Autonomous evolving agent method and system for traffic risk early warning and disposal

Abstract

This application discloses an autonomous evolved intelligent agent method and system for traffic risk early warning and handling, relating to the field of traffic risk early warning and collaborative accident handling technology. It employs a pre-trained traffic event recognition model to identify targets in video stream data, obtaining target recognition results. Based on these results, normalization results, risk scores, risk warning items, and handling suggestions are derived. Then, a first candidate event object is generated based on the above content. The first candidate event object is deduplicated to obtain a second candidate event object. The second candidate event object is sent to an instant messaging group for review, resulting in the target event object. The pre-trained traffic event recognition model is optimized based on the target event object. This improves the timeliness of accident detection, reduces duplicate alarms, enhances the consistency of risk assessment, shortens the on-site response chain, and strengthens the traceability and optimizability of the entire accident handling process.

Description

Technical Field

[0001] This application relates to the field of traffic risk early warning and accident collaborative handling technology, and in particular to an autonomous automated intelligent agent method and system for traffic risk early warning and handling. Background Technology

[0002] Highways are characterized by high vehicle speeds, continuous traffic flow, large accident impact areas, and a high risk of secondary accidents. If accidents are not detected and handled promptly after they occur, they can easily lead to rear-end collisions, congestion, risks to rescue operation areas, and reduced road capacity. Therefore, highway operation and management require timely detection, rapid confirmation, and closed-loop management of incidents such as accidents, abnormal parking, vehicle fires, lane obstruction, queuing congestion, and pedestrians or non-motorized vehicles entering the highway.

[0003] Existing highways typically deploy a large number of roadside cameras, event detection equipment, information boards, dispatch platforms, emergency command systems, and communication tools. Traditional methods mainly rely on manual monitoring, manual alarm reception, manual notification, and manual data entry. In recent years, deep learning, multimodal large-scale models, and traffic video analytics technologies have been used for traffic event detection. Some systems can identify events such as abnormal parking, congestion, accidents, fires and smoke, lane obstruction, and pedestrian intrusion from videos and generate prompts or alarms.

[0004] However, actual highway accident handling requires more than just accident identification; it also necessitates event deduplication, evidence extraction, risk classification, on-site personnel notification, manual confirmation, command center reassessment, resource allocation, handling records, database archiving, and subsequent strategy optimization. Therefore, simple video recognition or event alerts are insufficient to meet the comprehensive operational needs of highway command and control. Summary of the Invention

[0005] In view of this, this application provides an autonomous automated agent method and system for traffic risk early warning and handling, which is used to solve the above-mentioned problems in the prior art.

[0006] The objective of this application can be achieved through the following technical solutions: The first aspect of this application is to provide an autonomous automated intelligent agent method for traffic risk early warning and handling, characterized by comprising: Acquire video stream data, which includes video information; A pre-trained traffic incident recognition model is used to identify targets in video stream data, and the target recognition results are obtained. Based on the target identification results, normalization results and risk scores are determined, and risk warning items and handling suggestions are generated. The first candidate event object is generated based on video information, target recognition results, normalization results, risk score, risk warning items, and handling suggestions; The first candidate event object is deduplicated to obtain the second candidate event object; Send the second candidate event object to the instant messaging group for review, and obtain the target event object; The pre-trained traffic event recognition model is optimized based on the target event object.

[0007] In one optional embodiment, the target recognition result includes at least the following information: event stage information, accident visibility indication information, abnormal parking indication information, lane congestion indication information, vehicle congestion indication information, emergency rescue indication information, fire indication information, smoke indication information, personnel disembarkation indication information, rescue operation area exposure indication information, traffic restoration indication information, scene description information, and confidence level. The traffic event recognition model includes a target detection model and a visual language model. A pre-trained traffic event recognition model is used to perform target recognition on the video stream data to obtain the target recognition result, including: Input video stream data into a pre-trained object detection model and output object detection results; Based on the target detection results and preset rules, the status is determined and the following information is output: event stage information, accident visibility indication information, abnormal parking indication information, lane blockage indication information, vehicle congestion indication information, emergency rescue indication information, fire indication information, smoke indication information, personnel disembarkation indication information, rescue operation area exposure indication information, and traffic restoration indication information. The target detection results and preset text instructions are input into a pre-trained visual language model to generate image description information.

[0008] In one optional embodiment, the normalization result includes at least: rescue and response, suspected accidents, abnormal passage, and normal passage. The normalization result is determined based on the target recognition result, including: When the fire indication information and smoke indication information indicate the presence of fire and smoke and the emergency rescue indication information indicates that rescue has been carried out, or when the accident visibility indication information indicates the presence of an accident and the emergency rescue indication information indicates that rescue has been carried out, the normalized result is determined as the rescue and disposal. When the fire indication information and smoke indication information indicate the presence of fire smoke and the emergency rescue indication information indicates that no rescue has been carried out, or when the accident visibility indication information indicates the presence of an accident and the emergency rescue indication information indicates that no rescue has been carried out, the normalization result is determined to be a suspected accident. When the abnormal parking indicator indicates abnormal parking, the lane blockage indicator indicates lane blockage, or the vehicle blockage indicator indicates vehicle blockage, the normalization result is determined to be abnormal passage. When the abnormal parking indicator indicates that there is no abnormal parking, the lane blockage indicator indicates that the lane is not blocked, and the vehicle congestion indicator indicates that the vehicle is not congested, the normalization result is determined to be normal passage.

