Road accident and illegal parking detection and early warning method and system based on intelligent transportation

Abstract

The application discloses a road accident and illegal parking identification and early warning method and system based on intelligent traffic, relates to the technical field of intelligent traffic, and comprises the following steps: analyzing a preprocessed road monitoring video stream in real time, perceiving multi-dimensional accident features and multi-modal vehicle state features; constructing a parallel accident identification model and an illegal parking detection model, constructing an adaptive threshold based on camera parameters and environmental features, identifying accident scenes and illegal parking scenes by using the multi-dimensional accident features and the multi-modal vehicle state features in combination with the adaptive threshold; estimating the influence range of the accident scenes and the illegal parking scenes, and combining a multi-level early warning response mechanism to link sound and light alarms, navigation push and a rescue platform. The application constructs a more efficient and more accurate expressway accident prevention and control system by combining intelligent accident and illegal parking identification with a multi-level early warning linkage mechanism, and guarantees travel safety.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and more specifically, to a method and system for detecting and warning of road accidents and illegal parking based on intelligent transportation. Background Technology

[0002] With the rapid expansion of my country's expressway network and the continuous growth of motor vehicle ownership, road traffic accidents occur frequently, especially secondary accidents on expressways, which often cause more serious casualties and property losses. Traditional accident alarm and early warning methods have many shortcomings, and intelligent technology is urgently needed to improve rescue efficiency and proactive prevention and control capabilities. In the existing technology, expressway accident alarms mainly rely on drivers and passengers to make emergency calls. However, this method has the following problems: (1) Inaccurate positioning: the alarm caller often has difficulty accurately describing the accident location (such as the mileage, direction, etc.), which leads to delays in rescue; (2) Delayed information transmission: alarm information needs to be transferred through multiple levels, and traffic police, road administration, rescue and other departments cannot share data in real time, affecting the efficiency of collaboration; (3) Failure in extreme scenarios: if the driver is unconscious or unable to make a call, the accident may not be discovered for a long time, missing the best rescue time.

[0003] Traditional highway accident early warning systems primarily rely on manually placed warning signs or variable message signs, which suffer from slow response and limited coverage. In recent years, the Internet of Things (IoT), Artificial Intelligence (AI), and big data technologies have been gradually applied to traffic management; however, some existing solutions only focus on location (such as GPS) or video analysis, failing to form a closed loop of "alarm-early warning-coordination." Data from traffic management, road administration, and rescue departments is not shared, hindering collaborative decision-making. There is a lack of combined application of multi-dimensional early warning methods, such as audio-visual equipment, navigation push notifications, and remote broadcasting. With the development of 5G, vehicle-to-everything (V2X), and edge computing technologies, intelligent transportation systems are evolving towards real-time perception, proactive early warning, and multi-party collaboration. Therefore, there is an urgent need for a new road accident early warning method that integrates AI accident recognition, big data risk prediction, and multi-departmental coordinated response to address the shortcomings of traditional models and improve the intelligence level of highway safety management. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a method and system for detecting and warning of road accidents and illegal parking based on intelligent transportation. By combining AI-based accident and illegal parking identification with a multi-level early warning linkage mechanism, a more efficient and accurate highway accident prevention and control system is constructed, thereby reducing secondary accidents and ensuring travel safety.

[0005] The first aspect of this invention provides a method for detecting and warning of road accidents and illegal parking based on intelligent transportation, comprising the following steps: Acquire road surveillance video streams and preprocess them, analyze the preprocessed road surveillance video streams in real time, and perceive multi-dimensional accident characteristics and multi-modal vehicle status characteristics. Parallel accident recognition and illegal parking detection models are constructed. A contrastive learning strategy is used to train the models. An adaptive threshold is constructed based on camera parameters and environmental features to optimize the accident recognition and illegal parking detection models. In the two model branches, the multi-dimensional accident features and multi-modal vehicle state features are combined with adaptive thresholds to identify accident scenarios and illegal parking scenarios. Estimate the impact range of accident scenarios and illegal parking scenarios, and combine them with a multi-level early warning response mechanism to link audible and visual alarms, navigation push and rescue platform.

