A drone traffic congestion and crowd gathering detection method based on AI recognition

Abstract

The present invention discloses a method for detecting traffic congestion and gathering of people using drones based on AI recognition, comprising the following steps: S1, obtaining real-time video and three-dimensional spatial data of the target area, and constructing a data set; S2, preliminarily processing the data set and extracting key features; S3, transmitting the extracted key features to a cloud server in real time through a 5G network, and constructing and optimizing an AI model; S4, performing multi-target detection and classification on the received key features, and generating multi-target detection results; S5, using an intelligent scheduling algorithm of fuzzy logic control and particle swarm optimization to perform global scheduling optimization; S6, using a graph neural network and an adaptive boosting algorithm to identify possible equipment failures in advance; S7, real-time monitoring of the flight status and recognition results of the drone; S8, displaying the recognition results and analysis reports through a visual interface. The present invention combines AI recognition with intelligent scheduling to optimize drone path planning and maintenance, and has the advantages of high efficiency, safety, and strong adaptability.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for detecting drone traffic congestion and personnel gathering based on AI recognition. Background Art

[0002] With the rapid development of drone technology, drones are increasingly used in traffic monitoring, disaster response, environmental monitoring and other fields. However, with the increase in the number of drones and the complexity of application scenarios, how to efficiently and accurately manage and monitor the flight of drones in complex environments has become a key issue. The current drone management system mainly relies on traditional flight control and monitoring technology. Although it can meet basic mission requirements, it still has many shortcomings when dealing with complex and changing environments. Especially in dynamic tasks such as traffic congestion and crowd gathering detection, the limitations of existing technologies are more obvious.

[0003] Existing drone management systems usually rely on preset flight paths and static task allocation. Although this method is simple and easy to use, it is often difficult to cope with dynamically changing environments in actual applications. During the execution of a mission, a drone may encounter sudden traffic jams or gatherings of people. Traditional static path planning and task scheduling methods are difficult to respond to these changes in real time, causing the drone to be stuck in traffic jams or unable to effectively cover the target area. In addition, traditional drone monitoring systems usually rely on limited sensor data and rule sets for anomaly detection. This method has low detection accuracy when facing complex and changing environments, and is prone to false positives or false negatives, which affects the performance of the drone's mission.

[0004] In terms of predictive maintenance of drones, existing technologies mostly adopt regular maintenance strategies based on fixed time intervals. Although this maintenance method can ensure the normal operation of drones to a certain extent, it also has obvious shortcomings. First, the regular maintenance method cannot timely reflect the real-time operating status of the equipment, resulting in some potential faults not being discovered in time, increasing the risk of equipment failure. Secondly, the regular maintenance method lacks flexibility and is prone to excessive maintenance or untimely maintenance, which not only increases maintenance costs, but may also cause equipment damage or even mission failure due to untimely maintenance.

[0005] In addition, in the multi-target detection and classification tasks of drones, existing technologies mainly rely on traditional computer vision algorithms, such as Sobel edge detection and Hough transform. Although these methods perform well in simple scenes, their performance significantly decreases in complex scenes, especially when multiple targets are densely packed and severely occluded. Traditional methods have difficulty effectively distinguishing overlapping targets or handling simultaneous detection and classification tasks of multiple types of targets, resulting in insufficient accuracy and reliability of detection results.

[0006] Most current UAV path planning technologies use traditional algorithms, such as the A* algorithm and the Dijkstra algorithm. These algorithms can provide better path planning solutions in simple scenarios, but they often show limitations when faced with multiple targets, complex environments, and dynamically changing scenarios. Traditional algorithms are usually unable to consider dynamic changes in the environment in real time, such as traffic conditions, sudden obstacles, etc., which makes path planning inflexible and difficult to adapt to the real-time needs of UAVs in complex environments. In addition, traditional path planning algorithms also have shortcomings in terms of computational complexity and resource consumption, especially in large-scale scenarios or high-density environments. Traditional algorithms are difficult to quickly generate effective path planning solutions.

[0007] Based on the above problems, with the development of artificial intelligence technology, more and more studies have begun to apply AI technology to the management and monitoring of drones. However, existing AI applications are mostly focused on specific tasks or scenarios, lacking systematicity and integration. For example, in multi-target detection and classification, although some studies have introduced deep learning technologies such as convolutional neural networks, these methods can often only handle single tasks or small-scale scenarios and cannot effectively cope with large-scale, multi-task complex environments. At the same time, in terms of path planning and task scheduling, existing AI applications are mostly focused on a single optimization goal, such as the shortest path or the lowest energy consumption, and lack comprehensive optimization of multiple objectives and multiple constraints.

[0008] Therefore, how to provide a method for detecting drone traffic congestion and crowd gathering based on AI recognition is an urgent problem that technicians in this field need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose a method for detecting traffic congestion and gathering of people by drones based on AI recognition. The present invention combines advanced technologies such as AI recognition technology, intelligent scheduling algorithm, and graph neural network, and optimizes the path planning, task scheduling, and equipment maintenance of drones in complex environments, significantly improving the efficiency and accuracy of drones in multi-target detection, fault prediction, and real-time monitoring. By adopting fuzzy logic control and particle swarm optimization algorithm, the present invention realizes the dynamic adjustment of drone paths and the intelligent generation of maintenance suggestions, and has the advantages of high efficiency, safety, and strong adaptability, and can better cope with complex and changing mission environments.

