Intersection state perception and traffic safety evaluation method based on unmanned aerial vehicle and large model

Abstract

The invention discloses an intersection state sensing and traffic safety evaluation method based on an unmanned aerial vehicle and a large model, and belongs to the field of traffic safety evaluation. Comprising the following steps: constructing a target detection and tracking model based on an unmanned aerial vehicle visual angle, collecting intersection state and traffic flow data by using an unmanned aerial vehicle aerial video, counting traffic conflict data, constructing intersection potential conflict indexes, constructing a comprehensive grade evaluation model, improving an initial model by using a large model, and outputting an intersection comprehensive safety evaluation grade. Using an improved RT-DETR algorithm and a Deepsort model to improve the detection precision of the traffic participants and carrying out training; the intersection state and traffic flow data are extracted, and the overall risk quantification of the intersection is realized through a plurality of potential conflict indexes and by adjusting the weight of each index through a large model. According to the method, dynamic regulation and control are carried out on the safety index weight in combination with a large model, comprehensive evaluation of potential risks, abnormal behaviors and the overall safety level of the intersection is achieved, and the method has important significance.

Description

Technical Field

[0001] This invention relates to the fields of computer vision technology and traffic safety technology, specifically to a method for intersection state perception and traffic safety evaluation based on unmanned aerial vehicles (UAVs) and large-scale models. Background Technology

[0002] With the accelerating pace of urbanization, urban transportation networks are becoming increasingly complex. Road intersections, as key nodes in these networks, are characterized by high traffic volume and complex routes, making them high-risk areas for traffic accidents. The mixing of pedestrians, non-motorized vehicles, and motorized vehicles presents particularly significant challenges. Improving traffic safety at intersections is crucial for ensuring the safety of all road users and for perfecting the traffic management system.

[0003] Traditional urban traffic safety assessment methods primarily rely on fixed monitoring equipment or statistical analysis of historical traffic accident data. These methods suffer from limitations such as limited observation perspectives, delayed data updates, and insufficient spatial coverage, making it difficult to comprehensively and in real-time reflect the dynamic traffic conditions at intersections. Furthermore, fixed cameras are often limited by their installation location and angle, resulting in insufficient ability to capture traffic behavior at intersection edges and in perpendicular directions, thus affecting the accuracy of the assessment.

[0004] In recent years, drones have been increasingly adopted in urban traffic management and road condition monitoring due to their advantages of mobility, wide field of view, and convenient deployment. In the field of traffic safety, drone video data provides a new approach to acquiring macroscopic traffic flow information and fine-grained traffic behavior characteristics. Compared to ground-based monitoring systems, the altitude advantage of drones allows them to have a bird's-eye view, covering the entire intersection area and capturing the spatial distribution and movement trajectories of various traffic participants, making them particularly suitable for traffic behavior perception and conflict analysis in complex road conditions. However, the rich and heterogeneous data acquired by drones, including images and trajectories, contains more complex behavioral interaction semantics and potential risk characteristics than traditional observation data, placing higher demands on the feature extraction capabilities and intelligence level of subsequent data analysis models.

[0005] Currently, some studies have attempted to utilize image data collected by drones for traffic safety assessment, but these studies mostly focus on single-dimensional indicators (such as traffic flow and speed), failing to fully utilize the global spatiotemporal information from the drone's aerial perspective. More importantly, in urban intersection scenarios, there are complex nonlinear interactions among various traffic participants (pedestrians, motor vehicles, and non-motorized vehicles), resulting in significant dynamics and uncertainties in risk sources. Existing methods are insufficient in deeply understanding the semantics and dynamic evolution of these complex interactions, as well as in conducting comprehensive quantitative assessments of multi-dimensional risks. In particular, they lack the technical means to integrate multimodal information and possess strong reasoning and learning capabilities for end-to-end intelligent analysis and decision-making.