[0006] In this solution, the pre-processed road surveillance video stream is analyzed in real time to perceive multi-dimensional accident characteristics, including: In the accident feature perception branch, local feature encoding path and spatial distribution feature encoding path are constructed, the region of interest in the preprocessed video frame is extracted, the local feature encoding path is imported, and the ResNeXt-101 network improved by dynamic convolution kernel is used as the backbone network to adaptively extract local feature maps. The local feature map is embedded using coordinate attention. One-dimensional global average pooling is performed on the height and width directions respectively to generate two directional feature vectors, which are then concatenated to obtain coordinate features. The coordinate features are oriented and encoded. The encoded features are normalized into attention weights in the height and width directions. The local feature map is then weighted pixel by pixel in the height and width directions. The weighted local feature map is then subjected to channel max pooling to generate a spatial heatmap to locate key damage areas. By fusing local feature maps of different scales through hollow space pyramid pooling, a local feature vector of a preset dimension is generated. The spatial distribution feature encoding path is imported into the preprocessed video frames. The SwingTransformer is used as the backbone network. The long-distance spatial dependence is captured through the window self-attention mechanism, and deformable convolution is superimposed to adapt to the irregular object distribution. A scene graph model is constructed by using a graph neural network. The feature vectors corresponding to vehicles and obstacles are used as nodes. An edge structure is constructed based on the Euclidean distance and motion correlation between nodes. Feature encoding is performed using graph representation learning to generate spatially distributed feature vectors of a preset dimension. The local feature vector and the spatially distributed feature vector are weighted and fused using the Cosine attention function to generate multi-dimensional accident features.

[0007] In this solution, the preprocessed road monitoring video stream is analyzed in real time to perceive multimodal vehicle state characteristics, including: In the multimodal vehicle state feature perception branch, vehicle detection boxes are obtained from the preprocessed video frames, smoothed vehicle trajectory sequences are obtained based on Kalman filtering, and the vehicle trajectory sequences are imported into a noise reduction autoencoder, with Gaussian noise added to the input layer. The vehicle trajectory sequence after adding Gaussian noise is encoded using an LSTM encoder to obtain temporal dependent features, which are then mapped to the latent space using a fully connected layer. The features in the latent space are then reconstructed using an LSTM decoder to generate the reconstructed trajectory. The reconstructed trajectory is imported into a multilayer perceptron for low-dimensional intent encoding to obtain real-time intent classification labels. The preprocessed video frames are then processed using a lightweight U-Net network, and multi-scale context perception is performed through hollow spatial pyramid pooling to obtain lane segmentation results. Multimodal vehicle state features are obtained based on vehicle trajectory sequences, real-time intent classification labels, and lane segmentation results.

[0008] In this solution, parallel accident recognition and illegal parking detection models are constructed, and a contrastive learning strategy is used to train the models, including: Set up a common and shared input layer, construct an accident feature perception branch and a multimodal vehicle state feature perception branch, connect the independent accident recognition task head and illegal parking detection task head, and generate an accident recognition model framework and an illegal parking detection model framework. Accident identification instances and illegal parking detection instances are obtained through data retrieval. Positive and negative sample pairs are constructed to generate accident identification datasets and illegal parking detection datasets, respectively. A comparative learning strategy is implemented to train the model based on the accident identification datasets and illegal parking detection datasets. The cross-entropy loss is used to achieve initial convergence. The contrast loss of the two task heads is alternately optimized. Difficult negative samples are sampled in each training round to strengthen learning. The accident recognition model and illegal parking detection model are obtained by minimizing the loss. In the accident identification task head, positive sample pairs are established based on different perspective segments of the same accident, and negative sample pairs are constructed based on accident and normal driving segments. Accident identification comparison loss is then performed. Represented as: ; in Samples and samples The feature vectors represent positive sample pairs. The feature vectors of other samples in the batch represent negative samples. For temperature hyperparameters, This represents the total number of samples in the training batch. In the illegal parking detection task header, anchor samples are constructed based on illegal parking segments, positive samples are constructed based on illegal parking frames of the same vehicle, and negative samples are constructed based on legal parking frames. The illegal parking detection triplet loss is calculated. Represented as: ; in For anchor samples, As a positive sample, For negative samples, This serves as a boundary margin, controlling the minimum distance between positive and negative samples.

[0009] In this solution, an adaptive threshold is constructed based on camera parameters and environmental features to optimize the accident recognition model and the illegal parking detection model, including: Camera intrinsic and extrinsic parameters are obtained to construct camera parameter features, and light intensity, traffic flow density, and weather conditions are obtained to construct environmental features. The camera parameter features and environmental features are normalized and fused, and weighted fusion is used to generate an environmental state index. Set basic thresholds for accident identification and illegal parking duration, and dynamically adjust the basic thresholds using the environmental state index; An adaptive threshold is constructed using a two-layer MLP structure, and joint training is performed based on maximizing accuracy. The loss function for adaptive threshold joint training is... Represented as: ; in This is the current dynamic threshold of the model. The theoretically optimal threshold, The false alarm penalty coefficient, False alarm rate; Historical dynamic threshold adjustment records of each road segment node are obtained, federated learning is performed based on the historical dynamic threshold adjustment records, the threshold adjustment experience of each road segment node is uploaded, and the central server aggregates and generates regional threshold strategies.