[0010] According to an embodiment of the present invention, a method for detecting drone traffic congestion and personnel gathering based on AI recognition includes the following steps:

[0011] S1. Use drones equipped with multispectral cameras and lidar to obtain real-time video and three-dimensional spatial data of the target area and build a data set;

[0012] S2: Perform edge computing locally on the drone to preliminarily process the data set and extract key features of vehicles, people, and obstacles.

[0013] S3, transmit the extracted key features to the cloud server in real time through the 5G network, combine historical data for large-scale data analysis and modeling, and build and optimize the AI ​​model;

[0014] S4. Use the optimized AI model to perform multi-target detection and classification on the received key features, identify and mark traffic congestion and crowd gathering in the area, and generate multi-target detection results;

[0015] S5. Based on the multi-target detection results, the intelligent scheduling algorithm of fuzzy logic control and particle swarm optimization is used to optimize the flight path and monitoring strategy of the drone in real time, and global scheduling optimization is performed in combination with the status of airport resources and mission requirements;

[0016] S6. Through the AI-driven predictive maintenance module, the operating status of the drone is continuously monitored, and the graph neural network and adaptive boosting algorithm are used to identify possible equipment failures in advance and issue maintenance recommendations;

[0017] S7, combined with the intelligent warning system, monitors the flight status and recognition results of the drone in real time, triggers an alarm when an abnormal situation is detected, and notifies the operator to intervene;

[0018] S8. Display the recognition results and analysis reports through a visual interface, and provide task progress tracking and adjustment functions.

[0019] Optionally, the S2 specifically includes:

[0020] S21. Preprocess the data sets obtained from the multispectral camera and lidar through the edge computing unit carried by the drone, including data format conversion, noise filtering and redundant data removal;

[0021] S22, in the preprocessed data set, using a weighted average filter to smooth the video data and the three-dimensional space data to reduce high-frequency noise and sudden abnormal values ​​in the data to ensure the stability and reliability of the data;

[0022] S23, through the feature extraction algorithm based on gray level co-occurrence matrix, the texture analysis is performed on the pre-processed video data to extract the surface texture features of vehicles, people and obstacles ,in represents the i-th texture feature extracted from the image;

[0023] S24. Use the edge detection algorithm based on the Sobel operator to detect the edge of the target in the video data, extract the contour features of the target, and extract the straight line and circle features through the Hough transform ,in represents the i-th feature extracted from edge detection;

[0024] S25. For lidar data, an improved RANSAC algorithm combined with weighted least squares method is used to perform plane fitting to extract the shape features of vehicles, people and obstacles from the three-dimensional space data. ,in Represents the i-th shape feature extracted from the point cloud data, and separates different targets through point cloud segmentation technology to generate a 3D contour point cloud for each target:

[0025] ;

[0026] Where P represents the point cloud dataset, a, b, c and d represent the coefficients of the fitted plane, Represents the weight of a point (x, y, z):

[0027] ;

[0028] in, and represents the adjustment parameter, Represents the local density of a point (x, y, z) in a point cloud.

[0029] Optionally, the S3 specifically includes:

[0030] S31, transmit key feature data extracted from the local edge computing unit of the drone to the cloud server in real time through the 5G network;

[0031] S32. On the cloud server, the received feature data is stored in real time, and the data sharding technology is used to manage large-scale data. The data sharding matrix S=\left [ {{s}_{ij}} \right ] is defined. ,in Indicates the jth storage location of the i-th data block after sharding, and calculates the storage index of the data block through the hash function H:

[0032] ;

[0033] in, Represents a data block The hash value of , N represents the total number of storage nodes;

[0034] S33. On the cloud server, the received feature data is clustered and analyzed in combination with historical data. The density-based clustering algorithm is used to perform preliminary clustering of vehicles, personnel and obstacles. The clusters after clustering are defined as , where k represents the number of the cluster:

[0035] ;

[0036] in, Indicates is the center and the radius is Neighborhood, MinPts represents the minimum number of samples in the cluster;

[0037] S34. Based on the clustering results, the feature data is input into an AI model based on a graph convolutional neural network combined with a self-attention mechanism for large-scale data modeling, and the input matrix Z=\left [ {{z}_{mn}} \right ] of the AI ​​model is constructed. ,in Represents clusters The core formula for updating the nth eigenvalue in is:

[0038] ;

[0039] in, Indicates The feature matrix of the layer, Indicates The weight matrix of the layer, A represents the adjacency matrix of the graph, represents the query matrix in the self-attention mechanism, represents the key matrix in the self-attention mechanism, represents the activation function;

[0040] S35. Use historical data from cloud servers to optimize the AI ​​model and combine it with a multi-task learning framework to use a multi-objective evolutionary algorithm to optimize the model’s hyperparameters. Perform joint optimization:

[0041] ;

[0042] in, , and represents the weight coefficient, Accuracy represents the accuracy of the AI ​​model, F1-score represents the F1 score of the AI ​​model, and Complexity represents the complexity measurement function of the AI ​​model;

[0043] S36. Save the optimized AI model to the cloud server.