[0006] To address the aforementioned issues, this paper proposes a dynamic evaluation method for urban intersection traffic safety that integrates UAV stereo perception and large-scale model technology. This method aims to overcome the bottlenecks of traditional models in understanding and analyzing complex scenarios. Leveraging the powerful multimodal information processing, pattern recognition, knowledge fusion, and model optimization capabilities of large-scale models, it deeply mines the global spatiotemporal information collected by UAVs, enabling intelligent identification, quantification, and comprehensive evaluation of multidimensional interactive risks in mixed traffic flows. This significantly improves the level of refined urban road management and the real-time early warning capability for traffic risks. Summary of the Invention

[0007] This invention aims to address at least one of the technical problems existing in the prior art. To this end, one objective of this invention is to propose a method for intersection state perception and traffic safety evaluation based on unmanned aerial vehicles (UAVs) and large-scale models. This method uses a deep learning model to extract the dynamic characteristics of each traffic participant at the intersection and proposes traffic safety evaluation indicators that consider multi-dimensional interactions, effectively reflecting multiple aspects such as potential risks, abnormal behaviors, violations, and traffic pressure at the intersection. Finally, a grade evaluation model is constructed based on the above indicators and optimized using a large-scale model, mapping the intersection safety status to a grade output and providing an overall safety level determination. This method does not rely on historical accident data and is applicable to real-time monitoring and evaluation of urban road traffic safety.

[0008] To address the aforementioned problems, this invention provides a method for intersection state perception and traffic safety evaluation based on unmanned aerial vehicles (UAVs) and large-scale models, characterized in that the method includes: S1: Construct a traffic target detection and tracking model for intersections based on the perspective of drones; S2: Take off the drone from the intersection to be inspected and fly over the intersection to capture video of traffic operation at the intersection; S3: Input the collected intersection traffic operation video into the trained model, collect intersection status and traffic flow data, and identify and extract detailed movement information of traffic participants; S4: Collect traffic conflict data and construct potential conflict indicators for intersections; S5: Construct a comprehensive rating model and use the larger model to improve the initial model; S6: Calculate using the improved model and output the comprehensive safety evaluation level of the intersection.

[0009] Preferably, in step S1, an aerial view of urban intersection image data is collected using a drone and labeled using the LabelImg tool to build a basic dataset. Finally, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0010] Preferably, in step S1, an improved intersection traffic participant target detection model based on RT-DETR is constructed, and the improvement is achieved by improving the Backbone part and the AIFI part.

[0011] Preferably, in step S1, the improvement to the Backbone section is as follows: the Block module in the original RT-DETR backbone network is replaced with an M-Block module. The introduced M-Block module follows the design concept of the MetaFormer structure, and its definition can be summarized as follows: Where X is the input feature map, F MSIM For spatial interaction modules, F FEU This is a feature enhancement unit.

[0012] Preferably, in step S1, the improvement to the AIFI part is to replace the FNN module of the original RT-DETR backbone network with an SP-FFN module. The definition of the introduced SP-FFN module can be summarized as follows: Where X is the feature representation, Q is the frequency domain quantization matrix, F is the real two-dimensional Fourier transform, and Fold represents the concatenation operation. Represents the GEGLU gating mechanism. This is the Hadamard product in element-wise multiplication.

[0013] Preferably, in step S2, the drone takes off and hovers at an altitude of 80-120 meters, the drone captures a 15-minute video of the intersection, the video resolution is 1920*1080, the frame rate is 60 frames per second, and the format is standard mp4.

[0014] Preferably, in step S3, the traffic participants collected include three categories: pedestrians, non-motorized vehicles, and motorized vehicles. The traffic flow data includes target ID, target category, target detection box coordinates, target detection box height and width, and the frame in which the target detection box is located. The detailed motion information of the traffic participants includes speed, acceleration, and direction of motion.

[0015] Preferably, in step S4, the potential conflict indexes at the intersection include the intersection interaction conflict potential index, the intersection movement disorder index, the intersection abnormal driving behavior index, the intersection congestion pressure index, and the intersection violation interference index. In step S4, Set to 0.2, directional sector number to 8, a0 to 2.5, and T to 60s.

[0016] Preferably, in step S5, the user inputs intersection information into the large model by using a large model to improve the initial model, and the large model infers the weight settings and thought process in the grade evaluation model. Intersection information includes intersection geometric attributes, intersection functional attributes, intersection traffic data, and intersection conflict data; The weights output by the large model include w1, w2, w3, w4, and w5.

[0017] Preferably, in step S6, a weighted strategy is constructed using expert experience.