[0009] In one optional embodiment, determining a risk score based on the target identification result includes: The specific method for determining risk scores based on target identification results is as follows:

[0010]

[0011] in, Indicates overall risk. Indicates secondary risks. Indicates fire warning information. This indicates a smoke warning message. This indicates that the accident is visible. This indicates lane congestion information. Indicates emergency rescue instructions. This indicates a traffic congestion warning. This indicates information indicating where passengers should disembark. This indicates the exposure information in the rescue operation area. This indicates that traffic has resumed. , , , , 5, 6, , 2, 3, 4 represents the weighting coefficients.

[0012] In one optional embodiment, deduplication of the first candidate event object is performed to obtain a second candidate event object, including: Based on the video stream data time, camera number, road segment information, event type information, keyframe similarity, video feature similarity, image description similarity, and event stage continuity indication information, determine whether there is a common candidate event object among the first candidate event objects. If there is a common candidate event object among the first candidate event objects, then retain one candidate event object as the second candidate event object.

[0013] In one optional embodiment, a second candidate event object is sent to an instant messaging group for review to obtain the target event object, including: The second candidate event object is labeled based on the risk score to obtain the third candidate event object; The third candidate event object is sent to the instant messaging group for review, thus obtaining the target event object.

[0014] The second aspect of this application is to provide an autonomous automated intelligent agent system for traffic risk early warning and handling, characterized in that it includes: The acquisition module is used to acquire video stream data, which includes video information. The target recognition module is used to perform target recognition on video stream data using a pre-trained traffic event recognition model to obtain target recognition results; The first generation module is used to determine the normalization result and risk score based on the target identification result, and generate risk warning items and disposal suggestions; The second generation module is used to generate the first candidate event object based on video information, target recognition results, normalization results, risk scores, risk warning items, and handling suggestions; The deduplication module is used to remove duplicates from the first candidate event object to obtain the second candidate event object; The review module is used to send the second candidate event object to the instant messaging group for review of the second candidate event object and obtain the target event object; The optimization module is used to optimize the pre-trained traffic event recognition model based on the target event object.

[0015] A third aspect of this application is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the method as described in the first aspect.

[0016] A fourth aspect of this application is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the method as described in the first aspect.

[0017] The fifth aspect of this application is to provide a computer program product in which the computer program / instructions, when executed by a processor, implement the method as described in the first aspect.

[0018] Compared with existing technologies, the autonomous evolved intelligent agent method for traffic risk early warning and handling provided in this application uses a pre-trained traffic incident recognition model to identify targets in video stream data, obtaining target recognition results; based on the target recognition results, normalization results and risk scores are determined, and risk warning items and handling suggestions are generated; based on video information, target recognition results, normalization results, risk scores, risk warning items, and handling suggestions, a first candidate event object is generated; the first candidate event object is deduplicated to obtain a second candidate event object; the second candidate event object is sent to an instant messaging group for review, obtaining the target event object; and the pre-trained traffic incident recognition model is optimized based on the target event object. This improves the timeliness of accident detection, reduces duplicate alarms, enhances the consistency of risk assessment, shortens the on-site response chain, and strengthens the traceability and optimizability of the entire accident handling process. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating an autonomous intelligent agent method for traffic risk early warning and handling provided in an embodiment of this application; Figure 2 A structural block diagram of an autonomous intelligent agent system for traffic risk early warning and handling provided in an embodiment of this application; Figure 3 This is a structural block diagram of an electronic device for implementing an autonomous intelligent agent method for traffic risk early warning and handling, as provided in an embodiment of this application. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0022] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] It should be understood that in the embodiments of this application, "at least one" means one or more, and "more than one" means two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the related objects before and after it are in an "or" relationship. "Contains A, B and / or C" means containing any one, two, or three of A, B, and C.

[0024] It should be understood that in the embodiments of this application, "B corresponding to A", "B corresponding to A", "A corresponds to B" or "B corresponds to A" means that B is associated with A, and B can be determined based on A. Determining B based on A does not mean that B is determined solely based on A; B can also be determined based on A and / or other information.

[0025] To address the technical problems existing in related technologies, this application provides an autonomous automated agent method and system for traffic risk early warning and handling.

[0026] The autonomous evolved intelligent agent method for traffic risk early warning and handling provided in this application can be executed by an electronic device, such as a terminal or a server. The terminal can be a smartphone, tablet, laptop, or other similar device. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. It is understood that this application does not specifically limit the executing entity of the autonomous evolved intelligent agent method for traffic risk early warning and handling.

[0027] The technical solution of this application will be described in detail below through specific embodiments. It should be noted that the following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments described below are used to explain the technical solution of this application and are not intended to limit actual use.

[0028] To address the technical problems existing in related technologies, embodiments of this application provide an autonomous evolved intelligent agent method for traffic risk early warning and handling, such as... Figure 1 As shown, Figure 1 This is a flowchart illustrating an autonomous automated agent method for traffic risk warning and mitigation provided in an embodiment of this application. It should be noted that the steps shown may be executed in a different logical order than those shown in the flowchart. The method may include the following steps S101 to S107.

[0029] Step S101: Obtain video stream data.

[0030] In one alternative embodiment, video stream data can be acquired in real time from cameras deployed on the highway side, historical surveillance recordings can be retrieved from a hard drive, or video stream data can be obtained from a video surveillance management platform, etc. This application does not limit this.

[0031] It should be noted that video stream data includes video information.

[0032] In one optional embodiment, metadata such as the corresponding camera number, camera coverage section identifier, traffic flow direction, and camera timestamp can be bound to each video stream data.

[0033] In another alternative embodiment, video stream data is acquired according to camera number, road segment, direction, and time window to form independently detectable video segments; alternatively, a single video file can be directly used as a detection segment.

[0034] Step S102: Use a pre-trained traffic incident recognition model to perform target recognition on the video stream data to obtain the target recognition result.