[0044] Optionally, the S4 specifically includes:

[0045] S41, load the optimized AI model received from the cloud server into the computing environment, input the transmitted key feature data into the input layer of the model, and construct the feature matrix X=\left [ {{x}_{ij}} \right ] , Where represents the jth eigenvalue of the i-th sample;

[0046] S42. In the middle layer of the AI ​​model, a multi-layer convolutional neural network is applied for feature extraction, and local patterns of feature data are extracted through convolution operations;

[0047] S43. During the classification process, each target category is weightedly classified by an adaptive weighted classifier. The classifier dynamically adjusts the weight according to the diversity of input features and the imbalance between classes. The calculation formula of the classification probability vector is:

[0048] ;

[0049] in, represents the adaptive weight for the i-th class, Indicates the classification score corresponding to the input feature:

[0050] ;

[0051] in, represents the adjustment parameter, represents the number of samples in the i-th category, represents the mean of the overall classification score;

[0052] S44. After target detection, overlapping detection frames are filtered using a non-maximum suppression algorithm based on a non-linear suppression factor. A non-linear suppression factor is introduced on the basis of the traditional non-maximum suppression algorithm to better retain high-confidence detection frames:

[0053] ;

[0054] in, and Represents two detection boxes, represents the adjustment parameter of the inhibitory factor, The adjustment parameter representing the area difference, and Represents the area of ​​the detection box, and IoU represents the intersection over union ratio of two detection boxes. This factor effectively suppresses overlapping detection boxes with low confidence and large area differences, thereby improving the accuracy of the detection results;

[0055] S45, output the multi-target detection results processed by the non-maximum suppression algorithm of the nonlinear suppression factor, generate the final target recognition and classification results, and annotate the detection results, and define the target category label set as ,in Represents the category label of the i-th detection target;

[0056] S46, transmitting the labeled multi-target detection results to the visualization module of the system.

[0057] Optionally, the S5 specifically includes:

[0058] S51. Take the multi-target detection results as input to construct a multi-target optimization path planning problem. The objective function Including path length, obstacle avoidance safety and resource utilization:

[0059] ;

[0060] in, Indicates the path In time Length, represents the risk level of the rth obstacle, represents the minimum distance between the obstacle and the drone path, represents a small amount that prevents the denominator from being zero, represents the parameter controlling the risk weight, represents the total path planning time, Indicates the total number of obstacles;

[0061] S52. In the particle swarm optimization algorithm, each particle represents a flight path plan, and the position update adopts nonlinear dynamic equations and combines environmental disturbances and historical optimal path information:

[0062] ;

[0063] in, represents particle j at position The state function at represents the potential energy function between the particle position and the global optimal position, represents the adjustment factor, It represents the time step. The formula realizes the dynamic tuning of particle swarm by combining the dynamic change of particle state with the guidance of potential energy function.

[0064] S53. Introduce an adaptive acceleration mechanism based on fuzzy membership function in the particle swarm optimization algorithm to dynamically adjust the acceleration factor according to the complexity and convergence trend of the current path. and :

[0065] ;

[0066] in, Indicates the path The complexity membership function of represents the current convergence rate, represents the reference convergence rate, represents the adjustment parameter, It represents the adjusted acceleration factor. The adaptive mechanism realizes the dynamic adjustment of the acceleration factor by combining the path complexity and convergence characteristics.

[0067] S54: In each iteration, the update mechanism of S52 and the adaptive acceleration mechanism of S53 are used to update the speed and position of the particles until the convergence condition is met or the preset maximum number of iterations is reached, and the optimal flight path solution is obtained. ;

[0068] S55, the optimal flight path plan As a result of the UAV's flight path planning, the UAV's flight path is updated in real time, and the monitoring strategy is dynamically adjusted according to mission requirements and environmental changes.

[0069] Optionally, the S6 specifically includes:

[0070] S61. Construct a graph structure G=(V,E), where V represents the set of sensor nodes, E represents the set of edges between nodes, and each node Represents a sensor data feature, each Representation Node and nodes The correlation between

[0071] S62. Build a graph neural network model to extract global features of the sensor network, update node features, and extract global feature vectors ;

[0072] S63. Extract the global feature vector from the graph neural network Input into a Gaussian process-based adaptive boosting algorithm for fault prediction, using Gaussian process as the basis function for adaptive enhancement:

[0073] ;

[0074] Where X represents the input feature vector, represents the t-th Gaussian process basis function, represents the weight coefficient;

[0075] S64. Kernel function of Gaussian process basis function for:

[0076] ;

[0077] in, represents the signal variance, represents the characteristic length scale matrix, represents the noise variance, represents the Dirac delta function;

[0078] S65. Calculate the fault prediction probability of each sensor node through the GP-Boosting algorithm , and according to the set fault threshold Make a judgment:

[0079] ;

[0080] in, represents the cumulative distribution function of the standard normal distribution, represents the mean of the fault prediction values, represents the standard deviation of the fault prediction value, if , then the decision node There is a potential fault and maintenance recommendations are issued;

[0081] S66. Summarize the fault prediction results of all nodes and generate an overall health status assessment report of the UAV, which includes possible fault nodes, fault types, and recommended maintenance measures.

[0082] The beneficial effects of the present invention are:

[0083] First, the present invention realizes efficient processing and key feature extraction of real-time video and three-dimensional spatial data by introducing advanced sensors such as multispectral cameras and lidars, combined with edge computing technology. Compared with traditional static path planning and task scheduling methods, the intelligent scheduling algorithm combining fuzzy logic control and particle swarm optimization adopted by the present invention can optimize the flight path and monitoring strategy of drones in real time in a dynamic environment, effectively respond to sudden traffic congestion and crowd gathering, and ensure efficient execution of tasks.

