Highway traffic incident detection method and system based on size model cooperation

Abstract

The present application belongs to the field of intelligent transportation, and relates to a highway traffic event detection method, system, device and storage medium based on size model cooperation. At present, single small model is used to detect traffic events on highways, which has the problems of insufficient feature extraction capability and high false alarm rate. The present application uses high-definition cameras along the highway to collect video data, and small models perform preliminary detection with fast computing power, extract event video clips and generate preliminary detection results. After receiving the data, large models analyze, extract deep features and make accurate judgments. If it is confirmed to be a real event, the final detection report is generated. This method improves the real-time performance and accuracy of traffic event detection, reduces the cost and false alarm rate, and provides support for efficient management of highways. At the same time, the present application protects the size model cooperation architecture and model interaction mechanism.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to a method and system for detecting highway traffic incidents based on the collaboration of large and small models. Background Technology

[0002] With the continuous growth of highway traffic volume, timely and accurate detection of traffic incidents has become a crucial aspect of ensuring road safety and smooth traffic flow. Currently, many highway owners have deployed incident monitoring systems, most of which use single, small models (such as the YOLO series models) for detection. However, due to their low model complexity and limited parameters, these small models have significant shortcomings in feature extraction capabilities when facing complex scenarios and diverse traffic incidents. For example, when detecting incidents such as road construction (cone detection), non-motorized vehicle intrusion, and pedestrian intrusion—which involve small targets with indistinct features—the accuracy of small models is low, leading to a large number of false alarms.

[0003] For highway owners, these false alarms not only waste a lot of human and material resources on verifying false incidents, but also distract management departments from paying attention to and handling real traffic incidents, seriously affecting the efficiency of handling traffic incidents and management costs, and causing great trouble to the daily operation and management of highways. Summary of the Invention

[0004] The embodiments of the present invention provide a method and system for detecting highway traffic incidents based on large and small model collaboration, so as to improve the accuracy and timeliness of traffic incident detection, reduce the false alarm rate of the system, and provide strong support for the efficient management of highways.

[0005] To achieve the above objectives, the present invention adopts the following technical solution.

[0006] A highway traffic incident detection method based on size-model collaboration includes:

[0007] S1 continuously collects and transmits highway video data in real time by deploying high-definition cameras and establishing transmission links;

[0008] S2 loads a small model and preprocesses video data to detect traffic scenes, extract event segments, and generate preliminary detection results.

[0009] S3 transmits the preliminary detection results output by the small model to the large model through the event fusion and confirmation service module. The large model then analyzes, verifies, and performs in-depth analysis of the preliminary detection results and feeds them back to the event fusion and confirmation service module, which in turn generates the final report.

[0010] S4 will submit the final report, categorized and forwarded to the external management center, enabling the external management center to initiate contingency plans and provide feedback records.

[0011] Preferably, step S1 includes:

[0012] Based on the factors of highways, S11 plans the location of high-definition cameras and ensures that the camera's field of view can cover key areas; at the same time, considering the differences in traffic flow on different road sections, the density of cameras is appropriately increased in areas with high traffic flow to obtain more comprehensive traffic information.

[0013] Factors affecting highways include: the characteristics of the highway sections, the location of curves, and the distribution of entrances and exits;

[0014] Key areas include: accident-prone road sections and transportation hubs;

[0015] S12 places high-definition cameras at the planned locations and adjusts the camera angles and focal lengths to ensure that the high-definition cameras capture stable and clear images.

[0016] S13 establishes a data transmission network for high-definition cameras through a wired network transmission link, and configures corresponding network devices and network parameters to ensure that video data can be transmitted stably and quickly to the subsequent processing system.

[0017] The S14 continuously collects video data using a high-definition camera at a set frame rate. The collected video data is then transmitted in real time to the computing device where the small model is located via a pre-established data transmission link, providing timely and accurate data support for subsequent traffic incident detection.

[0018] Preferably, step S2 includes:

[0019] S21 loads a pre-trained small model from the storage medium, and at the same time loads the configuration file required for the model to run;

[0020] The configuration file includes: model hyperparameters and class label information; it ensures that the small model can run normally under the current hardware environment and detection task.