[0035] In one optional embodiment, the target recognition result includes at least the following information: event stage information, accident visibility indication information, abnormal parking indication information, lane blockage indication information, vehicle congestion indication information, emergency rescue indication information, fire indication information, smoke indication information, personnel disembarkation indication information, rescue operation area exposure indication information, traffic restoration indication information, image description information, and confidence level.

[0036] It should be noted that the event phase information is used to determine the progress of accident handling, such as the accident occurrence phase, accident handling phase, and accident recovery phase. Accident visibility information indicates whether there is any visible collision, deformation, or damage. Abnormal parking information indicates whether a vehicle has unexpectedly stopped in a non-parking area. Lane obstruction information indicates whether a lane is blocked by an accident. Vehicle congestion information indicates the length of the traffic jam caused by the accident. Emergency rescue information indicates whether rescue has been received. Risk information indicates whether a fire has occurred. Scene description information describes the accident scene.

[0037] In one optional embodiment, the traffic incident recognition model includes an object detection model and a visual language model. The pre-trained traffic incident recognition model is used to perform object recognition on video stream data to obtain object recognition results, including: Input video stream data into a pre-trained object detection model and output object detection results; Based on the target detection results and preset rules, the status is determined and the following information is output: event stage information, accident visibility indication information, abnormal parking indication information, lane blockage indication information, vehicle congestion indication information, emergency rescue indication information, fire indication information, smoke indication information, personnel disembarkation indication information, rescue operation area exposure indication information, and traffic restoration indication information. The target detection results and preset text instructions are input into a pre-trained visual language model to generate image description information.

[0038] It should be noted that the preset text instructions are used to determine whether the descriptive information output by the model is biased towards safety warnings, traffic management, or event review.

[0039] In one specific embodiment, the preset text instruction may be a detailed description instruction for the target, a causal and state reasoning instruction for abnormal events, or a structured report instruction for a specific scenario, etc., and this application does not limit it in this way.

[0040] In a more specific embodiment, the preset text instruction input might be: "Based on the detection results, describe the fire situation of the truck and determine whether the safety protection measures of the on-site rescue personnel are in place."

[0041] In one optional embodiment, before using a pre-trained traffic event recognition model to perform target recognition on the video stream data and obtain the target recognition result, video segments within a preset time window before and after the trigger time point can be extracted when an anomaly is triggered. Video frames are then uniformly extracted from the video segments as keyframes according to a set ratio, for example, keyframes can be extracted at positions such as 10%, 35%, 60%, and 85% of the video length. It should be noted that the keyframes can be determined based on the camera frame rate, event type, and computing resources.

[0042] In one specific embodiment, for a real-time video stream, when an anomaly is triggered, video segments within a preset time window before and after the trigger time point are extracted. Specifically, this includes: using a lightweight target detection model to scan each frame of the video stream to identify whether there are preset key targets in the video frame, such as pedestrians, vehicles, and smoke; if the preset key targets are identified in the video frame, then video segments within a preset time window before and after the trigger time point are extracted.

[0043] In another optional embodiment, a pre-trained traffic event recognition model is used to perform target recognition on the video stream data to obtain target recognition results. This includes: inputting keyframes into a target detection model and outputting target detection results, which include object category, bounding box, and confidence level; and using a target tracking algorithm to track the target object; performing state determination based on the target detection results and preset rules, and outputting event stage information, accident visibility indication information, abnormal parking indication information, lane congestion indication information, vehicle congestion indication information, emergency rescue indication information, fire indication information, smoke indication information, personnel disembarkation indication information, rescue operation area exposure indication information, and traffic restoration indication information; and inputting the target detection results and preset text instructions into a pre-trained visual language model to generate scene description information.

[0044] In one specific embodiment, if two or more vehicle category targets are detected, and their bounding box intersection-union ratio (IU) is greater than a preset IU threshold, and their relative speed is close to 0, then a collision accident is determined. If an abnormal aspect ratio is detected in a vehicle category target, it is determined to be a rollover. If any of the above conditions are met, the accident visibility indication information is output as true, and the accident type is marked.

[0045] In another specific embodiment, the tracking target is a vehicle, and within N consecutive frames, its centroid coordinate displacement is less than a preset displacement threshold (e.g., 2 frames, 5 frames, or 10 frames). It is determined whether the target's position coordinates fall within a polygon mask of a preset non-parking area (e.g., driving lane, overtaking lane); and whether the average speed of surrounding vehicles is greater than a preset speed threshold (the preset speed threshold can be 0.07 m / s, 0.1 m / s, or 0.3 m / s). If the tracking target's centroid coordinate displacement is less than the preset displacement threshold, the target's position coordinates fall within the preset non-parking area, and the average speed of surrounding vehicles is less than the preset speed threshold, then an abnormal parking indication is output as true. It should be noted that the preset displacement threshold and preset speed threshold can be set according to actual conditions, and this application does not limit them.

[0046] In another specific embodiment, if a non-vehicle object (such as a fallen rock or cargo) or a stationary accident vehicle occupies a preset proportion of the lane width (the preset proportion can be 1 / 4, 1 / 3, or 1 / 2), lane congestion indication information is output. The vehicle density and average speed within the region of interest of a specific lane are calculated. If the vehicle density is greater than a preset vehicle density threshold (approximately 20-30 vehicles per kilometer per lane, etc.) and the average vehicle speed is less than a preset average vehicle speed threshold (e.g., 20 km / h, 25 km / h, or 32 km / h, etc.), vehicle congestion indication information is output. This vehicle congestion indication information can be classified as light, moderate, or severe. The specific classification of vehicle congestion indication information can be set according to actual conditions, and this application does not limit this.