[0084] In addition, by introducing graph neural networks and adaptive boosting algorithms, the present invention significantly improves the accuracy and timeliness of equipment failure prediction. The application of graph neural networks enables drones to capture complex correlations between sensor data, while the adaptive boosting algorithm enhances sensitivity to abnormal situations and prediction accuracy by dynamically adjusting the weights of the model. Compared with traditional regular maintenance methods, this predictive maintenance mechanism is more flexible and efficient. It can identify potential equipment failures in advance, reduce the probability of failures, extend the service life of equipment, and reduce maintenance costs and the risk of mission interruption.

[0085] The present invention also solves the problem of large-scale data transmission and storage by combining 5G communication technology and cloud computing, ensuring the data processing capability of drones in complex environments. The introduction of multi-task learning framework and multi-objective evolutionary algorithm enables AI models to be adaptively optimized in complex and changeable mission scenarios, improving the efficiency and accuracy of multi-target detection and classification. Especially when dealing with scenes with dense multi-targets and severe occlusion, the present invention significantly enhances the model's ability to distinguish different targets and detection accuracy by combining graph convolutional neural networks with self-attention mechanisms. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0087] Figure 1 This is a flow chart of a method for detecting drone traffic congestion and personnel gathering based on AI recognition proposed by the present invention;

[0088] Figure 2 A schematic diagram of multi-target detection and classification for an AI model of a method for detecting drone traffic congestion and gathering of people based on AI recognition proposed by the present invention;

[0089] Figure 3 This is a real-time optimization flowchart of the drone flight path and monitoring strategy of the drone traffic congestion and personnel gathering detection method based on AI recognition proposed by the present invention. DETAILED DESCRIPTION

[0090] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0091] refer to Figure 1-3 , a drone traffic congestion and personnel gathering detection method based on AI recognition, comprising the following steps:

[0092] S1. Use drones equipped with multispectral cameras and lidar to obtain real-time video and three-dimensional spatial data of the target area and build a data set;

[0093] S2: Perform edge computing locally on the drone to preliminarily process the data set and extract key features of vehicles, people, and obstacles.

[0094] S3, transmit the extracted key features to the cloud server in real time through the 5G network, combine historical data for large-scale data analysis and modeling, and build and optimize the AI ​​model;

[0095] S4. Use the optimized AI model to perform multi-target detection and classification on the received key features, identify and mark traffic congestion and crowd gathering in the area, and generate multi-target detection results;

[0096] S5. Based on the multi-target detection results, the intelligent scheduling algorithm of fuzzy logic control and particle swarm optimization is used to optimize the flight path and monitoring strategy of the drone in real time, and global scheduling optimization is performed in combination with the status of airport resources and mission requirements;

[0097] S6. Through the AI-driven predictive maintenance module, the operating status of the drone is continuously monitored, and the graph neural network and adaptive boosting algorithm are used to identify possible equipment failures in advance and issue maintenance recommendations;

[0098] S7, combined with the intelligent warning system, monitors the flight status and recognition results of the drone in real time, triggers an alarm when an abnormal situation is detected, and notifies the operator to intervene;

[0099] S8. Display the recognition results and analysis reports through a visual interface, and provide task progress tracking and adjustment functions.

[0100] In this implementation, S2 specifically includes:

[0101] S21. Preprocess the data sets obtained from the multispectral camera and lidar through the edge computing unit carried by the drone, including data format conversion, noise filtering and redundant data removal;

[0102] S22, in the preprocessed data set, using a weighted average filter to smooth the video data and the three-dimensional space data to reduce high-frequency noise and sudden abnormal values ​​in the data to ensure the stability and reliability of the data;

[0103] S23, through the feature extraction algorithm based on gray level co-occurrence matrix, the texture analysis is performed on the pre-processed video data to extract the surface texture features of vehicles, people and obstacles ,in represents the i-th texture feature extracted from the image;

[0104] S24. Use the edge detection algorithm based on the Sobel operator to detect the edge of the target in the video data, extract the contour features of the target, and extract the straight line and circle features through the Hough transform ,in represents the i-th feature extracted from edge detection;

[0105] S25. For lidar data, an improved RANSAC algorithm combined with weighted least squares method is used to perform plane fitting to extract the shape features of vehicles, people and obstacles from the three-dimensional space data. ,in Represents the i-th shape feature extracted from the point cloud data, and separates different targets through point cloud segmentation technology to generate a 3D contour point cloud for each target:

[0106] ;

[0107] Where P represents the point cloud dataset, a, b, c and d represent the coefficients of the fitted plane, Represents the weight of a point (x, y, z):

[0108] ;

[0109] in, and represents the adjustment parameter, Represents the local density of a point (x, y, z) in a point cloud.