[0010] In this solution, the multi-dimensional accident features and multi-modal vehicle state features are combined with adaptive thresholds to identify accident scenarios and illegal parking scenarios, including: The current multi-dimensional accident features and multi-modal vehicle state features are obtained, and the multi-dimensional accident features and multi-modal vehicle state features are imported into the accident recognition task head and the illegal parking detection task head respectively. The threshold is adaptively adjusted according to the environmental state. In accident scenario determination, a multilayer perceptron is used to determine the accident scenario and warning level, and output the accident type and accident warning. In illegal parking scenario determination, it is determined whether the parking area, stationary time and driving intention corresponding to the multimodal vehicle state characteristics meet the triggering conditions, and the illegal parking scenario is output based on the judgment result.

[0011] This plan estimates the impact range of accident and illegal parking scenarios and integrates a multi-level early warning response mechanism with audible and visual alarms, navigation notifications, and a rescue platform, including: Based on the accident scene, obtain the accident type, number of vehicles involved, average vehicle speed, road curvature, and visibility. Add node attributes to the graph structure corresponding to the current video frame, use spatial graph convolution to perform neighborhood aggregation to update node features, and use 1D dilated convolution to obtain temporal dependency features. The outputs of spatial graph convolution and dilated convolution are concatenated, and two layers of GNN are used for prediction. The risk transmission of node feature mutation and spatial edge is modeled to predict the radius of the core influence area and the radius of the potential influence area. Based on the illegal parking scenario, obtain the type of illegal parking lane, vehicle size, and current traffic flow; assess the current road capacity based on the illegal parking lane type and vehicle size; and predict the impact length based on a fluid dynamics model. Based on the impact range of the accident scenario and illegal parking scenario, a corresponding warning level is generated according to a preset threshold range. A preset response action is obtained according to the warning level, and the response action is sent to the road test equipment and cloud platform for execution.

[0012] The second aspect of this invention provides a road accident and illegal parking identification and early warning system based on intelligent transportation. The system includes: a data acquisition and preprocessing module, a multimodal feature perception module, an identification and detection module, an adaptive threshold adjustment module, an impact range estimation module, and a multi-level early warning and collaborative response module. The data acquisition and preprocessing module acquires the road monitoring video stream, performs noise reduction, enhancement, and frame alignment preprocessing on the road monitoring video stream, extracts key frames and adjusts the resolution to adapt to the input requirements of the recognition and detection module. The multimodal feature perception module analyzes the preprocessed road monitoring video stream in real time to perceive multi-dimensional accident features and multimodal vehicle status features. The identification and detection module constructs parallel accident identification models and illegal parking detection models, and uses a contrastive learning strategy to train the models. The accident identification model identifies accident scenarios through a temporal behavior analysis network and calculates the severity level, while the illegal parking detection model identifies illegal parking behavior. The adaptive threshold adjustment module dynamically adjusts the judgment threshold based on camera parameters and environmental data, achieving intelligent adjustment in both time and space dimensions. The impact range estimation module constructs a road coordinate system through perspective transformation to estimate the impact range of accident scenarios and illegal parking scenarios. The aforementioned early warning and collaborative response module combines a multi-level early warning and response mechanism with audible and visual alarms, navigation push notifications, and a rescue platform.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention improves the efficiency of road traffic safety management by deeply integrating multimodal perception technology with an adaptive decision-making mechanism. Through a parallel accident recognition and illegal parking detection model architecture, combined with high-precision spatiotemporal feature analysis, it significantly enhances the ability to identify abnormal events in complex scenarios. Specifically, the spatiotemporal graph convolutional network models vehicle-to-vehicle interactions and accident propagation trends, making the predicted impact range physically interpretable and overcoming the limitations of traditional fixed-threshold warnings. Regarding the assessment of illegal parking impact, a dynamic capacity calculation model based on fluid dynamics theory fully considers the coupling relationship between lane type, vehicle size, and real-time traffic flow, achieving a scientific quantification of congestion propagation.

[0014] A multi-level early warning and response mechanism constructs a tiered response system from basic alerts to emergency intervention through the coordinated optimization of environmental perception and adaptive threshold adjustment. It links roadside equipment, navigation platforms, and rescue systems to form a closed-loop management system, which not only improves the accuracy and timeliness of event identification but also enables precise dispatch of emergency resources through intelligent impact range prediction. The dynamic threshold mechanism effectively balances the contradiction between false alarms and missed alarms, the fusion of spatiotemporal features enhances adaptability to complex scenarios, and online learning capabilities ensure the system's continuous evolution. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0016] Figure 1 A flowchart of a road accident and illegal parking identification and early warning method based on intelligent transportation is shown; Figure 2 A flowchart illustrating the perception of multi-dimensional accident characteristics is shown. Figure 3 A flowchart illustrating the perception of multimodal vehicle state characteristics is shown. Figure 4 A block diagram of a road accident and illegal parking identification and early warning system based on intelligent transportation is shown. Detailed Implementation

[0017] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0019] Figure 1 A flowchart of a road accident and illegal parking identification and early warning method based on intelligent transportation is shown.