[0047] In another specific embodiment, a target classified as smoke or flame is detected, and the confidence level is greater than a preset confidence threshold (the preset confidence threshold can be 80%, 85%, or 93%, etc.); and the Euclidean distance between the smoke target and the nearest vehicle target is calculated. If the distance is less than a preset distance threshold (the preset distance threshold can be 5 meters, 6 meters, or 8 meters, etc.), then it is determined that the vehicle is on fire, and the fire indication information and smoke indication information are output as true, and the fire source vehicle information is associated.

[0048] In another specific embodiment, in frame t-1, the person target is located within or adjacent to the vehicle target's bounding box; in frame t, the distance between the person target and the vehicle target is greater than a preset distance threshold, and the output of the person getting off the vehicle instruction information is true.

[0049] In another specific embodiment, if a predefined special vehicle category (ambulance, fire truck, police car, breakdown vehicle) is detected entering the accident scene area, an emergency rescue instruction message is output as true.

[0050] In another specific embodiment, upon detecting a rescue vehicle or an accident vehicle, a safety buffer zone is defined centered on the vehicle. Within this buffer zone, a scan is performed to check for targets such as safety cones or warning signs. If the number of safety cones is zero, or the distance between safety cones is less than a preset safety distance, it is determined that protection is lacking, and a rescue operation area exposure indication is output as true. It should be noted that the preset safety distance is determined based on actual traffic rules.

[0051] In another specific embodiment, if the accident vehicle or obstacle target that existed in the previous moment disappears (is dragged away) in the current frame, the average speed of vehicles in the lane is greater than a preset average vehicle speed threshold, and the traffic density is less than a preset traffic density threshold, the output of the resumption of traffic indication information is true, and the event end time is marked.

[0052] It should be noted that the object detection model can be YOLO-Pose or YOLO-Seg, or other object detection models; this application does not impose any limitations on this. When data is insufficient, calculating precise speed, dwell time, vehicle distance, or queue length is not mandatory.

[0053] In another specific embodiment, when a collision or sudden braking is detected, the event stage information is marked as the occurrence stage; when a rescue vehicle is detected, the event stage information is marked as the handling stage; and when towing operations or traffic flow is restored, the event stage information is marked as the recovery stage.

[0054] In another specific embodiment, the visual language model can be a model such as MiniCPM-V, LLaVA, or Qwen-VL, and this application does not limit it.

[0055] It should be noted that when lane line polygons or camera calibration results are lacking, only suspected lane occupancy, queuing, smoke, fire, emergency vehicles, and signs of recovery are identified; precise lane numbers are not mandatory. The visual language model can be deployed on a local server or a remote server.

[0056] In another optional embodiment, the training process of the object detection model is as follows: collect traffic monitoring videos, extract images containing scenes such as accidents, congestion, smoke, and rescue, label the object categories using a labeling tool, and train the object detection model using sample data labeled with object categories to obtain a pre-trained object detection model.

[0057] In another alternative embodiment, the training process of the visual language model is as follows: Construct image-text pairs, where the images are accident scene photos and the text is the corresponding accident descriptions. Use the image-text pairs to train a visual language model to obtain a pre-trained visual language model.

[0058] In one specific embodiment, information such as camera viewpoint, road segment, direction, weather prompts, manual notes, and optional areas of interest can also be used as visual language model prompts or post-processing context.

[0059] Step S103: Determine the normalization result and risk score based on the target identification result, and generate risk warning items and handling suggestions.

[0060] In one optional embodiment, the normalization result includes at least: rescue and response, suspected accidents, abnormal passage, and normal passage. The normalization result is determined based on the target recognition result, including: When the fire indication information and smoke indication information indicate the presence of fire and smoke and the emergency rescue indication information indicates that rescue has been carried out, or when the accident visibility indication information indicates the presence of an accident and the emergency rescue indication information indicates that rescue has been carried out, the normalized result is determined as the rescue and disposal. When the fire indication information and smoke indication information indicate the presence of fire smoke and the emergency rescue indication information indicates that no rescue has been carried out, or when the accident visibility indication information indicates the presence of an accident and the emergency rescue indication information indicates that no rescue has been carried out, the normalization result is determined to be a suspected accident. When the abnormal parking indicator indicates abnormal parking, the lane blockage indicator indicates lane blockage, or the vehicle blockage indicator indicates vehicle blockage, the normalization result is determined to be abnormal passage. When the abnormal parking indicator indicates that there is no abnormal parking, the lane blockage indicator indicates that the lane is not blocked, and the vehicle congestion indicator indicates that the vehicle is not congested, the normalization result is determined to be normal passage.

[0061] In another alternative embodiment, determining a risk score based on the target identification result includes: The specific method for determining risk scores based on target identification results is as follows:

[0062]

[0063] in, Indicates overall risk. Indicates secondary risks. Indicates fire warning information. This indicates a smoke warning message. This indicates that the accident is visible. This indicates lane congestion information. Indicates emergency rescue instructions. This indicates a traffic congestion warning. This indicates information indicating where passengers should disembark. This indicates the exposure information in the rescue operation area. This indicates that traffic has resumed. , , , , 5, 6, , 2, 3, 4 represents the weighting coefficients.

[0064] It should be noted that the weighting coefficients can be given empirically or calibrated based on historical manually confirmed data, false alarm samples, and missed alarm samples.

[0065] In another optional embodiment, risk warning items and handling suggestions are generated, including: When fire smoke is detected, it recommends contacting fire trucks; when a rescue operation area is detected, it generates a risk warning for the rescue operation area; when signs of restoring traffic are detected, it generates a transition prompt for restoring traffic.

[0066] In one specific embodiment, the proposed action could be to coordinate with fire trucks, dispatch patrol vehicles, suggest issuing or verifying abnormal warnings on information boards ahead, and continuously monitor for the risk of collisions with vehicles approaching from behind, etc. This application does not limit the scope of the action.