[0110] In this implementation, S3 specifically includes:

[0111] S31, transmit key feature data extracted from the local edge computing unit of the drone to the cloud server in real time through the 5G network;

[0112] S32. On the cloud server, the received feature data is stored in real time, and the data sharding technology is used to manage large-scale data. The data sharding matrix S=\left [ {{s}_{ij}} \right ] is defined. ,in Indicates the jth storage location of the i-th data block after sharding, and calculates the storage index of the data block through the hash function H:

[0113] ;

[0114] in, Represents a data block The hash value of , N represents the total number of storage nodes;

[0115] S33. On the cloud server, the received feature data is clustered and analyzed in combination with historical data. The density-based clustering algorithm is used to perform preliminary clustering of vehicles, personnel and obstacles. The clusters after clustering are defined as , where k represents the number of the cluster:

[0116] ;

[0117] in, Indicates is the center and the radius is Neighborhood, MinPts represents the minimum number of samples in the cluster;

[0118] S34. Based on the clustering results, the feature data is input into an AI model based on a graph convolutional neural network combined with a self-attention mechanism for large-scale data modeling, and the input matrix Z=\left [ {{z}_{mn}} \right ] of the AI ​​model is constructed. ,in Represents clusters The core formula for updating the nth eigenvalue in is:

[0119] ;

[0120] in, Indicates The feature matrix of the layer, Indicates The weight matrix of the layer, A represents the adjacency matrix of the graph, represents the query matrix in the self-attention mechanism, represents the key matrix in the self-attention mechanism, represents the activation function;

[0121] S35. Use historical data from cloud servers to optimize the AI ​​model and combine it with a multi-task learning framework to use a multi-objective evolutionary algorithm to optimize the model’s hyperparameters. Perform joint optimization:

[0122] ;

[0123] in, , and represents the weight coefficient, Accuracy represents the accuracy of the AI ​​model, F1-score represents the F1 score of the AI ​​model, and Complexity represents the complexity measurement function of the AI ​​model;

[0124] S36. Save the optimized AI model to the cloud server.

[0125] In this implementation, S4 specifically includes:

[0126] S41, load the optimized AI model received from the cloud server into the computing environment, input the transmitted key feature data into the input layer of the model, and construct the feature matrix X=\left [ {{x}_{ij}} \right ] , Where represents the jth eigenvalue of the i-th sample;

[0127] S42. In the middle layer of the AI ​​model, a multi-layer convolutional neural network is applied for feature extraction, and local patterns of feature data are extracted through convolution operations;

[0128] S43. During the classification process, each target category is weightedly classified by an adaptive weighted classifier. The classifier dynamically adjusts the weight according to the diversity of input features and the imbalance between classes. The calculation formula of the classification probability vector is:

[0129] ;

[0130] in, represents the adaptive weight for the i-th class, Indicates the classification score corresponding to the input feature:

[0131] ;

[0132] in, represents the adjustment parameter, represents the number of samples in the i-th category, represents the mean of the overall classification score;

[0133] S44. After target detection, overlapping detection frames are filtered using a non-maximum suppression algorithm based on a non-linear suppression factor. A non-linear suppression factor is introduced on the basis of the traditional non-maximum suppression algorithm to better retain high-confidence detection frames:

[0134] ;

[0135] in, and Represents two detection boxes, represents the adjustment parameter of the inhibitory factor, The adjustment parameter representing the area difference, and Represents the area of ​​the detection box, and IoU represents the intersection over union ratio of two detection boxes. This factor effectively suppresses overlapping detection boxes with low confidence and large area differences, thereby improving the accuracy of the detection results;

[0136] S45, output the multi-target detection results processed by the non-maximum suppression algorithm of the nonlinear suppression factor, generate the final target recognition and classification results, and annotate the detection results, and define the target category label set as ,in Represents the category label of the i-th detection target;

[0137] S46, transmitting the labeled multi-target detection results to the visualization module of the system.

[0138] In this implementation manner, S5 specifically includes:

[0139] S51. Take the multi-target detection results as input to construct a multi-target optimization path planning problem. The objective function Including path length, obstacle avoidance safety and resource utilization:

[0140] ;

[0141] in, Indicates the path In time Length, represents the risk level of the rth obstacle, represents the minimum distance between the obstacle and the drone path, represents a small amount that prevents the denominator from being zero, represents the parameter controlling the risk weight, represents the total path planning time, Indicates the total number of obstacles;

[0142] S52. In the particle swarm optimization algorithm, each particle represents a flight path plan, and the position update adopts nonlinear dynamic equations and combines environmental disturbances and historical optimal path information:

[0143] ;

[0144] in, represents particle j at position The state function at represents the potential energy function between the particle position and the global optimal position, represents the adjustment factor, It represents the time step. The formula realizes the dynamic tuning of particle swarm by combining the dynamic change of particle state with the guidance of potential energy function.

[0145] S53. Introduce an adaptive acceleration mechanism based on fuzzy membership function in the particle swarm optimization algorithm to dynamically adjust the acceleration factor according to the complexity and convergence trend of the current path. and :

[0146] ;

[0147] in, Indicates the path The complexity membership function of represents the current convergence rate, represents the reference convergence rate, represents the adjustment parameter, It represents the adjusted acceleration factor. The adaptive mechanism realizes the dynamic adjustment of the acceleration factor by combining the path complexity and convergence characteristics.

[0148] S54: In each iteration, the update mechanism of S52 and the adaptive acceleration mechanism of S53 are used to update the speed and position of the particles until the convergence condition is met or the preset maximum number of iterations is reached, and the optimal flight path solution is obtained. ;

[0149] S55, the optimal flight path plan As a result of the UAV's flight path planning, the UAV's flight path is updated in real time, and the monitoring strategy is dynamically adjusted according to mission requirements and environmental changes.

