cosine similarity calculation module is used to calculate the cosine similarity between any two feature vectors for each preliminary target detection result; The second graph structure construction module is used to construct a graph structure for each preliminary target detection result; the nodes of the graph structure are the multiple targets, and the two nodes constituting the edge of the graph structure are the two targets corresponding to the cosine similarity that meets the set conditions; The second graph neural network is used to capture the semantic relationship between the objects in each graph structure and generate a plurality of second hidden vectors; The splicing module is used to splice the multiple second hidden vectors to obtain a spliced ​​hidden vector; The category prediction module includes a dot product module and a fully connected layer; the dot product module is used to implement the dot product operation of the first hidden vector and the concatenated hidden vector; the result of the dot product operation passes through the fully connected layer to output the category probability distribution of the target.

2. The method according to claim 1, characterized in that During the training process of the model, the total loss function used is: in, is the total loss function, , , , They are , , and The corresponding weight parameter, is the loss corresponding to the YOLOv8 network, is the loss corresponding to the Faster R-CNN network, is the loss corresponding to the SSD network, is the label loss.

3. A traffic condition detection device based on multi-channel target detection and graph neural network, characterized in that: include: A training data set acquisition module is used to acquire a model training data set; The samples in the model training data set are traffic condition pictures; A model building module, used to build a traffic condition detection model; the traffic condition detection model includes a multi-channel target detection module, a target fusion learning module, a target individual learning module and a category prediction module; The multi-channel target detection module is used to perform multi-channel target detection on the traffic condition picture to obtain multiple preliminary target detection results; the target fusion learning module is used to fuse the multiple preliminary target detection results, and construct a graph structure based on the fusion results, and use a graph neural network to obtain a first hidden vector according to the graph structure; The target individual learning module is used to construct a graph structure according to each of the preliminary target detection results, and use a graph neural network to obtain a plurality of second hidden vectors according to the graph structure, and concatenate the plurality of second hidden vectors to obtain a concatenated hidden vector; the category prediction module is used to output a category probability distribution of the target according to the first hidden vector and the concatenated hidden vector; A model training module, used to train the traffic condition detection model based on the model training data set to obtain a trained traffic condition detection model; A detection module, used for inputting the traffic condition picture to be detected into the trained traffic condition detection model to obtain the traffic condition detection result; The multi-channel target detection module includes a YOLOv8 network, a Faster R-CNN network and an SSD network; the traffic condition picture is respectively input into the YOLOv8 network, the Faster R-CNN network and the SSD network to obtain a plurality of preliminary target detection results; The target fusion learning module includes a repeated region removal module, a first ResNet module, a first cosine similarity calculation module, a first graph structure construction module and a first graph neural network; The repeated region removal module is used to remove repeated targets in the multiple preliminary target detection results to obtain processed target detection results; the processed target detection results include multiple targets; The first ResNet module is used to convert each target in the processed target detection result into a feature vector; The first cosine similarity calculation module is used to calculate the cosine similarity between any two feature vectors; The first graph structure construction module is used to construct a graph structure, the nodes of the graph structure are the multiple targets, and the two nodes constituting the edge of the graph structure are the two targets corresponding to the cosine similarity that meets the set conditions; The first graph neural network is used to capture the semantic relationship between the objects in the graph structure and generate a first hidden vector; The target separate learning module includes a second ResNet module, a second cosine similarity calculation module, a second graph structure construction module, a second graph neural network and a splicing module; The second ResNet module is used to convert each target in each preliminary target detection result into a feature vector; each of the preliminary target detection results includes multiple targets; The second cosine similarity calculation module is used to calculate the cosine similarity between any two feature vectors for each preliminary target detection result; The second graph structure construction module is used to construct a graph structure for each preliminary target detection result; the nodes of the graph structure are the multiple targets, and the two nodes constituting the edge of the graph structure are the two targets corresponding to the cosine similarity that meets the set conditions; The second graph neural network is used to capture the semantic relationship between the objects in each graph structure and generate a plurality of second hidden vectors; The splicing module is used to splice the multiple second hidden vectors to obtain a spliced ​​hidden vector; The category prediction module includes a dot product module and a fully connected layer; the dot product module is used to implement the dot product operation of the first hidden vector and the concatenated hidden vector; the result of the dot product operation passes through the fully connected layer to output the category probability distribution of the target.

4. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the traffic condition detection method based on multi-channel target detection and graph neural network as described in any one of claims 1-2.