[0018] The advantages of this invention compared to the prior art are: 1. Accurate perception of intersection operation status based on "UAV + target detection": Using the high-altitude top-down view of UAVs to obtain global spatiotemporal information of intersections, combined with an improved target detection and tracking model, the accuracy of identification and trajectory tracking of small targets such as pedestrians and non-motorized vehicles in complex environments (such as evening and heavy fog) is significantly improved. This provides a reliable data foundation for subsequent traffic conflict data statistics and index calculation, effectively making up for the shortcomings of traditional methods that rely on historical accident data statistics, such as poor real-time performance and large sample bias.

[0019] 2. Constructing a comprehensive safety evaluation index system for intersections that considers multi-dimensional interactions: To break away from the reliance of traditional methods on single indicators (such as accident rate) or simple conflict indicators (such as TTC and PET), this method innovatively proposes a multi-dimensional safety evaluation index system for intersections that covers potential conflicts, disordered movement, abnormal behavior, congestion pressure, and illegal interference. This system quantifies the complex nonlinear dynamic interaction behaviors between pedestrians, non-motorized vehicles, and motorized vehicles, accurately captures potential conflict patterns, and thus more comprehensively reflects the potential risks and overall safety level of intersections.

[0020] 3. Utilizing large-scale models to achieve dynamic optimization of indicator weights that consider personalized needs: Innovatively, large-scale models (LLM) are introduced into the comprehensive safety level evaluation. By leveraging their powerful knowledge fusion and semantic understanding capabilities, the weight priority of various safety indicators in the final evaluation is intelligently adjusted according to the actual traffic conditions of the intersection and user needs. This accurately captures the dynamic correlation and risk evolution patterns among various indicators in complex traffic scenarios, achieving multi-dimensional comprehensive quantitative assessment of risks. Compared with the traditional method of manually pre-setting weights, this significantly improves the accuracy, adaptability, and interpretability of the evaluation results. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a structural diagram of the improved M-Block of the present invention; Figure 3 This is a structural diagram of the improved SP-FFN of the present invention. Detailed Implementation

[0023] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0024] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0025] The present invention will now be described in further detail with reference to the accompanying drawings.

[0026] The present invention provides a method for intersection state perception and traffic safety evaluation based on unmanned aerial vehicles (UAVs) and large-scale models, comprising the following steps: S1: Construct a traffic target detection and tracking model for intersections based on the perspective of drones; S2: Take off the drone from the intersection to be inspected and fly over the intersection to capture video of traffic operation at the intersection; S3: Input the collected intersection traffic operation video into the trained model, collect intersection status and traffic flow data, and identify and extract detailed movement information of traffic participants; S4: Collect traffic conflict data and construct potential conflict indicators for intersections; S5: Construct a comprehensive rating model and use the larger model to improve the initial model; S6: Calculate using the improved model and output the comprehensive safety evaluation level of the intersection; Preferably, the steps for building the target detection and tracking model are as follows: using a drone to collect aerial video of the intersection and construct a traffic participant dataset; using an improved RT-DETR algorithm and DeepSort model to construct a neural network model; inputting the constructed traffic participant dataset into the model for training; and optimizing the model based on the training results to obtain training weights.

[0027] Preferably, collecting intersection status and traffic flow data includes the following steps: S31: Use an object detection model to obtain the object detection bounding boxes and category labels of traffic participants in each frame.

[0028] S32: Obtain trajectory data of traffic participants in the intersection video based on the target detection and tracking results. S33: Calculate the motion status information (speed, acceleration, direction of motion) of each traffic participant using trajectory data. Preferably, the steps for calculating the potential conflict index at an intersection are as follows: S41: Extract potential conflict objects within the intersection and classify them according to the type of traffic conflict participants.

[0029] S42: Calculate the potential index of intersection interaction conflicts.

[0030] S43: Calculate the intersection motion disorder index.

[0031] S44: Calculate the abnormal driving behavior index at the intersection.

[0032] S45: Calculate the intersection congestion pressure index.

[0033] S46: Calculate the intersection violation interference index.

[0034] Preferably, the steps for calculating the comprehensive safety evaluation level of an intersection are as follows: standardize the safety indices of five intersections; define the calculation formula for the comprehensive safety index of road intersections; set the normalization weights for each index; establish safety level classification rules; calculate the comprehensive safety score and output the level.