[0150] In this implementation manner, S6 specifically includes:

[0151] S61. Construct a graph structure G=(V,E), where V represents the set of sensor nodes, E represents the set of edges between nodes, and each node Represents a sensor data feature, each Representation Node and nodes The correlation between

[0152] S62. Build a graph neural network model to extract global features of the sensor network, update node features, and extract global feature vectors ;

[0153] S63. Extract the global feature vector from the graph neural network Input into a Gaussian process-based adaptive boosting algorithm for fault prediction, using Gaussian process as the basis function for adaptive enhancement:

[0154] ;

[0155] Where X represents the input feature vector, represents the t-th Gaussian process basis function, represents the weight coefficient;

[0156] S64. Kernel function of Gaussian process basis function for:

[0157] ;

[0158] in, represents the signal variance, represents the characteristic length scale matrix, represents the noise variance, represents the Dirac delta function;

[0159] S65. Calculate the fault prediction probability of each sensor node through the GP-Boosting algorithm , and according to the set fault threshold Make a judgment:

[0160] ;

[0161] in, represents the cumulative distribution function of the standard normal distribution, represents the mean of the fault prediction values, represents the standard deviation of the fault prediction value, if , then the decision node There is a potential fault and maintenance recommendations are issued;

[0162] S66. Summarize the fault prediction results of all nodes and generate an overall health status assessment report of the UAV, which includes possible fault nodes, fault types, and recommended maintenance measures.

[0163] Embodiment 1:

[0164] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a busy road section in the city center. This road section has many intersections, heavy traffic volume, and is prone to traffic jams during rush hours. At the same time, due to the concentration of commercial and office spaces in the area, large-scale gatherings of people often occur, which poses a great challenge to traffic management. Traditional traffic monitoring systems are difficult to accurately obtain global information in a short period of time due to limited viewing angles and untimely responses, making it difficult to effectively manage emergencies, affecting the smoothness of traffic flow and the safety of personnel.

[0165] During the morning rush hour, drones equipped with multispectral cameras and lidar equipment flew 100 meters above the road, covering an area of ​​about 5 square kilometers. The drones acquired real-time video and three-dimensional spatial data on the road, and performed preliminary processing locally to extract key features of vehicles and people. These feature data were transmitted to the cloud server in real time via the 5G network. Combined with historical data, the cloud server used deep learning models for large-scale analysis and modeling.

[0166] Through multi-target detection and classification of cloud-based AI models, the drone identified multiple traffic jams and areas where people gathered. At this time, 10 kilometers away from the area, another group of drones was monitoring the traffic flow and crowd dynamics in the area. Through fuzzy logic control and particle swarm optimization algorithms, the system can dynamically adjust the flight path of the drone based on real-time detection results, and optimize monitoring strategies for different levels of traffic congestion and density of people gathering. In high-risk areas, drones are instructed to lower their flight altitudes to provide more detailed images and data support, while in low-risk areas, the coverage frequency is reduced by adjusting the path to save resources.

[0167] ;

[0168] Table 1 compares the key data and effect indicators of the conventional technology and the present invention in drone traffic congestion and crowd gathering detection in detail. It can be seen from the comparison that the present invention has significant advantages in many aspects.

[0169] First, in terms of the detection of traffic jams and crowded areas, the present invention significantly increases the number of detections. Specifically, the present invention can detect 10 traffic jams and 3 crowded areas, which are 67% and 200% higher than the traditional technology, respectively. This shows that the present invention has greatly improved the detection accuracy and coverage in complex environments.

[0170] Secondly, the present invention has significant advantages in data processing capabilities. By utilizing 5G communication and cloud computing technology, the present invention can process 10TB of data, while traditional technology can only process 2TB of data, increasing the data processing capacity by 400%. This improvement in capabilities enables the present invention to better cope with the real-time data processing needs in large-scale and complex environments.

[0171] In terms of the application of AI models, the present invention not only achieves multiple dynamic optimizations, but also significantly improves the detection accuracy from 75% of traditional technologies to 90% through a multi-task learning framework and a multi-objective evolutionary algorithm, while reducing the false alarm rate from 20% to 5%, an increase of 15% and a decrease of 75% respectively. These improvements significantly reduce unnecessary interventions caused by false alarms and improve the reliability of the overall system.

[0172] In terms of the response speed and path planning of drones, the present invention realizes the real-time optimization of drone flight paths and monitoring strategies through fuzzy logic control and particle swarm optimization algorithm. The results show that the relief time of congestion points is shortened from the traditional 30 minutes to 15 minutes, and the response time is reduced from 20 minutes to 5 minutes, which are shortened by 50% and 75% respectively. This rapid response capability greatly improves the efficiency of the system.

[0173] In addition, the present invention also realizes predictive maintenance of drones through graph neural networks and adaptive boosting algorithms. Data shows that the equipment failure rate in traditional technologies is 10%, while the present invention did not have any failures during the mission execution, and there were 2 successful cases of predictive maintenance, which completely avoided mission interruptions caused by equipment failures. This improvement not only improves the reliability of drones, but also reduces maintenance costs.