[0021] S22 preprocesses the video data acquired in real time in sub-step S1 into a format acceptable to the small model; normalizes the images and maps pixel values ​​to a specific range, which can accelerate the training and inference speed of the small model.

[0022] Preprocessing operations include resizing the video images to fit the requirements of the small model input layer;

[0023] S23 inputs the pre-processed video data frame by frame into the small model for inference operations, and performs feature extraction and analysis on the traffic scene in the video frame;

[0024] S24 determines the location information of the event based on the output of the small model and extracts the video segment containing the event from the video stream; when the small model detects a target that matches the preset traffic event type from the video segment, the event determination mechanism is triggered.

[0025] Based on the extracted video segments and time determination mechanism, S25 obtains preliminary detection results, including: event type, clarifying whether it is an abnormal parking, non-motorized vehicle intrusion, or other specific type of event; recording the timestamp of the event, accurate to the second or even millisecond, for easy subsequent tracing; and determining the location information of the event.

[0026] Location information of the event includes: relative location coordinates based on the camera, or actual geographical location information obtained by combining with a map system.

[0027] Preferably, sub-step S23 includes:

[0028] The small model identifies various target objects in the image, including key traffic elements such as pedestrians, vehicles, non-motorized vehicles, construction cones, and spilled materials.

[0029] Post-processing judgments are made based on the identified traffic elements, and information such as the category, movement trajectory, and duration of the traffic elements are comprehensively used to determine events including abnormal parking, pedestrian intrusion, non-motorized vehicle intrusion, motorcycle intrusion, occupation of emergency lanes, vehicles driving in the wrong direction, traffic congestion, construction areas, traffic accidents, and road debris.

[0030] Preferably, step S3 includes:

[0031] S31 sends the preliminary detection results to the large model through a high-speed network transmission channel, so that the large model can receive, analyze and verify them.

[0032] S32 uses a large model to perform in-depth and comprehensive feature extraction on video clips, adopts different strategies for different types of traffic events, and makes accurate judgments on events based on the extracted deep features.

[0033] S33 If the large model confirms that the event is a real traffic anomaly, then the final detection report is generated using the judgment results of the large model.

[0034] Preferably, in sub-step S32:

[0035] The large model adopts the Transformer architecture, which has a self-attention mechanism that can associate information globally and simultaneously and comprehensively analyze the position, driving direction, collision degree and other features of all involved vehicles, as well as the relationships between vehicles.

[0036] Large models can learn the general features of various objects from massive image data. When detecting road traffic incidents, they can identify and judge road debris or complex scenes that occur less than a preset threshold number of times in the training data more quickly and accurately.

[0037] Preferably, step S4 includes:

[0038] S41 obtains the final detection report generated by the large model and determines whether it is an abnormal event based on the judgment results in the report;

[0039] For emergency events, S42 adopts a high-priority reporting method, including multiple channels such as dedicated network lines and SMS, to ensure that event information can be delivered to the event detection business system in the shortest possible time. For general events, event information can be reported to the management center through a common network interface, while ensuring the accuracy and integrity of information transmission.

[0040] S43 parses and stores the received data, and activates the corresponding emergency plan according to the type and severity of the event, including: for vehicle fire incidents, immediately notifying the fire department and traffic police department, and guiding nearby vehicles to evacuate; for minor congestion incidents, dispatching traffic control personnel to the scene to direct traffic, or releasing real-time traffic information through the traffic guidance system to guide vehicles to detour.

[0041] During incident handling, S44 collects real-time progress information and provides feedback to external parties. After the incident is completed, it records the handling process and results in detail, including the time, location, type, handling measures, and handling time of the incident, forming an incident handling archive for subsequent data analysis and experience summarization, and continuously optimizing traffic incident handling procedures and emergency plans.

[0042] Preferably, the large and small model collaborative architecture includes: a highway traffic incident detection architecture that works collaboratively with the large and small models, wherein the small model is responsible for initial rapid detection and the large model is responsible for secondary accurate confirmation. Through the collaborative work of the two, the efficiency and accuracy of highway traffic incident detection are effectively improved.

[0043] Model interaction mechanism: After the small model detects a suspected event, it extracts video clips and generates a preliminary detection report, which is then sent to the large model via a high-speed network transmission channel. The large model receives the data, performs analysis, deep feature extraction, and precise judgment, and then feeds back the results—a complete interactive process. Event fusion and confirmation services are the core control mechanisms for the entire interaction process.