[0020] like Figure 1 As shown, this embodiment provides a method for identifying and issuing early warnings of road accidents and illegal parking based on intelligent transportation, including: S102, acquire road monitoring video stream and preprocess it, analyze the preprocessed road monitoring video stream in real time, and perceive multi-dimensional accident features and multi-modal vehicle status features. S104. Construct parallel accident recognition and illegal parking detection models, train the models using a contrastive learning strategy, construct adaptive thresholds based on camera parameters and environmental features, and optimize the accident recognition and illegal parking detection models. S106, In the two model branches, the multi-dimensional accident features and multi-modal vehicle state features are combined with adaptive thresholds to identify accident scenarios and illegal parking scenarios. S108 estimates the impact range of accident scenarios and illegal parking scenarios, and combines a multi-level early warning response mechanism to link audible and visual alarms, navigation push and rescue platform.

[0021] It should be noted that the real-time acquisition of road monitoring video streams involves extracting keyframes from the video stream at a fixed frame rate and performing time alignment to avoid the impact of jitter. Non-local mean denoising or the BM3D algorithm is used to reduce interference from low light and rain / fog, histogram equalization is used to enhance contrast and improve visibility in nighttime or backlit scenes, and frames are uniformly scaled to the model input size and pixel values ​​are normalized.

[0022] Figure 2 A flowchart illustrating the perception of multi-dimensional accident characteristics is shown.

[0023] According to an embodiment of the present invention, in the accident feature perception branch, a local feature encoding path and a spatial distribution feature encoding path are constructed to improve detection robustness through feature complementarity. Regions of interest are extracted from preprocessed video frames and imported into the local feature encoding path to extract high-granularity features at the vehicle individual level, such as collision deformation and scattered parts. A dynamically modified ResNeXt-101 network with dynamic convolutional kernels is used as the backbone network to adaptively extract local feature maps. Coordinate attention is used to embed the local feature maps into coordinates, a lightweight attention mechanism that enhances the model's sensitivity to spatial location by decomposing two-dimensional global pooling into two one-dimensional directional codes with almost no increase in computation. Local feature maps from the backbone network are acquired, and one-dimensional global average pooling is performed on the height and width directions to generate two directionally aware feature vectors, which are then concatenated to obtain coordinate features. These coordinate features are then directionally aware and encoded. Convolutional compression of the number of channels reduces computational cost, and the Swish activation function enhances non-linear expressive power. The encoded features are normalized to attention weights in the height and width directions, and pixel-wise weighting is applied to the local feature maps in both the height and width directions. Channel max pooling is then performed on the weighted local feature maps to generate a spatial heatmap for locating key damage areas and contour response regions. Finally, local feature maps of different scales are fused using hollow spatial pyramid pooling to generate local feature vectors of a preset dimension.

[0024] The spatial distribution feature encoding path is imported into the preprocessed video frames to analyze the global spatial relationships of the accident scene. A SwinTransformer is used as the backbone network, employing a window self-attention mechanism to capture long-distance spatial dependencies and superimposing deformable convolutions to adapt to irregular object distributions, such as scattered goods. A scene graph model is constructed using a graph neural network, with feature vectors corresponding to vehicles and obstacles as nodes. An edge structure is built based on the Euclidean distance and motion correlation between nodes, and graph representation learning is used for feature encoding to generate spatial distribution feature vectors of a preset dimension. The Cosine attention function is then used to weight and fuse the local feature vectors and spatial distribution feature vectors to align semantics, generating multi-dimensional accident features.

[0025] Figure 3 A flowchart illustrating the perception of multimodal vehicle state characteristics is shown.

[0026] According to an embodiment of the present invention, in the multimodal vehicle state feature perception branch, the vehicle detection box (center coordinates) in the preprocessed video frame is obtained. Width and height ,speed A linear motion model is introduced, and a smoothed vehicle trajectory sequence is obtained based on Kalman filtering. This vehicle trajectory sequence is then imported into a denoising autoencoder, with Gaussian noise added to the input layer. Intent features are extracted based on the denoising autoencoder, and an LSTM encoder is used to encode the Gaussian noise-added vehicle trajectory sequence to obtain temporal dependency features. These features are then mapped to a latent space using a fully connected layer, and an LSTM decoder is used to reconstruct the features in the latent space, generating a reconstructed trajectory. This reconstructed trajectory is then imported into a multilayer perceptron for low-dimensional intent encoding to obtain real-time intent classification labels, including normal, lane change, emergency braking, and illegal U-turn. A lightweight U-Net network is used to process the preprocessed video frames, and MobileNetV3 is used as the encoder. Multi-scale context awareness is achieved through dilated spatial pyramid pooling to obtain the class probability of each pixel and output lane segmentation results. Multimodal vehicle state features are obtained based on the vehicle trajectory sequence, real-time intent classification labels, and lane segmentation results.