[0174] Finally, the present invention also demonstrated its effectiveness in handling crowd gathering events. Through accurate identification and real-time monitoring, the present invention successfully prevented a potential safety accident, further verifying its superiority in complex dynamic environments.

[0175] In summary, it can be clearly seen from Table 1 that the performance of the present invention in drone traffic congestion and personnel gathering detection far exceeds that of traditional technologies, especially in terms of detection accuracy, response speed, data processing capability and equipment reliability, showing great improvements and innovations, and providing strong technical support for drone applications in complex environments.

[0176] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for detecting drone traffic congestion and personnel gathering based on AI recognition, characterized in that: The steps include: S1. Use drones equipped with multispectral cameras and lidar to obtain real-time video and three-dimensional spatial data of the target area and build a data set; S2: Perform edge computing locally on the drone to preliminarily process the data set and extract key features of vehicles, people, and obstacles. S3, transmit the extracted key features to the cloud server in real time through the 5G network, combine historical data for large-scale data analysis and modeling, and build and optimize the AI ​​model; S4. Use the optimized AI model to perform multi-target detection and classification on the received key features, identify and mark traffic congestion and crowd gathering in the area, and generate multi-target detection results; S5. Based on the multi-target detection results, the intelligent scheduling algorithm of fuzzy logic control and particle swarm optimization is used to optimize the flight path and monitoring strategy of the drone in real time, and global scheduling optimization is performed in combination with the status of airport resources and mission requirements; S6. Through the AI-driven predictive maintenance module, the operating status of the drone is continuously monitored, and the graph neural network and adaptive boosting algorithm are used to identify possible equipment failures in advance and issue maintenance recommendations; S7, combined with the intelligent warning system, monitors the flight status and recognition results of the drone in real time, triggers an alarm when an abnormal situation is detected, and notifies the operator to intervene; S8. Display the recognition results and analysis reports through a visual interface, and provide task progress tracking and adjustment functions.

2. According to the method of claim 1, the method is characterized in that: The S2 specifically includes: S21. Preprocess the data sets obtained from the multispectral camera and lidar through the edge computing unit carried by the drone, including data format conversion, noise filtering and redundant data removal; S22, in the preprocessed data set, using a weighted average filter to smooth the video data and the three-dimensional space data; S23, using a feature extraction algorithm based on gray-level co-occurrence matrix, texture analysis is performed on the pre-processed video data to extract surface texture features of vehicles, personnel and obstacles. texture =[f1,f2,…,f n ], where f i represents the i-th texture feature extracted from the image; S24, using the edge detection algorithm based on the Sobel operator, edge detection is performed on the target in the video data, the contour features of the target are extracted, and the straight line and circle features are extracted through the Hough transform. edge =[e1,e2,…,e m ], where e i represents the i-th feature extracted from edge detection; S25. For the lidar data, an improved RANSAC algorithm combined with weighted least squares method is used to perform plane fitting to extract the shape features of vehicles, people and obstacles from the three-dimensional space data. shape =[s1,s2,…,s k ], where s i Represents the i-th shape feature extracted from the point cloud data, and separates different targets through point cloud segmentation technology to generate a 3D contour point cloud for each target: Where P represents the point cloud dataset, a, b, c and d represent the coefficients of the fitted plane, and w x,y,z Represents the weight of a point (x, y, z): Among them, α and β represent adjustment parameters, and density(x,y,z) represents the local density of point (x,y,z) in the point cloud.

3. The method for detecting drone traffic congestion and personnel gathering based on AI recognition according to claim 1 is characterized in that: The S3 specifically includes: S31, transmit key feature data extracted from the local edge computing unit of the drone to the cloud server in real time through the 5G network; S32. On the cloud server, the received feature data is stored in real time, and the large-scale data is managed using data sharding technology. The data sharding matrix S = [s ij ], where s ij Indicates the jth storage location of the i-th data block after sharding, and calculates the storage index of the data block through the hash function H: s ij =H(x ij )modN; Among them, H(x ij ) represents the data block x ij The hash value of , N represents the total number of storage nodes; S33. On the cloud server, perform cluster analysis on the received feature data in combination with historical data, use density-based clustering algorithm to perform preliminary clustering on vehicles, personnel and obstacles, and define the clusters after clustering as C. k , where k represents the number of the cluster: C k ={x i ∣MinPts≤|N(x i ,∈)|}; Among them, N(x i ,∈) represents x i is the neighborhood with the center and radius ∈, MinPts represents the minimum number of samples in the cluster; S34. According to the clustering results, the feature data is input into an AI model based on a graph convolutional neural network combined with a self-attention mechanism for large-scale data modeling, and the input matrix Z of the AI ​​model is constructed as follows: mn ], where z mn Represents cluster C m The core formula for updating the nth eigenvalue in is: H (l+1) =σ(A·H (l) ·W (l) +H (l) ·Q (l) ·H (l)T ·K (l) ); Among them, H (l) represents the feature matrix of the lth layer, W (l) represents the weight matrix of the lth layer, A represents the adjacency matrix of the graph, Q (l) represents the query matrix in the self-attention mechanism, K (l) represents the key matrix in the self-attention mechanism, σ represents the activation function; S35. Use the historical data of the cloud server to optimize the AI ​​model training, and combine it with the multi-task learning framework to use a multi-objective evolutionary algorithm to jointly optimize the model's hyperparameter λ: Among them, ω1, ω2 and ω3 represent weight coefficients, Accuracy represents the accuracy of the AI ​​model, F1-score represents the F1 score of the AI ​​model, and Complexity represents the complexity measurement function of the AI ​​model; S36. Save the optimized AI model to the cloud server.