[0067] In another optional embodiment, the event type, stage, risk level, main risk warnings and handling suggestions are output by means of rule weights, machine learning classifiers, expert rule tables, road segment risk profiles, historical accident statistics, and fusion scoring of weather conditions and traffic conditions.

[0068] Step S104: Generate the first candidate event object based on video information, target recognition results, normalization results, risk score, risk warning items, and handling suggestions.

[0069] In one optional embodiment, the first candidate event object includes at least the following information: event number, video stream data corresponding to the event, camera number, road segment, event type, target recognition result, risk score, handling suggestion, and fields to be reviewed.

[0070] It should be noted that the event number is used for subsequent event queries. The "To be reviewed" field represents the marked content that needs to be reviewed.

[0071] In another optional embodiment, for cases where the model output is missing, inconsistent in type, or has low confidence, the missing items are filled in, the uncertainty marker is increased, and subsequent manual review is prompted.

[0072] Default rules include: missing Boolean anomaly signals are assigned the value false by default; missing event stages or risk levels are assigned "uncertain" or "pending review"; missing confidence levels are assigned a preset low confidence level; and fields with inconsistent formats are mapped to a unified enumeration value according to the field dictionary. When there are situations such as missing fields, low confidence, inconsistent judgments between different segments, or conflicts between model stages and anomaly signals, the system increases the uncertainty flag, marking the event as requiring manual review or increasing its review priority.

[0073] Step S105: Remove duplicates from the first candidate event object to obtain the second candidate event object.

[0074] In one optional embodiment, deduplication of the first candidate event object is performed to obtain a second candidate event object, including: Based on the video stream data time, camera number, road segment information, event type information, keyframe similarity, video feature similarity, image description similarity, and event stage continuity indication information, determine whether there is a common candidate event object among the first candidate event objects. If there is a common candidate event object among the first candidate event objects, then retain one candidate event object as the second candidate event object.

[0075] In one specific embodiment, the process sequentially determines whether the time difference is less than a preset time difference threshold based on the video stream data time, whether it is the same camera based on the camera number, whether it is the same road segment based on the road segment, whether it is the same event type based on the event type information, and determines keyframe similarity, video feature similarity, image description similarity, and event phase continuity. If two candidate events satisfy the following conditions: the video stream data time difference is less than a preset time difference threshold (the preset time difference threshold can be 5 seconds, 8 seconds, or 10 seconds, etc.), the camera number is determined to be the same camera, the road segment is the same road segment, the event type is the same event type, the keyframe similarity, video feature similarity, image description similarity, and event phase continuity meet preset repetition conditions, then they are merged into the same target accident event, and the new content is appended to the event timeline; otherwise, a new target accident event is created.

[0076] In a more specific embodiment, the preset repetition condition may be that the similarity is greater than a preset similarity threshold (the similarity threshold may be 0.85, 0.9, or 0.92, etc.), which can be set according to the actual situation, and this application does not limit it. The above-mentioned preset threshold is set according to the actual situation, and this application does not limit it.

[0077] In this step, candidate accident events are sorted according to camera, road segment, viewpoint, and video clip order, and the stages of suspected accidents, rescue and response, and traffic restoration in consecutive clips are linked together to form an event timeline. For repeated abnormal signals of the same event in adjacent clips, the duration, stage, and evidence frame are updated instead of generating multiple alarms repeatedly.

[0078] Step S106: Send the second candidate event object to the instant messaging group to review the second candidate event object and obtain the target event object.

[0079] In one optional embodiment, a second candidate event object is sent to an instant messaging group for review to obtain the target event object, including: The second candidate event object is labeled based on the risk score to obtain the third candidate event object; The third candidate event object is sent to the instant messaging group for review, thus obtaining the target event object.

[0080] In one specific embodiment, the second candidate event object is labeled based on the risk score to obtain the third candidate event object, including: If the risk score is higher than the preset risk score threshold, the field to be reviewed will be designated as requiring manual review.

[0081] It should be noted that the preset risk scoring threshold can be 85, 90 or 92 points, etc., and can be set according to the actual situation. This application does not set it.

[0082] In a more specific embodiment, when multiple consecutive segments maintain a high risk score, a high-priority event candidate is output; when the risk score is low but the uncertainty is high, a low-priority review candidate is output.

[0083] In another specific embodiment, a third candidate event object is sent to an instant messaging group for review to obtain the target event object. This includes: sending the third candidate event object to the instant messaging group in the form of a card message; on-site personnel or monitors clicking the card to enter an H5 micro-application or web workbench to view event details, video evidence, risk level, handling suggestions, and handling timeline; and performing operations such as confirmation, reversal, supplementing handling information, and archiving. The event status is updated according to manual operation, and the handling progress and case closure report are synchronized to the group.

[0084] In another specific embodiment, text messages with links, image messages, video messages, work order detail pages, voice broadcasts, or a combination of multiple channels are used for notification. The manual review terminal is not limited to H5 (mobile web micro-application) micro-applications or web workbenches; it can also be a mobile application, mini-program, command center large screen, desktop client, or an embedded page in an existing business system. This is as long as it can complete event viewing, confirmation, false alarm closure, field modification, handling status update, and archive write-back.

[0085] In another alternative embodiment, a document database, time-series database, object storage, data lake, search engine, message stream storage, or a combination of these storage methods can be used. The unified event object is not limited to JSON (Structured Data Interchange Format, JavaScript Object Notation) files; it can also be split into multiple table structures such as an event master table, evidence table, timeline table, resource table, operation log table, and archive summary table, or transmitted in the form of interface objects, message objects, database records, etc. As long as the AI ​​(Artificial Intelligence) initial judgment, video evidence, manual review, resource disposal, status transition, final conclusion, and archive records can be completely preserved, a closed-loop event accumulation can be achieved.