[0027] It should be noted that the parallel accident recognition model and illegal parking detection model are constructed using a contrastive learning strategy for training. This includes: setting a common, shared input layer; constructing an accident feature perception branch and a multimodal vehicle state feature perception branch; connecting independent accident recognition task heads and illegal parking detection task heads; and generating the accident recognition model framework and illegal parking detection model framework. Accident recognition instances and illegal parking detection instances are obtained through data retrieval, and positive and negative sample pairs are constructed to generate the accident recognition dataset and illegal parking detection dataset, respectively. The accident recognition dataset includes real accident video clips, labeled with accident type, occurrence frame, and impact range, as well as similar scenarios such as normal driving and sudden braking without collision. The illegal parking detection dataset includes continuous frame sequences of illegal parking, temporary parking, and slow-moving vehicles. The model is trained using a contrastive learning strategy based on the aforementioned accident recognition dataset and illegal parking detection dataset.

[0028] Initial convergence is achieved using cross-entropy loss. The contrastive loss of the two task heads is alternately optimized. Hard negative samples are sampled in each training round to reinforce learning. The trained accident recognition model and illegal parking detection model are obtained by minimizing the loss. In the accident recognition task head, positive sample pairs are established based on different perspective segments of the same accident, and negative sample pairs are constructed based on accident and normal driving segments. This brings the positive sample pairs of the same accident closer together and separates them from other samples. The feature-based accident recognition contrastive loss... Represented as: ; in Samples and samples The feature vectors represent positive sample pairs. The feature vectors of other samples in the batch represent negative samples. For temperature hyperparameters, This represents the total number of samples in the training batch. In the illegal parking detection task header, anchor samples are constructed based on illegal parking segments, positive samples are constructed based on illegal parking frames of the same vehicle, and negative samples are constructed based on legal parking frames. This ensures that the features of illegal parking samples are close to those of similar illegal parking samples and far away from non-illegal parking samples, thus minimizing the illegal parking detection triplet loss. Represented as: ; in For anchor samples, As a positive sample, For negative samples, This serves as a boundary margin, controlling the minimum distance between positive and negative samples.

[0029] It should be noted that by analyzing camera physical parameters and environmental conditions in real time, the judgment thresholds for accident recognition and illegal parking detection are dynamically adjusted to adapt to the false alarm-recall balance requirements in different scenarios. Camera intrinsic parameters such as focal length, optical center coordinates, and distortion coefficients, as well as extrinsic parameters such as installation height, pitch angle, and yaw angle, are acquired to construct camera parameter features. These camera parameter features are used to calculate the perspective transformation matrix from image coordinates to world coordinates, which is used to accurately measure the actual position and speed of the vehicle. Illumination intensity, traffic density, and weather conditions are acquired to construct environmental features. The camera parameter features and environmental features are normalized and fused, and a weighted fusion is used to generate an environmental state index, where the weights are obtained through regression learning from historical data. Basic thresholds for accident recognition and illegal parking duration are set, and the environmental state index is used to dynamically adjust the basic thresholds, where the accident recognition threshold... Represented as: ; in As the basic threshold for accident identification, This is an environmental status index. This is the standard focal length, compensating for differences in field of view between different cameras. The focal length of the camera; Threshold for determining illegal parking duration Represented as: ; in This is the base threshold for the duration of illegal parking. Used to compensate for detection delays caused by installation height.

[0030] An adaptive threshold is constructed using a two-layer MLP structure, and joint training is performed based on maximizing accuracy. The loss function for adaptive threshold joint training is... Represented as: ; in This is the current dynamic threshold of the model. The theoretically optimal threshold, The false alarm penalty coefficient, False alarm rate; Historical dynamic threshold adjustment records of each road segment node are obtained, federated learning is performed based on the historical dynamic threshold adjustment records, the threshold adjustment experience of each road segment node is uploaded, and the central server aggregates and generates regional threshold strategies.