4. The method for detecting drone traffic congestion and personnel gathering based on AI recognition according to claim 1 is characterized in that: The S4 specifically includes: S41, load the optimized AI model received from the cloud server into the computing environment, and input the transmitted key feature data into the input layer of the model to construct the feature matrix X = [x ij ], where x ij represents the jth eigenvalue of the i-th sample; S42. In the middle layer of the AI ​​model, a multi-layer convolutional neural network is applied for feature extraction, and local patterns of feature data are extracted through convolution operations; S43. During the classification process, each target category is weightedly classified by an adaptive weighted classifier. The classifier dynamically adjusts the weight according to the diversity of input features and the imbalance between classes. The calculation formula of the classification probability vector is: Among them, w i represents the adaptive weight for the i-th class, z i Indicates the classification score corresponding to the input feature: Among them, α1 represents the adjustment parameter, M i represents the number of samples in the i-th category, and μ represents the mean of the overall classification score; S44. After target detection, overlapping detection frames are filtered using a non-maximum suppression algorithm based on a nonlinear suppression factor: Among them, B i and B j represents two detection boxes, γ1 represents the adjustment parameter of the suppression factor, λ1 represents the adjustment parameter of the area difference, S i and S j Represents the area of ​​the detection box, and IoU represents the intersection-over-union ratio of two detection boxes; S45, output the multi-target detection results processed by the non-maximum suppression algorithm of the nonlinear suppression factor, generate the final target recognition and classification results, and annotate the detection results, and define the target category label set as L = {L1, L2, ..., L k }, where L i Represents the category label of the i-th detection target; S46, transmitting the labeled multi-target detection results to the visualization module of the system.

5. The method for detecting drone traffic congestion and personnel gathering based on AI recognition according to claim 1 is characterized in that: The S5 specifically includes: S51. Take the multi-target detection results as input and construct a multi-target optimization path planning problem. The objective function G(Θ) includes path length, obstacle avoidance safety, and resource utilization: Where L(Θ(τ)) represents the length of the path Θ(τ) at time τ, Γ r represents the risk level of the rth obstacle, Ψ r represents the minimum distance between the obstacle and the drone path, ∈ represents a small amount to prevent the denominator from being zero, ζ represents the parameter for controlling the risk weight, Λ represents the total path planning time, and M1 represents the total number of obstacles; S52. In the particle swarm optimization algorithm, each particle represents a flight path plan, and the position update adopts nonlinear dynamic equations and combines environmental disturbances and historical optimal path information: in, represents particle j at position The state function at represents the potential energy function between the particle position and the global optimal position, ν represents the adjustment factor, and Δτ represents the time step; S53. An adaptive acceleration mechanism based on fuzzy membership function is introduced into the particle swarm optimization algorithm to dynamically adjust the acceleration factors ξ1 and ξ2 according to the complexity and convergence trend of the current path: Among them, μ complexity (Θ) represents the complexity membership function of path Θ, R(τ) represents the current convergence rate, R ref represents the reference convergence rate, κ represents the adjustment parameter, ξ k represents the adjusted acceleration factor; S54, in each iteration, using the update mechanism of S52 and the adaptive acceleration mechanism of S53, update the speed and position of the particle until the convergence condition is met or the preset maximum number of iterations is reached, and the optimal flight path solution Θ is obtained. opt ; S55, the optimal flight path plan Θ opt As a result of the UAV's flight path planning, the UAV's flight path is updated in real time, and the monitoring strategy is dynamically adjusted according to mission requirements and environmental changes.

6. The method for detecting drone traffic congestion and personnel gathering based on AI recognition according to claim 1 is characterized in that: The S6 specifically includes: S61. Construct a graph structure G = (V, E), where V represents the set of sensor nodes, E represents the set of edges between nodes, and each node v i ∈V represents a sensor data feature, and each edge e ij ∈E represents node v i and node v j The correlation between S62. Build a graph neural network model to extract global features of the sensor network, update node features, and extract global feature vectors S63. Extract the global feature vector from the graph neural network Input into a Gaussian process-based adaptive boosting algorithm for fault prediction, using Gaussian process as the basis function for adaptive enhancement: Among them, x represents the input feature vector, G t (x) represents the t-th Gaussian process basis function, β t represents the weight coefficient; S64, the kernel function K(x,x) of the Gaussian process basis function ′ )for: in, represents the signal variance, Λ represents the characteristic length scale matrix, represents the noise variance, δ(xx ′ ) represents the Diracdelta function; S65. Calculate the fault prediction probability P of each sensor node through the GP-Boosting algorithm f (v i ), and according to the set fault threshold θ f Make a judgment: Where Φ represents the cumulative distribution function of the standard normal distribution, μ f represents the mean of the fault prediction value, σ f represents the standard deviation of the fault prediction value. If P f (v i )>θ f , then determine node v i There is a potential fault and maintenance recommendations are issued; S66. Summarize the fault prediction results of all nodes and generate an overall health status assessment report of the UAV, which includes possible fault nodes, fault types, and recommended maintenance measures.

Images