[0044] On the second aspect, this invention provides a highway traffic incident detection system based on large-scale model collaboration, comprising:

[0045] The video acquisition module is used to continuously collect and acquire real-time video data of highways through deployed high-definition cameras and established transmission links.

[0046] The video analysis and processing module is used for:

[0047] Load a small model and preprocess video data to detect traffic scenes, extract event segments, and generate preliminary detection results;

[0048] The event fusion and confirmation service module is used to transmit the preliminary detection results output by the small model to the large model, so that the large model can be parsed, verified, and deeply analyzed before feeding back to the event fusion and confirmation service module, and the video analysis and processing module generates the final report.

[0049] The output module is used to: output the final report and categorize the final report for submission to the external management center, enabling the external management center to initiate contingency plans and provide feedback on the processing records.

[0050] As can be seen from the technical solutions provided by the embodiments of the present invention above, the present invention belongs to the field of intelligent transportation and relates to a method, system, device, and storage medium for highway traffic incident detection based on large and small model collaboration. Currently, highways mostly use a single small model to detect traffic incidents, which suffers from insufficient feature extraction capabilities and a high false alarm rate. The present invention utilizes high-definition cameras along the highway to collect video data. The small model performs preliminary detection using its fast computing power, extracting event video segments and generating preliminary detection results. The large model receives the data and performs parsing, deep feature extraction, and precise judgment. If it confirms a real event, it generates a final detection report. This method, through large and small model collaboration, improves the real-time performance and accuracy of traffic incident detection, optimizes system resource allocation, reduces costs, and decreases the false alarm rate, providing support for efficient highway management. Simultaneously, the present invention protects the large and small model collaborative architecture and model interaction mechanism.

[0051] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments 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 based on these drawings without creative effort.

[0053] Figure 1 A flowchart of the highway traffic incident detection method based on size model collaboration provided by the present invention;

[0054] Figure 2 The system architecture diagram of the highway traffic incident detection method based on size model collaboration provided by the present invention is shown below.

[0055] Figure 3 A schematic diagram showing the change in accuracy of abnormal parking events before and after the implementation of the highway traffic incident detection method based on big-small model collaboration provided by the present invention;

[0056] Figure 4 A schematic diagram showing the change in the total number of reported traffic incidents before and after the implementation of the highway traffic incident detection method based on size model collaboration provided by the present invention;

[0057] Figure 5 The logical block diagram of the highway traffic incident detection system based on size model collaboration provided by the present invention is shown. Detailed Implementation

[0058] Embodiments of the present invention are described in detail below, examples of which 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 the present invention, and should not be construed as limiting the present invention.

[0059] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0060] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0061] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0062] See Figure 1 and 2 This invention provides a method and system for detecting highway traffic incidents based on large-scale model collaboration, wherein the method includes the following steps:

[0063] Step S1: Data Acquisition. Through the rational planning and deployment of high-definition cameras and the establishment of transmission links, highway video data is continuously acquired and transmitted in real time.

[0064] In a preferred embodiment provided by the present invention, step S1 specifically includes:

[0065] Sub-step S11: Camera Deployment Planning. Based on the characteristics of the highway sections, the location of curves, and the distribution of entrances and exits, rationally plan the installation locations of high-definition cameras. Ensure that the camera's field of view covers key areas, such as accident-prone sections and traffic hubs, avoiding blind spots. Simultaneously, considering the differences in traffic flow across different road sections, appropriately increase the camera density in areas with higher traffic volume to obtain more comprehensive traffic information.

[0066] Sub-step S12: Camera Equipment Selection and Installation. Select a high-definition camera with high resolution (e.g., 18P or higher), wide dynamic range, and good low-light performance. High resolution helps to clearly capture details of traffic events, wide dynamic range allows for normal imaging in strong light or backlight conditions, and low-light performance ensures that effective video data can be obtained even at night or in low light. Install the camera according to the planned location, adjust the camera angle and focal length to ensure stable and clear image capture;

[0067] Sub-step S13: Establishing a data transmission link. Establish a reliable data transmission link for the video data captured by the camera. Prioritize wired network transmission, such as fiber optic cables, to ensure data transmission stability and high bandwidth. For areas where fiber optic cables cannot be laid, wireless transmission technologies, such as 5G, can be combined to achieve remote data transmission. Simultaneously, configure appropriate network equipment, such as switches and routers, and set reasonable network parameters to ensure that video data can be stably and quickly transmitted to the subsequent processing system.