[0086] In a more specific embodiment, the event orchestration and notification service can send an alarm card upon the first detection of a high-risk event, a progress card after manual confirmation or task assignment, and a case closure summary card after access is restored or the event is archived. Each card includes an event number, title, summary, details page address, target page label, key information fields, and action buttons.

[0087] Specifically, the initial alert card is used to reach on-site personnel or monitors while an incident is still pending verification. Example fields include incident number, suspected incident, overall risk, initial detection time, monitoring range, road segment, model stage, main risk warnings, and suggested actions. Card buttons include "View Incident Details" and "Confirm Now".

[0088] The incident progress card is used to synchronize information such as dispatching, rescue arrival, lockdown, and information board posting after the incident enters the incident tracking stage. Example fields include current status, update time, incident resources, information board status, current action, key risk warnings, and next step. Card buttons include "View Incident Progress" and "Update Incident Status."

[0089] Case closure summary cards are used to remind personnel to verify key timelines, resources allocated, and final conclusions during incident recovery or archiving. Example fields include final conclusion, final risk, recovery or archiving time, archiving status, key resources, key risk warnings, follow-up actions, and a summary of conclusions. Card buttons include "View Case Closure Summary" and "View Full Timeline."

[0090] It should be noted that the page displays an event overview, risk list, handling timeline, dispatch resources, and archived briefings. In actual deployment, the page address can be configured based on the intranet or cloud environment and is not limited to a specific address.

[0091] In one specific embodiment, the H5 micro-application and web-based workbench are used to receive detailed links from Lark cards and provide on-site personnel, monitors, and command center staff with access to view event details, conduct manual confirmation, make reversals, update handling status, and archive information. This workbench includes an event overview page, a risk warning page, a dispatch resource page, a handling record page, and an archive briefing page.

[0092] The Event Overview page displays the event number, event type, road segment, camera range, initial detection time, keyframes, short video clips, model stage, normalization stage, and image description. The Risk Warning page displays the overall risk, secondary risks, major risk alerts, risk signals, and scoring reasons. The Dispatch Resources page displays the suggested quantity, vehicle number, estimated arrival time, actual arrival time, and status of resources such as patrol cars, fire trucks, and breakdown vehicles. The Handling Record page displays a timeline of events including manual confirmation, reversal of judgment, dispatch, arrival at the scene, lockdown, information board posting, fire handling, and restoration of traffic. The Archived Briefing page displays the final conclusion, final risk, handling resources, actions taken, restoration time, and archived status.

[0093] The web-based workbench writes manual operations back to the event database. Manual operations include confirming an event's validity, marking it as a false alarm, modifying the event type, changing the risk level, adding information about road occupancy, adding resources for handling the situation, updating the handling status, confirming the restoration of access, and archiving the data. Each operation generates an operation log, recording the operator, operation time, modified fields, the value before modification, the value after modification, and the reason for the modification.

[0094] Step S107: Optimize the pre-trained traffic event recognition model based on the target event object.

[0095] In one optional embodiment, feedback samples are constructed from archived events, false alarm closure records, manual field correction records, handling results, and archived briefings. The false alarms, missed alarms, field reclassifications, and handling suggestion adoption rates are statistically analyzed under different camera types, road segments, event types, time periods, and weather conditions to generate policy optimization candidates. These candidates require offline playback verification of historical samples and manual review. Once verified, they are released as new policy versions, retaining the version number, scope of effect, reason for change, verification metrics, and rollback entry point.

[0096] In another specific embodiment, methods such as model fine-tuning, model selection, model parameter updates, risk weight updates, camera context configuration updates, attention area updates, false alarm filtering rule updates, message frequency limiting strategy updates, handling suggestion template updates, or multi-strategy replay verification can be employed. Replay verification is not limited to offline historical sample sets; it can also employ shadow mode, gray-scale road segment verification, manual sampling verification, or simulation sample verification. Through feedback sample construction, effect verification, manual review, version release, and rollback management, the controllable optimization objective of this application can be achieved.

[0097] In a more specific embodiment, the model's initial judgment, rule inference, and final human conclusion are aligned to form a feedback sample that includes the original video clips, keyframes, model output, rule output, manually corrected fields, reasons for correction, handling results, and final business status. Statistical analysis is performed on false alarms, missed alarms, event type reclassifications, risk level reclassifications, road occupation field reclassifications, suggested action reclassifications, and handling suggestion adoption under different camera, road segment, event type, time period, and weather conditions. Based on the statistical results of these differences, executable optimization candidates are generated. The optimization of candidate strategies includes updating camera contextual hints, updating focus area hints, adjusting risk score weights, adjusting manual review thresholds, adding false alarm filtering rules, updating prompt word templates, updating notification frequency limiting rules, updating bulletin templates, and updating handling suggestion templates. The evolved candidates are played back offline on archived events, false alarm samples, and manually corrected samples to compare event recognition accuracy, false alarm rate, false negative rate, manual correction rate, alarm timeliness, and handling suggestion adoption rate before and after evolution. After the evolved candidates pass playback verification and manual review, they are released as new strategy versions, retaining the version number, effective date, applicable cameras or road sections, reason for change, verification indicators, and rollback entry point. This design differs from solutions that directly train or automatically replace models online, ensuring the controllability, traceability, and rollback capability of strategy changes in high-speed accident handling scenarios.