[0035] To more clearly illustrate the specific embodiments of the present invention, an example is provided below: This invention relates to a method for safety assessment of urban intersections based on the perspective of unmanned aerial vehicles (UAVs), the flowchart of which is shown below. Figure 1 As shown, it includes: Step 1: Construct a target detection and tracking model based on the UAV's perspective. The RT-DETR backbone network adopts a four-layer (basic residual block, Basic Block) structure. To enhance the feature extraction capability for targets of different sizes from the UAV's perspective, this invention has improved its backbone network and AIFI module.

[0036] In the backbone network, an enhanced structural block M-Block is proposed, which integrates frequency domain enhancement mechanism and multi-branch spatial awareness mechanism. The improved M-Block module follows the Metafromer structure and consists of two parts: Multi-Scale Spatial Interaction Module (MSIM) and Feature Enhancement Unit (FEU). Its overall process can be represented by the following function transformation process: Where X∈RB×C×H×W is the input feature map, Y is the output feature map, N1 and N2 represent normalization operations, β, Y∈R1×C×1×1 are learnable scaling factors used to adjust the intensity of the residual response, FMSIM is the spatial interaction module, and FFEU is the feature enhancement unit.

[0037] Let the input feature map be X∈RB×C×H×W, where C represents the number of channels, and H and W are the height and width of the feature map, respectively. First, the input features are subjected to two-dimensional normalization to alleviate gradient explosion and vanishing phenomena in deep networks and improve training stability. in and These represent the mean and variance of the channel dimension, respectively. is the numerical stability constant.

[0038] The normalized features first enter the Multi-Scale Spatial Interaction Module (MSIM) to extract salient region features from the spatial dimension, specifically including the following steps: (1). Preprocessing and multi-branch convolution: First, depthwise separable convolutions are enabled to reduce the number of parameters, and channel expansion is performed on the preprocessed feature map: Where r is the channel expansion factor, set to 2 to better enhance the nonlinear representation capability of features.

[0039] Then, multi-branch convolution is performed on X3 to perceive spatial features at different scales: Each branch uses a different expansion rate di to model different receptive field ranges.

[0040] (2). Gated nonlinear activation (SimpleGate) and simplified channel attention (SCA): SimpleGate is a lightweight nonlinear gating mechanism that enables adjustment of spatial feature saliency, facilitating the extraction of meaningful spatial information by the model. Specifically, it performs element-wise multiplication after splitting the feature channels in half. in This represents the Hadamard product (element-by-element multiplication).

[0041] Then, a simplified channel attention module is constructed using global average pooling and 1×1 convolution, with the following formula: After 1×1 compressed convolution, the output is identical to the original input channel: The final feature output after spatial augmentation is generated using residual connections and learnable scaled silver beta: Furthermore, to address the issue that the spatial domain cannot efficiently represent edge and global structural information, this project uses a Feature Enhancement Unit (FEU) as an additional attention mechanism to improve the model's robustness to changes in image quality. The specific steps are as follows: (1). Preprocessing and Feature Enhancement First, normalize the feature output Y1 from the previous step again: To enhance the image's perception capability in low-light scenes such as nighttime, the input feature map is transformed from the spatial domain to the frequency domain: in and These represent the amplitude and phase spectra, respectively, and F denotes the two-dimensional discrete Fourier transform (FastFourier). Transform), which is defined as follows: (2). Frequency domain amplitude enhancement and feature recovery Attention enhancement and spectrum reconstruction are performed on the frequency amplitude, while preserving the phase and only a linear transformation is applied to the magnitude of the spectrum: Then, it is transformed back to the spatial domain using the inverse fast Fourier transform (IFFT): Right now: Finally, the frequency domain information is fused with the original residual features, where Y is a learnable parameter: The above structure can improve the feature representation ability of the model while maintaining computational efficiency.

[0042] In the AIFI section, further improvements are made to enhance the model's robustness in complex environments by incorporating a hybrid RT-DETR encoding. The feedforward neural network module in the AIFI module of the encoder has been improved to SP-FFN, which specifically includes the following steps: (1). Module Input and Spatial Transformation First, channel expansion is performed using a 1×1 convolution kernel to obtain intermediate features, and then two sub-channels are generated: Where r is the expansion ratio (e.g., 2), F is split into two parts: The expanded features are then processed using depthwise separable convolution (DConv) to enhance local receptive fields and spatial awareness, followed by nonlinear fusion. Where O represents element-wise multiplication and GELU is the Gaussian error linear unit activation function, this structure retains the nonlinear expressive power of the traditional FFN while enhancing the explicit coupling between features.