[0068] Sub-step S14: Video data acquisition and real-time transmission. The camera continuously acquires video data at a set frame rate (e.g., 25fps or 3fps). The acquired video data is transmitted in real-time to the computing device where the small model resides through the established data transmission link, providing timely and accurate data support for subsequent traffic incident detection.

[0069] Step S2: Preliminary detection using a small model. Load the small model and preprocess the video data, detect traffic scenes, extract event segments, and generate preliminary detection results.

[0070] In a preferred embodiment provided by the present invention, step S2 specifically includes:

[0071] Sub-step S21: Model loading. When the system starts, a pre-trained small model, such as a lightweight YOLO series model, is loaded from the storage medium. At the same time, the configuration file required for the model to run is loaded. The configuration file contains the model's hyperparameters, class label information, etc., to ensure that the small model can run normally under the current hardware environment and detection task.

[0072] Sub-step S22: Video data input and preprocessing. The video data acquired in real time in step S1 is preprocessed according to a format acceptable to the small model. Preprocessing operations include adjusting the size of the video images to meet the requirements of the input layer of the small model; normalizing the images to map pixel values ​​to a specific range to accelerate model training and inference.

[0073] Sub-step S23: Traffic Scene Detection. Preprocessed video data is input frame-by-frame into the small model for inference operations. Feature extraction and analysis are performed on the traffic scenes in the video frames. During this process, the small model identifies various target objects in the image, such as pedestrians, vehicles, non-motorized vehicles, construction cones, and spilled materials—key traffic elements. Based on the identified traffic elements, post-processing judgments are performed. By comprehensively utilizing information such as the category, movement trajectory, and duration of traffic elements, events including but not limited to abnormal parking, pedestrian intrusion, non-motorized vehicle intrusion, motorcycle intrusion, occupation of emergency lanes, vehicles driving in the wrong direction, traffic congestion, construction areas, traffic accidents, and road debris can be determined.

[0074] Sub-step S24: Event Determination and Segment Extraction. When the small model detects a target that matches the preset traffic event type, the event determination mechanism is triggered. Based on the output of the small model, the location information of the event (such as the coordinate range of the target in the image) is determined, and a video segment containing the event is extracted from the video stream. The duration of the video segment can be set according to actual needs, generally including a short period before the event and a period after the event to provide more complete event information;

[0075] Sub-step S25: Preliminary detection result generation. For the extracted video clips, preliminary detection results are generated. The results include the event type, specifying which type of event it is, such as abnormal parking or non-motorized vehicle intrusion; the timestamp of the event is recorded, accurate to the second or even millisecond, for easy subsequent tracing; and the location information of the event is determined, which can be based on the relative position coordinates of the camera or the actual geographical location information obtained by combining with a map system.

[0076] Sub-step S3: Secondary confirmation of the large model. Event fusion and confirmation service ( Figure 1 The initial results of the small model are transmitted to the large model. After parsing, verifying and performing in-depth analysis, the large model feeds back to the event fusion and confirmation service, which then generates the final report.

[0077] In a preferred embodiment provided by the present invention, step S3 specifically includes:

[0078] Sub-step S31: Data transmission and reception. The preliminary detection results are sent to the large model via a high-speed network transmission channel. The large model receives the data and performs analysis and verification.

[0079] Sub-step S32: Deep Feature Extraction and Analysis. The large-scale model performs deep and comprehensive feature extraction on video clips, employing different strategies for different types of traffic events. Based on the extracted deep features, it makes accurate judgments about the events. Specifically, the large-scale model adopts a Transformer architecture, and its self-attention mechanism can correlate information globally. It can simultaneously and comprehensively analyze the position, direction of travel, collision severity, and other features of all involved vehicles, as well as the relationships between vehicles. Furthermore, the large-scale model learns the general features of various objects from massive image data. When detecting road traffic events, it can identify and judge road debris or complex scenes (such as traffic scenes in severe weather) that are less common in the training data more quickly and accurately.