[0031] It should be noted that the current multi-dimensional accident features and multi-modal vehicle state features are acquired and then imported into the accident recognition task head and the illegal parking detection task head, respectively. Thresholds are adaptively adjusted according to the environmental conditions. In accident scene determination, a multilayer perceptron is used to determine the accident scene and warning level, outputting the accident type and warning. The warning level is determined through a comprehensive scoring system using the multi-dimensional accident features. The formula for obtaining and calculating this information is as follows: , This is the weight matrix. Real-time warnings push accident coordinates to navigation software, provide voice prompts for vehicles behind to slow down, and link with emergency lane monitoring cameras to automatically track vehicles associated with the accident. In illegal parking scenario determination, it judges whether the parking area, stationary time, and driving intention corresponding to the multimodal vehicle state characteristics meet the triggering conditions. Based on the judgment result, it outputs the illegal parking scenario, for example, if the vehicle's IoU with the no-parking zone > 0.7; and the stationary duration > 0.7. If the driving intent is classified as illegal parking, then it is determined to be an illegal parking scenario.

[0032] It should be noted that, based on the accident scene, the accident type, number of vehicles involved, average speed, road curvature, and visibility are obtained. Node attributes are added to the graph structure corresponding to the current video frame. Spatial graph convolution is used for neighborhood aggregation to update node features, and 1D dilated convolution is used to obtain temporal dependency features. The outputs of the spatial graph convolution and dilated convolution are concatenated to obtain spatiotemporal features. Two layers of GNN are used for prediction, modeling the risk propagation of node feature mutations and spatial edges, and predicting the radius of the core impact area and the radius of the potential impact area; the radius of the potential impact area... Represented as: ; in This is the normalized value of traffic density. The traffic density influence coefficient. The radius of the core danger zone was determined by the maximum length of the vehicle involved in the accident. and relative velocity Determine, Preferred .

[0033] The spatiotemporal features confirm the authenticity of the accident through abrupt changes in vehicle trajectory and spatial interaction, identify vehicles directly involved in the collision and affected vehicles, analyze spatiotemporal patterns such as traffic density and speed distribution, and predict the direction of secondary accidents. Based on the illegal parking scenario, the type of illegally parked lane, vehicle size, and current traffic flow are obtained. The current road capacity is assessed based on the illegally parked lane type and vehicle size. The basic capacity is calculated based on the single-lane capacity under ideal conditions, lane width correction coefficient, large vehicle mixing correction coefficient, and illegally parked vehicle congestion correction coefficient. The impact length is estimated based on a fluid dynamics model, treating traffic flow as an incompressible viscous fluid and illegally parked vehicles as equivalent to throttle valves in a pipeline. Based on the accident scenario and the impact range of the illegal parking scenario, corresponding warning levels are generated based on preset threshold ranges. Preset response actions are obtained based on the warning levels and sent to road test equipment and the cloud platform for execution. For example, the response action corresponding to a level 1 warning level is to display an accident ahead on an LED screen and provide navigation voice prompts to slow down; the response action corresponding to a level 2 warning level is to activate flashing lights, mark the accident area on a map, and pre-dispatch the rescue platform.

[0034] Figure 4 The diagram shows the architecture of a road accident and illegal parking identification and early warning system based on intelligent transportation.

[0035] The second embodiment of the present invention provides a road accident and illegal parking identification and early warning system based on intelligent transportation. The system includes: a data acquisition and preprocessing module, a multimodal feature perception module, an identification and detection module, an adaptive threshold adjustment module, an impact range estimation module, and a multi-level early warning and collaborative response module. The data acquisition and preprocessing module acquires the road monitoring video stream, performs noise reduction, enhancement, and frame alignment preprocessing on the road monitoring video stream, extracts key frames and adjusts the resolution to adapt to the input requirements of the recognition and detection module. The multimodal feature perception module analyzes the preprocessed road monitoring video stream in real time to perceive multi-dimensional accident features and multimodal vehicle status features. The identification and detection module constructs parallel accident identification models and illegal parking detection models, and uses a contrastive learning strategy to train the models. The accident identification model identifies accident scenarios through a temporal behavior analysis network and calculates the severity level, while the illegal parking detection model identifies illegal parking behavior. The adaptive threshold adjustment module dynamically adjusts the judgment threshold based on camera parameters and environmental data, achieving intelligent adjustment in both time and space dimensions. The impact range estimation module constructs a road coordinate system through perspective transformation to estimate the impact range of accident scenarios and illegal parking scenarios. The aforementioned early warning and collaborative response module combines a multi-level early warning and response mechanism with audible and visual alarms, navigation push notifications, and a rescue platform.

[0036] The third embodiment of this application also provides a computer-readable storage medium storing computer program code, which, when executed by the processor, causes the electronic device to perform the relevant method steps in the above method embodiments.