[0043] (2) Enhanced frequency domain filtering The gated output is then reduced in dimensionality using a 1×1 convolution to restore its original dimensions. To extract frequency information from local regions, the input feature map is divided into multiple non-overlapping patches. First, boundaries are added, and then the patches are split, with each patch denoted as P×P. Perform a two-dimensional real-number fast Fourier transform (Real2DFFT) on each split patch to obtain its spectral representation. To enhance the model's sensitivity to specific frequency structures, a learnable frequency domain quantization matrix Q∈RC×1×1×P×(P / 2+1) is introduced to weight and filter the spectrogram, enabling the model to actively learn which frequency components are more helpful in identifying the target features of traffic participants (such as contour edges, high-frequency textures, etc.). After inverse FFT of the filtered spectral data, patch reconstruction is performed. Finally, after restoring the output tensor Yf to its original data type, the final output Y is obtained. final ∈RB×C×H×W.

[0044] Compared to the original RT-DETR backbone network, the improved M_Block module achieves adjustable spatial perception enhancement by introducing multi-branch spatial convolution and gating mechanisms. Simultaneously, it combines Fourier analysis to regulate feature frequency components, significantly improving image processing capabilities and further enhancing the model's expressive power and stability. Building upon this, the improved SP-FFN module completes the process of "spatial feature enhancement - nonlinear enhancement - frequency domain filtering - reconstruction fusion," greatly improving spatiotemporal modeling capabilities. This results in significantly improved accuracy in recognizing distant, small-sized targets (such as pedestrians and vehicles) at urban intersections from a drone's perspective, exhibiting a clear advantage in extremely complex scenarios such as nighttime and heavy fog, while maintaining low computational complexity.

[0045] Intersection images captured by drones were processed, and traffic participants (including pedestrians, motor vehicles, and non-motorized vehicles) were labeled. The images were then divided into training, validation, and test sets in an 8:1:1 ratio. The datasets were used for model training and testing, and the model weights were saved.

[0046] Step 2: Take off the drone from the intersection to be inspected and record video of traffic flow at the intersection. Use the drone to take off from the intersection, maintaining a flight altitude between 80-120 meters, fixing the position at the center of the intersection, recording the latitude and longitude coordinates, adjusting the camera angle to a standard overhead view, and ensuring a consistent shooting angle. Once the drone is stable, begin recording. The video format should be standard MP4, with a consistent resolution of 1920*1080 and a consistent frame rate of 60 frames per second.

[0047] Step 3: Input the collected intersection traffic operation video into the trained model to collect intersection status and traffic flow data, and identify and extract detailed movement information of traffic participants. Furthermore, using the improved neural network model trained in step 1 and the multi-object tracking model DeepSort, the traffic participants in the drone's aerial video of the intersection are detected and tracked to obtain the trajectory coordinate data of each target. The final output is a CSV file, and its specific output format is as follows: in, Indicates the first i The data is a CSV file containing the trajectory coordinates of several intersections. In the matrix, in Indicates the first [item] in the intersection video. n The ID number of each traffic participant. cn Indicates video i The Middle n Categories of traffic participants xn and yn They represent the first n The x and y coordinates of the top-left corner of the target detection bounding box. wn and hn These represent the width and height of the corresponding detection box, respectively. fn Indicates the video's first n Frame time.

[0048] Furthermore, the trajectory coordinate data can be used to extract corresponding information such as speed, acceleration, and direction of motion. For the i-th traffic participant, the trajectory is divided into n small segments using a segmentation method. Each segment approximates a straight line. The ratio of the Euclidean distance difference of each segment to the time frame is calculated and denoted as the location speed of that segment. The speed is defined as follows: in, vi The average speed (km / h) of motor vehicles passing through the intersection. and They represent the first i The coordinates of the start and end points of the segment trajectory. ...

[0049] Its acceleration is defined as: Its direction of motion is defined as: Step 4: Collect traffic conflict data and construct potential conflict indicators for intersections. First, extract potential conflict objects within the intersection. Specifically, collect events with a TTC value < 2.5s extracted from the video, establish a set of potential interactive conflict object pairs, and then classify them into four categories according to the type of traffic conflict participants: motor vehicle-motor vehicle, pedestrian-non-motor vehicle, non-motor vehicle-motor vehicle, and pedestrian-motor vehicle.