[0080] Sub-step S33: Result Feedback and Fusion. If the large model confirms the event as a real traffic anomaly, the system uses the large model's judgment results to generate the final detection report;

[0081] Sub-step S4: Event Reporting and Handling. Based on the final report, determine the abnormal event, classify and report it to the management center, activate the contingency plan, and provide feedback records.

[0082] In a preferred embodiment provided by the present invention, step S4 specifically includes:

[0083] Sub-step S41: Abnormal Event Judgment and Classification. The system obtains the final detection report generated by the large model and determines whether it is an abnormal event based on the judgment results in the report. If it is determined to be an abnormal event, the event type is further classified, such as into emergency events (e.g., vehicle fire, serious traffic accident) and general events (e.g., ordinary vehicle malfunction, minor congestion), so that different handling strategies can be adopted subsequently;

[0084] Sub-step S42: Event reporting mechanism activated. For different types of abnormal events, the corresponding reporting process is initiated. For emergency events, a high-priority reporting method is used, such as through dedicated networks, SMS, and other channels, to ensure that event information reaches the event detection business system in the shortest possible time. For general events, event information can be reported to the management center through regular network interfaces, while ensuring the accuracy and completeness of information transmission.

[0085] Sub-step S43: Management Center Receives and Processes Information. The event detection system receives reported event information, parses and stores the received data. Based on the event type and severity, it activates the corresponding emergency plan. For example, for a vehicle fire, it immediately notifies the fire department and traffic police, and guides nearby vehicles to evacuate; for minor traffic congestion, it dispatches traffic control personnel to the scene to direct traffic, or publishes real-time road condition information through the traffic guidance system to guide vehicles to detour.

[0086] Sub-step S44: Incident Handling Feedback and Recording. During the incident handling process, collect progress information in real time and provide feedback to relevant departments and personnel. After the incident is completed, record the handling process and results in detail, including the incident's time, location, type, handling measures, and handling time, forming an incident handling file for subsequent data analysis and experience summarization to continuously optimize traffic incident handling procedures and emergency plans.

[0087] In a preferred embodiment of the present invention, a collaborative architecture for large and small models and a model interaction mechanism are designed to implement the above method.

[0088] Large and small model collaborative architecture: A highway traffic incident detection architecture based on the collaborative work of large and small models. The small model is responsible for the initial rapid detection, while the large model is responsible for the secondary accurate confirmation. Through the collaborative work of the two, the efficiency and accuracy of highway traffic incident detection are effectively improved.

[0089] Model interaction mechanism: After the small model detects a suspected event, it extracts video clips and generates a preliminary detection report, which is then sent to the large model through a high-speed network transmission channel; and the large model receives the data, performs analysis, deep feature extraction, and accurate judgment, and then feeds the results back to the system where the small model is located and integrates them to generate the final result.

[0090] The present invention also provides an embodiment to exemplarily demonstrate the effects of implementing the present invention.

[0091] The Hebei section of the Beijing-Xiong'an Expressway, a pilot project for building a strong transportation network, is an important component of the "four vertical and three horizontal" regional expressway network in Xiong'an New Area. Connecting Beijing's southwest Fifth Ring Road to the north and Xiong'an New Area to the south, this 75.498-kilometer-long expressway opened to traffic on May 29, 2021. The main line is an eight-lane, two-way expressway with a design speed of 120 kilometers per hour.

[0092] The incident detection system on the Beijing-Xiong'an Expressway was developed with the participation of multiple manufacturers, each with varying incident detection capabilities. This resulted in numerous false alarms and duplicate reporting of events, particularly for incident types such as abnormal parking and traffic accidents. To address this issue, improve traffic management efficiency, and ensure road safety and smooth traffic flow, the highway traffic incident detection method based on large and small model collaboration described in this patent was adopted to enhance the final detection results.

[0093] Data Acquisition: Utilizing the roadside AI cameras already deployed in the project, the acquired video data can be transmitted to the edge computing unit in real time at a frame rate of 25fps.