[0037] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0038] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0039] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying and issuing early warnings of road accidents and illegal parking based on intelligent transportation, characterized in that, Includes the following steps: Acquire road surveillance video streams and preprocess them, analyze the preprocessed road surveillance video streams in real time, and perceive multi-dimensional accident characteristics and multi-modal vehicle status characteristics. Parallel accident recognition and illegal parking detection models are constructed. A contrastive learning strategy is used to train the models. An adaptive threshold is constructed based on camera parameters and environmental features to optimize the accident recognition and illegal parking detection models. In the two model branches, the multi-dimensional accident features and multi-modal vehicle state features are combined with adaptive thresholds to identify accident scenarios and illegal parking scenarios. Estimate the impact range of accident scenarios and illegal parking scenarios, and combine them with a multi-level early warning response mechanism to link sound and light alarms, navigation push and rescue platform; An adaptive threshold is constructed based on camera parameters and environmental features to optimize the accident recognition model and illegal parking detection model, including: Camera intrinsic and extrinsic parameters are obtained to construct camera parameter features, and light intensity, traffic flow density, and weather conditions are obtained to construct environmental features. The camera parameter features and environmental features are normalized and fused, and weighted fusion is used to generate an environmental state index. Set basic thresholds for accident identification and illegal parking duration, and dynamically adjust the basic thresholds using the environmental state index; Adaptive thresholding is constructed using a two-layer MLP structure, and joint training is performed based on maximizing accuracy. Historical dynamic threshold adjustment records of each road segment node are obtained, federated learning is performed based on the historical dynamic threshold adjustment records, the threshold adjustment experience of each road segment node is uploaded, and the central server aggregates and generates regional threshold strategies.

2. The method for identifying and issuing early warnings of road accidents and illegal parking based on intelligent transportation according to claim 1, characterized in that, Real-time analysis of pre-processed road surveillance video streams to perceive multi-dimensional accident characteristics, including: In the accident feature perception branch, local feature encoding path and spatial distribution feature encoding path are constructed, the region of interest in the preprocessed video frame is extracted, the local feature encoding path is imported, and the ResNeXt-101 network improved by dynamic convolution kernel is used as the backbone network to adaptively extract local feature maps. The local feature map is embedded using coordinate attention. One-dimensional global average pooling is performed on the height and width directions respectively to generate two directional feature vectors, which are then concatenated to obtain coordinate features. The coordinate features are oriented and encoded. The encoded features are normalized into attention weights in the height and width directions. The local feature map is then weighted pixel by pixel in the height and width directions. The weighted local feature map is then subjected to channel max pooling to generate a spatial heatmap to locate key damage areas. By fusing local feature maps of different scales through hollow space pyramid pooling, a local feature vector of a preset dimension is generated. The spatial distribution feature encoding path is imported into the preprocessed video frames. The Swing Transformer is used as the backbone network. The long-distance spatial dependence is captured through the window self-attention mechanism, and deformable convolution is superimposed to adapt to the irregular object distribution. A scene graph model is constructed by using a graph neural network. The feature vectors corresponding to vehicles and obstacles are used as nodes. An edge structure is constructed based on the Euclidean distance and motion correlation between nodes. Feature encoding is performed using graph representation learning to generate spatially distributed feature vectors of a preset dimension. The local feature vector and the spatially distributed feature vector are weighted and fused using the Cosine attention function to generate multi-dimensional accident features.

3. The method for identifying and issuing early warnings of road accidents and illegal parking based on intelligent transportation according to claim 1, characterized in that, Real-time analysis of pre-processed road surveillance video streams to perceive multimodal vehicle state characteristics, including: In the multimodal vehicle state feature perception branch, vehicle detection boxes are obtained from the preprocessed video frames, smoothed vehicle trajectory sequences are obtained based on Kalman filtering, and the vehicle trajectory sequences are imported into a noise reduction autoencoder, with Gaussian noise added to the input layer. The vehicle trajectory sequence after adding Gaussian noise is encoded using an LSTM encoder to obtain temporal dependent features, which are then mapped to the latent space using a fully connected layer. The features in the latent space are then reconstructed using an LSTM decoder to generate the reconstructed trajectory. The reconstructed trajectory is imported into a multilayer perceptron for low-dimensional intent encoding to obtain real-time intent classification labels. The preprocessed video frames are then processed using a lightweight U-Net network, and multi-scale context perception is performed through hollow spatial pyramid pooling to obtain lane segmentation results. Multimodal vehicle state features are obtained based on vehicle trajectory sequences, real-time intent classification labels, and lane segmentation results.