[0098] In another specific embodiment, the system accesses multi-view video clips of a highway section, with the monitoring range including both long-range and close-up views. The risk warning agent extracts keyframes from the video clips and identifies signs of fire and smoke appearing near a certain moment in the close-up view, outputting "accident visible is real" or "fire or smoke is real." The stage normalization module corrects the event stage to "suspected accident" based on this visual signal.

[0099] The system generates a candidate event with the event number "E-FIRE-20220522-001". Because this event includes a fire smoke signal, the risk grading module sets its overall risk to "extremely high", the event type to "suspected vehicle fire event", the main risk warning to "fire smoke risk warning", and the recommended action to "activate fire truck". The event orchestration and notification service writes the initial draft of the event to the event database, with the status "pending confirmation", and sends an initial alarm card to the instant messaging group with the title "[Extremely High Risk] Suspected Vehicle Fire Event Pending Review".

[0100] Monitors or on-site personnel can click "View Event Details" on the card to access an H5 page, where they can view keyframes, video clips, event descriptions, risk levels, and suggested actions. After manual verification confirms the event, the system updates the status to "Confirmed" or "In Progress" and sends a progress card. Dispatchers record the simultaneous dispatch of patrol cars and fire trucks, and the system records one fire truck, one patrol car, their estimated arrival time, and subsequent arrival time on the dispatch resource page. When fire trucks and rescue vehicles continue to appear in the video, the event stage updates to "Rescue and Disposal," and the main risk warning updates to "Risk Warning for Rescue Operation Area."

[0101] Subsequently, when video or manual feedback indicates that the fire is under control, the main response is complete, and traffic begins to resume, the event stage is updated to "Resumption of Traffic." Command center personnel verify key time points, response resources, and final conclusions in the web-based workbench and then perform the archiving operation. The system generates a case closure report card titled "[Resumption Completed] Suspected Vehicle Fire Incident Case Closure Report" and writes the complete closed-loop event record to the database. This record includes AI initial assessment, risk classification, Lark broadcast, manual confirmation, dispatch resources, actual actions taken, resumption of traffic time, archiving status, and subsequent evolutionary feedback samples.

[0102] In this embodiment, keyframes, short video clips before and after the accident, camera information, road segment information, occurrence time, model output, and scene description are automatically extracted from the high-speed video stream to form a verifiable video evidence package. This eliminates the need for manual reviewers to search for key scenes again from long videos, reducing review costs and improving accident confirmation efficiency.

[0103] Based on time windows, camera identifiers, road segment locations, event types, and evidence similarity, multiple-triggered candidate events are deduplicated and merged to form a unique target accident event. This can reduce duplicate alarms and duplicate dispatches for the same accident and reduce stage misjudgments caused by model field conflicts.

[0104] Instead of directly using the original stage fields from the model, the system normalizes the event stages based on visual signals such as accident visibility, fire smoke, abnormal parking, lane congestion, queuing, emergency vehicles, towing assistance, and signs of recovery. This generates an overall risk level, secondary risk level, key risk indicators, handling priorities, and resource recommendations. Based on the event status, it generates Feishu cards such as initial alarms, handling progress reports, and case closure summaries, pushing the event to on-site personnel. On-site personnel or monitors use H5 mini-applications or web-based workbenches to confirm, close false alarms, modify field judgments, update handling status, and archive the data. The AI ​​initial judgment, manual review, handling resources, actual actions taken, restoration of traffic, and archiving status are all uniformly written into the closed-loop event record. The identification results no longer stop at a single alarm but form a business closed loop that is accessible, verifiable, modifiable, and archiveable.

[0105] Based on manual confirmation results, false alarm closure records, field correction records, handling results, and archived reports in the event database, system optimization candidates are generated. These candidates do not directly modify online models or business rules. Instead, they undergo feedback sample construction, difference statistics, offline playback verification of historical samples, and manual review before being released as new policy versions, retaining the version number, scope of effect, verification metrics, and rollback entry point. This ensures the verifiability, auditability, and rollback capability of policy changes while continuously optimizing identification and handling strategies.

[0106] In another optional embodiment, this application provides an autonomous evolved intelligent agent method for traffic risk early warning and handling. This method acquires real-time video streams or segmented video clips; generates keyframes, short videos before and after the accident, and basic metadata; performs traffic incident identification, stage normalization, event aggregation, risk classification, and handling suggestions; transforms candidate events into traversable business events, and completes deduplication, frequency limiting, status transition, and message template selection; sends event cards to on-site personnel via instant messaging groups; enables event detail viewing, manual confirmation, reversal, handling status updates, and archiving entry through H5 micro-applications or web-based work platforms; utilizes an event library / business library to store the event master table, evidence, timeline, resources, operation logs, and archived reports; and generates optimized candidates based on archived events and manual feedback. A unified event object is first formed, then the process of event orchestration, manual review, handling recording, and archiving is initiated. This avoids the direct impact of a single model's misjudgment on the business status and ensures that every manual confirmation and reversal can be accumulated as a data source for subsequent strategy optimization.

[0107] Corresponding to the autonomous automated agent method for traffic risk early warning and handling provided in the embodiments of this application, the embodiments of this application also provide an autonomous automated agent system for traffic risk early warning and handling, such as... Figure 2 As shown, the autonomous evolving intelligent agent system for traffic risk early warning and handling includes: The acquisition module 201 is used to acquire video stream data, which includes video information; The target recognition module 202 is used to perform target recognition on video stream data using a pre-trained traffic event recognition model to obtain target recognition results; The first generation module 203 is used to determine the normalization result and risk score based on the target identification result, and generate risk warning items and disposal suggestions; The second generation module 204 is used to generate a first candidate event object based on video information, target recognition results, normalization results, risk scores, risk warning items, and handling suggestions; The deduplication module 205 is used to deduplicat the first candidate event object to obtain the second candidate event object; The review module 206 is used to send the second candidate event object to the instant messaging group to review the second candidate event object and obtain the target event object; The optimization module 207 is used to optimize the pre-trained traffic event recognition model based on the target event object.