[0094] Preliminary detection using small models: Event detection is performed by small models integrated into the AI ​​camera and edge computing unit. The generated preliminary detection results include event type (e.g., abnormal parking), timestamp (accurate to milliseconds), and an accompanying video clip (approximately 1 minute).

[0095] Secondary Confirmation by the Large Model: The preliminary detection results generated by the small model are transmitted to the large model via a high-speed network. After receiving the data, the large model analyzes the video clips, employing targeted feature extraction strategies, focusing on analyzing features such as vehicle status and the surrounding environment. Through in-depth analysis, the large model confirms that this is a genuine abnormal parking event. Subsequently, the large model feeds back its judgment to the event fusion and confirmation service, and the system integrates the large model's results to generate the final detection report.

[0096] Incident Reporting and Handling: After the final detection report determines the incident to be an abnormal event, the system immediately reports the incident information to the incident detection business system. Upon receiving the information, the management center confirms the incident type and severity, and then activates the corresponding emergency plan. Nearby patrolling traffic police are dispatched to the scene, and simultaneously, information about the abnormal parking ahead is displayed on traffic guidance screens to remind drivers to slow down and be cautious. Upon arrival, the traffic police find the vehicle unable to start due to a malfunction, assist the driver in moving the vehicle to a safe area, and provide on-site warnings until the disabled vehicle is towed away and traffic returns to normal. The management center keeps a detailed record of the incident handling process, including the time of occurrence, reporting, and handling, the measures taken, and the on-site situation, to facilitate subsequent experience summarization and management process optimization.

[0097] like Figure 3Taking abnormal parking as an example, after adopting the highway traffic incident detection method based on big-small model collaboration, the detection accuracy increased from 66% to 87%.

[0098] like Figure 4 After adopting a highway traffic incident detection method based on large and small model collaboration, the number of reported incidents was significantly reduced.

[0099] Secondly, the present invention provides a highway traffic incident detection system based on large-scale model collaboration, comprising:

[0100] The video acquisition module 601 is used to continuously collect and acquire highway video data in real time through deployed high-definition cameras and established transmission links;

[0101] Video analysis and processing module 602 is used for:

[0102] Load a small model and preprocess video data to detect traffic scenes, extract event segments, and generate preliminary detection results;

[0103] The event fusion and confirmation service module 603 is used to transmit the preliminary results of the small model to the large model, so that the large model can provide event fusion and confirmation services after parsing, verification and in-depth analysis, and the video analysis and processing module generates the final report.

[0104] Output module 604 is used to: output the final report and classify the final report for submission to the management center, so that the management center can initiate contingency plan processing and provide feedback records.

[0105] In summary, this invention belongs to the field of intelligent transportation and relates to a method, system, device, and storage medium for highway traffic incident detection based on large and small model collaboration. Currently, highways mostly use a single small model to detect traffic incidents, which suffers from insufficient feature extraction capabilities and a high false alarm rate. This invention utilizes high-definition cameras along the highway to collect video data. The small model, with its fast computing power, performs preliminary detection, extracts event video segments, and generates preliminary detection results. The large model receives the data, performs parsing, deep feature extraction, and precise judgment. If it confirms a real event, it generates a final detection report. This method, through large and small model collaboration, improves the real-time performance and accuracy of traffic incident detection, optimizes system resource allocation, reduces costs, and decreases the false alarm rate, providing support for efficient highway management. Furthermore, this invention also protects the large and small model collaborative architecture and model interaction mechanism.