4. The method for identifying and issuing early warnings of road accidents and illegal parking based on intelligent transportation according to claim 1, characterized in that, Parallel accident recognition and illegal parking detection models are constructed, and a contrastive learning strategy is used to train the models, including: Set up a common and shared input layer, construct an accident feature perception branch and a multimodal vehicle state feature perception branch, connect the independent accident recognition task head and illegal parking detection task head, and generate an accident recognition model framework and an illegal parking detection model framework. Accident identification instances and illegal parking detection instances are obtained through data retrieval. Positive and negative sample pairs are constructed to generate accident identification datasets and illegal parking detection datasets, respectively. A comparative learning strategy is implemented to train the model based on the accident identification datasets and illegal parking detection datasets. The cross-entropy loss is used to achieve initial convergence. The contrast loss of the two task heads is alternately optimized. Difficult negative samples are sampled in each training round to strengthen learning. The accident recognition model and illegal parking detection model are obtained by minimizing the loss. In the accident identification task head, positive sample pairs are established based on different perspective segments of the same accident, and negative sample pairs are constructed based on accident and normal driving segments. Accident identification comparison loss is then performed. Represented as: ； in Samples and samples The feature vectors represent positive sample pairs. The feature vectors of other samples in the batch represent negative samples. For temperature hyperparameters, This represents the total number of samples in the training batch. In the illegal parking detection task header, anchor samples are constructed based on illegal parking segments, positive samples are constructed based on illegal parking frames of the same vehicle, and negative samples are constructed based on legal parking frames. The illegal parking detection triplet loss is calculated. Represented as: ； in For anchor samples, As a positive sample, For negative samples, This serves as a boundary margin, controlling the minimum distance between positive and negative samples.

5. The method for identifying and issuing early warnings of road accidents and illegal parking based on intelligent transportation according to claim 1, characterized in that, The system utilizes the aforementioned multi-dimensional accident features and multi-modal vehicle state features, combined with adaptive thresholds, to identify accident scenarios and illegal parking scenarios, including: The current multi-dimensional accident features and multi-modal vehicle state features are obtained, and the multi-dimensional accident features and multi-modal vehicle state features are imported into the accident recognition task head and the illegal parking detection task head respectively. The threshold is adaptively adjusted according to the environmental state. In accident scenario determination, a multilayer perceptron is used to determine the accident scenario and warning level, and output the accident type and accident warning. In illegal parking scenario determination, it is determined whether the parking area, stationary time and driving intention corresponding to the multimodal vehicle state characteristics meet the triggering conditions, and the illegal parking scenario is output based on the judgment result.

6. The method for identifying and issuing early warnings of road accidents and illegal parking based on intelligent transportation according to claim 1, characterized in that, Estimate the impact range of accident and illegal parking scenarios, and integrate a multi-level early warning response mechanism with audible and visual alarms, navigation notifications, and rescue platforms, including: Based on the accident scene, obtain the accident type, number of vehicles involved, average vehicle speed, road curvature, and visibility. Add node attributes to the graph structure corresponding to the current video frame, use spatial graph convolution to perform neighborhood aggregation to update node features, and use 1D dilated convolution to obtain temporal dependency features. The outputs of spatial graph convolution and dilated convolution are concatenated, and two layers of GNN are used for prediction. The risk transmission of node feature mutation and spatial edge is modeled to predict the radius of the core influence area and the radius of the potential influence area. Based on the illegal parking scenario, obtain the type of illegal parking lane, vehicle size, and current traffic flow; assess the current road capacity based on the illegal parking lane type and vehicle size; and predict the impact length based on a fluid dynamics model. Based on the impact range of the accident scenario and illegal parking scenario, a corresponding warning level is generated according to a preset threshold range. A preset response action is obtained according to the warning level, and the response action is sent to the road test equipment and cloud platform for execution.

7. A road accident and illegal parking identification and early warning system based on intelligent transportation, characterized in that, To implement the intelligent transportation-based road accident and illegal parking identification and early warning method as described in any one of claims 1-6, the system includes: a data acquisition and preprocessing module, a multimodal feature perception module, an identification and detection module, an adaptive threshold adjustment module, an impact range estimation module, and a multi-level early warning and collaborative response module; The data acquisition and preprocessing module acquires the road monitoring video stream, performs noise reduction, enhancement, and frame alignment preprocessing on the road monitoring video stream, extracts key frames and adjusts the resolution to adapt to the input requirements of the recognition and detection module. The multimodal feature perception module analyzes the preprocessed road monitoring video stream in real time to perceive multi-dimensional accident features and multimodal vehicle status features. The identification and detection module constructs parallel accident identification models and illegal parking detection models, and uses a contrastive learning strategy to train the models. The accident identification model identifies accident scenarios through a temporal behavior analysis network and calculates the severity level, while the illegal parking detection model identifies illegal parking behavior. The adaptive threshold adjustment module dynamically adjusts the judgment threshold based on camera parameters and environmental data, achieving intelligent adjustment in both time and space dimensions. The impact range estimation module constructs a road coordinate system through perspective transformation to estimate the impact range of accident scenarios and illegal parking scenarios. The aforementioned early warning and collaborative response module combines a multi-level early warning and response mechanism with audible and visual alarms, navigation push notifications, and a rescue platform.

Images