[0050] The Interaction Conflict Potential (ICP) index is defined as follows: in i, j For two traffic participants in a potential conflict. The time interval for them to reach potential conflict points is a constant, and C is the set of potential conflicting object pairs predicted. The larger the index, the higher the risk of conflict.

[0051] The Movement Disorder Index (MDI) for intersections is defined as follows: in, p k The entropy value is the probability distribution of traffic participants in each direction. When the different vehicle movement directions are more dispersed and uniform, the entropy value is higher, indicating that the traffic flow is chaotic and difficult to predict; conversely, if most vehicles travel along a few paths, the entropy value is lower.

[0052] The Abnormal Behavior Index (ABI) at intersections is defined as follows: Where N is the total number of vehicles during the statistical period, I is the indicator function, and a0 is the set acceleration threshold. This indicator reflects the frequency of sudden deceleration behavior by traffic participants; a higher value indicates a potential conflict or disturbance.

[0053] The intersection congestion pressure index (CPI) is defined as follows: in, The physical area occupied by traffic participants within a fixed time window. Ai Indicates the first i The area of ​​the union of the bounding boxes of all traffic participants in the frame, where T is the observation duration.

[0054] Define the intersection violation interference index as (ViolationInterferenceIndex, VII): in, Nvio The number of times pedestrians or non-motorized vehicles enter the restricted area (motor vehicle lane) per unit time, where T is the observation duration. A higher value indicates more frequent violations.

[0055] Step 5: Construct a comprehensive rating model and use the larger model to improve the initial model.

[0056] First, establish an initial comprehensive security evaluation model framework: (1). Define the input vector: Combine the five core indicators (ICP, MDI, ABI, CPI, VII) generated in step 4 into a feature vector. Simultaneously initialize the default weights. .

[0057] (2). The framework is improved by using a large model (LLM) based on the detailed motion information of traffic participants extracted in step 3.

[0058] Preferably, the DeepSeekR1 model is used to improve the model framework. Specifically, the user inputs a prompt word into the large model for initialization, including but not limited to the following formats: "Your current role is that of an expert in the field of transportation engineering. You are proficient in the service level safety evaluation system for urban road intersections, familiar with the threshold setting standards for traffic conflict parameters such as TTC and PET, and possess background knowledge and experience in model parameter tuning related to deep learning. Your goal is to accurately perceive the operating status of intersections and achieve precise quantitative assessment of intersection safety risks by analyzing and adaptively optimizing multi-source data input by users."

[0059] Your answer needs to pay special attention to adhering to the following core principles: Principle 1: Always prioritize the safety of all traffic participants, and then consider other factors such as traffic efficiency and potential risks.

[0060] Principle 2: Dynamically adapt to the actual characteristics of each intersection to avoid ambiguity in decision-making. If there is insufficient reliable data or arguments to support the decision, adopt a conservative strategy.

[0061] Principle 3: Maintain logical interpretability. Responses should demonstrate the thought process and provide relevant textual evidence to support the decision-making process, thus meeting the transparency requirements of security assessments.

[0062] After you finish outputting, please ask me if I have any suggestions for improvement. If so, please rewrite the output based on the above guidelines, taking those suggestions into account. No need to repeat content; let me know if you are ready. Meanwhile, users input structured feature descriptions of the intersection to be detected into the large model, including but not limited to: intersection geometric attributes: intersection type, number of lanes in each direction, and sidewalk width.

[0063] Intersection functional attributes: vehicle type at the intersection, whether it is near a school / commercial area, and peak-hour traffic flow ratio.

[0064] Intersection traffic flow data: traffic participant trajectory coordinate data, historical traffic flow statistics Intersection conflict data: historical accident records at intersections, and handling of traffic violations at intersections. The large model analyzes the risk mechanism of intersections based on the traffic engineering knowledge base and multi-dimensional features, matches similar scenario case libraries, and finally generates the weight adjustment strategy W1 and the thinking process.

[0065] Step 6: Calculate using the improved model and output the comprehensive safety evaluation level of the intersection.