[0106] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0107] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0108] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0109] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A highway traffic incident detection method based on large-scale model collaboration, characterized in that, A highway traffic incident detection architecture is constructed that allows for collaborative operation of large and small models. After the small model detects a suspected incident, it extracts video clips and generates a preliminary detection report, which is then analyzed, deep feature extracted, and judged by the large model. The method includes: S1 continuously collects and transmits highway video data in real time by deploying high-definition cameras and establishing transmission links; S2 loads a small model and preprocesses video data to detect traffic scenes, extracts event video clips, and generates preliminary detection results. S3 transmits the event video clips and the preliminary detection results output by the small model to the large model through the event fusion and confirmation service module. The large model then analyzes and verifies the preliminary detection results and performs in-depth analysis, feeding them back to the event fusion and confirmation service module, which in turn generates the final report. S4 will submit the final report, categorized and submitted to the external management center, enabling the external management center to initiate contingency plans and provide feedback records; Step S2 includes: Sub-step S21: When the system starts, load the pre-trained small model from the storage medium and load the configuration file required for the model to run. The configuration file contains the model's hyperparameters and class label information. Sub-step S22: Preprocess the real-time acquired video data in a format acceptable to the small model. The preprocessing operations include adjusting the size of the video image to make the video data meet the requirements of the input layer of the small model; and normalizing the video image to map the pixel values ​​to a specific range. Sub-step S23: The preprocessed video image data is input frame by frame into the small model for inference calculation. Features are extracted and analyzed from the traffic scene in the video image frames. The small model identifies various target objects in the video image and performs post-processing judgment based on the identified traffic elements. Various events are judged by comprehensively utilizing the category, movement trajectory and duration information of traffic elements. These events include abnormal parking, pedestrian intrusion, non-motorized vehicle intrusion, motorcycle intrusion, occupation of emergency lane, vehicle driving in the wrong direction, traffic congestion, construction area, traffic accident and road debris. Sub-step S24: When the small model detects a target that matches the preset traffic event type, the event determination mechanism is triggered. Based on the output of the small model, the location information of the event is determined, and a video segment containing the event is extracted from the video stream. The duration of the video segment is set according to actual needs, including a short period before the event and a period after the event. Sub-step S25: Generate preliminary detection results for the extracted video segments. The preliminary detection results include the event type, record the timestamp of the event, and determine the location information of the event. The location information is based on the relative position coordinates of the camera or the actual geographical location information obtained by combining with the map system. Step S3 includes: Sub-step S31: Send the event video clip and preliminary detection results to the large model through a high-speed network transmission channel. After receiving the event video clip and preliminary detection results, the large model performs analysis and verification. Sub-step S32: The large model extracts features from the event video clips and adopts different strategies for different types of traffic events. It judges the event based on the extracted deep features. The large model adopts the Transformer architecture and uses the self-attention mechanism to associate information in the global scope. It analyzes the position, driving direction and collision degree features of all involved vehicles, as well as the interrelationship between vehicles. The large model learns the general features of various objects from the image data. Sub-step S33: If the large model confirms that the event is a real traffic anomaly, the system uses the judgment result of the large model to generate the final detection report.

2. The method according to claim 1, characterized in that, Step S1 includes: Based on the factors of highways, S11 plans the location of high-definition cameras and ensures that the camera's field of view can cover key areas; at the same time, considering the differences in traffic flow on different road sections, the camera density is appropriately increased in areas with high traffic flow. Factors affecting highways include: the characteristics of the highway sections, the location of curves, and the distribution of entrances and exits; Key areas include: accident-prone road sections and transportation hubs; S12 places cameras at the planned locations and adjusts the camera angles and focal lengths; S13 establishes a data transmission network for the camera through a wired network transmission link, and configures the corresponding network equipment and network parameters to enable the video data to be transmitted to the subsequent processing system. The S14 continuously collects video data using a high-definition camera at a set frame rate. The collected video data is then transmitted in real time to the computing device where the small model is located via a pre-established data transmission link, providing data support for subsequent traffic incident detection.

3. The method according to claim 2, characterized in that, Step S4 includes: S41 obtains the final detection report generated by the large model and determines whether it is an abnormal event based on the judgment results in the report; For emergency events, S42 adopts a high-priority reporting method, including ensuring that event information can be delivered to the event detection business system in the shortest possible time through dedicated network lines and SMS channels; for general events, event information is reported to the management center through a common network interface. S43 parses and stores the received data, and activates the corresponding emergency plan according to the type and severity of the event, including: for vehicle fire incidents, notifying the fire department and traffic police department, and guiding nearby vehicles to evacuate; for traffic congestion incidents, dispatching traffic control personnel to the scene to direct traffic, or releasing real-time traffic information through the traffic guidance system to guide vehicles to detour. During incident handling, S44 collects real-time progress information and provides feedback to external parties. After the incident is completed, it records the handling process and results in detail, including the time, location, type, handling measures, and handling time of the incident, forming an incident handling archive for subsequent data analysis and experience summarization, and continuously optimizing traffic incident handling procedures and emergency plans.

Images