[0066] First, all indices are standardized using the following formula: Furthermore, a weighted strategy is constructed using expert experience, and the overall security score is defined as: Construct security level classification rules, and output the corresponding security level (AE) based on the classification rules. Furthermore, the safety level classification rules are as follows: 0.0≤S<0.4 is classified as E (dangerous), 0.40≤S<0.55 is classified as D (relatively dangerous), 0.55≤S<0.7 is classified as C (moderate), 0.70≤S<0.85 is classified as B (relatively safe), and 0.85≤S<1 is classified as A (very safe).

[0067] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for intersection state perception and traffic safety evaluation based on unmanned aerial vehicles (UAVs) and large-scale models, characterized in that, The method includes: S1: Construct a traffic target detection and tracking model for intersections based on the perspective of drones; S2: Take off the drone from the intersection to be inspected and fly over the intersection to capture video of traffic operation at the intersection; S3: Input the collected intersection traffic operation video into the trained model, collect intersection status and traffic flow data, and identify and extract detailed movement information of traffic participants; S4: Collect traffic conflict data and construct potential conflict indicators for intersections; S5: Construct a comprehensive rating model and use the larger model to improve the initial model; S6: Calculate using the improved model and output the comprehensive safety evaluation level of the intersection.

2. The method for intersection state perception and traffic safety evaluation based on UAVs and large models according to claim 1, characterized in that: In step S1, drones are used to collect aerial view images of urban intersections and the LabelImg tool is used to label and build a basic dataset. Finally, the dataset is divided into training set, validation set and test set in a ratio of 8:1:

1.

3. The method for intersection state perception and traffic safety evaluation based on UAVs and large models according to claim 1, characterized in that: In step S1, an improved intersection traffic participant target detection model based on RT-DETR is constructed. The improvement is achieved by modifying the Backbone and AIFI parts.

4. The method for intersection state perception and traffic safety evaluation based on UAVs and large models according to claim 3, characterized in that: In step S1, the improvement to the Backbone section is as follows: the Block module in the original RT-DETR backbone network is replaced with an M-Block module. The introduced M-Block module follows the design philosophy of the MetaFormer structure, and its definition can be summarized as follows: ； Where X is the input feature map, F MSIM For spatial interaction modules, F FEU This is a feature enhancement unit.

5. The method for intersection state perception and traffic safety evaluation based on UAVs and large models according to claim 3, characterized in that: In step S1, the improvement to the AIFI part is to replace the FNN module of the original RT-DETR backbone network with the SP-FFN module. The definition of the introduced SP-FFN module can be summarized as follows: ； Where X is the feature representation, Q is the frequency domain quantization matrix, F is the real two-dimensional Fourier transform, and Fold denotes the concatenation operation. Represents the GEGLU gating mechanism. This is the Hadamard product in element-wise multiplication.

6. The method for intersection state perception and traffic safety evaluation based on UAVs and large models according to claim 1, characterized in that: In step S2, the drone takes off and hovers at an altitude of 80-120 meters. The drone captures a 15-minute video of the intersection, with a resolution of 1920*1080, a frame rate of 60 frames per second, and a standard mp4 format.

7. The method for intersection state perception and traffic safety evaluation based on UAVs and large models according to claim 1, characterized in that: In step S3, the traffic participants collected include three categories: pedestrians, non-motorized vehicles, and motorized vehicles. The traffic flow data includes target ID, target category, target detection box coordinates, target detection box height and width, and the frame in which the target detection box is located. The detailed motion information of the traffic participants includes speed, acceleration, and direction of motion.

8. The method for intersection state perception and traffic safety evaluation based on UAVs and large models according to claim 1, characterized in that: In step S4, the potential conflict indicators at the intersection include the intersection interaction conflict potential index, the intersection movement disorder index, the intersection abnormal driving behavior index, the intersection congestion pressure index, and the intersection violation interference index; in step S4, Set to 0.2, directional sector number to 8, a0 to 2.5, and T to 60s.

9. The method for intersection state perception and traffic safety evaluation based on UAVs and large models according to claim 1, characterized in that: In step S5, the user inputs intersection information into the large model to improve the initial model. The large model then infers the weight settings and reasoning process in the grade evaluation model. The intersection information includes intersection geometric attributes, intersection functional attributes, intersection traffic data, and intersection conflict data. The weights output by the large model include W1, W2, W3, W4, and W5.

10. The method for intersection state perception and traffic safety evaluation based on UAVs and large models according to claim 1, characterized in that: In step S6, a weighted strategy is constructed using expert experience.

Images