[0108] Corresponding to the autonomous automated intelligent agent method for traffic risk early warning and handling provided in the embodiments of this application, the embodiments of this application also provide an electronic device for executing the autonomous automated intelligent agent method for traffic risk early warning and handling, such as... Figure 3 As shown, the electronic device includes: a processor 301; and a memory 302 for storing a program for an autonomous automated intelligent agent method for traffic risk warning and handling. After the device is powered on and the processor runs the program for the autonomous automated intelligent agent method for traffic risk warning and handling, the following steps are performed: Acquire video stream data, which includes video information; A pre-trained traffic incident recognition model is used to identify targets in video stream data, and the target recognition results are obtained. Based on the target identification results, normalization results and risk scores are determined, and risk warning items and handling suggestions are generated. The first candidate event object is generated based on video information, target recognition results, normalization results, risk score, risk warning items, and handling suggestions; The first candidate event object is deduplicated to obtain the second candidate event object; The second candidate event object is sent to an instant messaging group for review, thus obtaining the target event object; the pre-trained traffic event recognition model is then optimized based on the target event object.

[0109] Corresponding to the autonomous automated agent method for traffic risk warning and handling provided in the embodiments of this application, the embodiments of this application also provide a computer-readable storage medium storing a program for the autonomous automated agent method for traffic risk warning and handling, which is executed by a processor to perform the following steps: Acquire video stream data, which includes video information; A pre-trained traffic incident recognition model is used to identify targets in video stream data, and the target recognition results are obtained. Based on the target identification results, normalization results and risk scores are determined, and risk warning items and handling suggestions are generated. The first candidate event object is generated based on video information, target recognition results, normalization results, risk score, risk warning items, and handling suggestions; The first candidate event object is deduplicated to obtain the second candidate event object; Send the second candidate event object to the instant messaging group for review, and obtain the target event object; The pre-trained traffic event recognition model is optimized based on the target event object.

[0110] Corresponding to the autonomous evolved intelligent agent method for traffic risk early warning and handling provided in the embodiments of this application, the embodiments of this application also provide a computer program containing instructions, which, when executed by a computer, cause the computer to perform the following steps: Acquire video stream data, which includes video information; A pre-trained traffic incident recognition model is used to identify targets in video stream data, and the target recognition results are obtained. Based on the target identification results, normalization results and risk scores are determined, and risk warning items and handling suggestions are generated. The first candidate event object is generated based on video information, target recognition results, normalization results, risk score, risk warning items, and handling suggestions; The first candidate event object is deduplicated to obtain the second candidate event object; Send the second candidate event object to the instant messaging group for review, and obtain the target event object; The pre-trained traffic event recognition model is optimized based on the target event object.

[0111] It should be noted that for a detailed description of the autonomous automated intelligent agent system, electronic device, computer-readable storage medium and computer program product for traffic risk early warning and disposal provided in the embodiments of this application, please refer to the relevant description of the embodiments of the autonomous automated intelligent agent method for traffic risk early warning and disposal provided in the embodiments of this application, which will not be repeated here.

[0112] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

[0113] In a typical configuration, an electronic device includes one or more processors (Central Processing Units), input / output interfaces, network interfaces, and memory.

[0114] Memory may include non-persistent storage in computer-readable media, such as random access memory and / or non-volatile memory, like read-only memory or flash memory. Memory is an example of computer-readable media.

[0115] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable operations, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory, static random access memory, dynamic random access memory, other types of random access memory, read-only memory, electrically erasable programmable read-only memory, flash memory or other memory technologies, compact disc read-only memory, digital video disc or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include non-transitory computer-readable media, such as modulated data signals and carrier waves.

[0116] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, compact disc read-only memory, optical storage, etc.) containing computer-usable program code.

[0117] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

Claims

1. A method for autonomous automated intelligent agents for traffic risk early warning and handling, characterized in that, include: Acquire video stream data, wherein the video stream data includes video information; A pre-trained traffic incident recognition model is used to perform target recognition on the video stream data to obtain target recognition results; Based on the target identification results, normalization results and risk scores are determined, and risk warning items and handling suggestions are generated; A first candidate event object is generated based on the video information, target recognition results, normalization results, risk score, risk warning items, and handling suggestions; The first candidate event object is deduplicated to obtain the second candidate event object; The second candidate event object is sent to an instant messaging group for review, thereby obtaining the target event object; The pre-trained traffic event recognition model is optimized based on the target event object.

2. The autonomous evolved intelligent agent method for traffic risk early warning and handling according to claim 1, characterized in that, The target recognition result includes at least the following information: event stage information, accident visibility indication information, abnormal parking indication information, lane congestion indication information, vehicle congestion indication information, emergency rescue indication information, fire indication information, smoke indication information, personnel disembarkation indication information, rescue operation area exposure indication information, traffic restoration indication information, image description information, and confidence level. The traffic event recognition model includes a target detection model and a visual language model. The target recognition result obtained by using a pre-trained traffic event recognition model to perform target recognition on the video stream data includes: The video stream data is input into a pre-trained target detection model, which outputs the target detection results. Based on the target detection results and preset rules, the state is determined, and the event stage information, accident visibility indication information, abnormal parking indication information, lane blockage indication information, vehicle congestion indication information, emergency rescue indication information, fire indication information, smoke indication information, personnel disembarkation indication information, rescue operation area exposure indication information, and traffic restoration indication information are output. The target detection results and preset text instructions are input into a pre-trained visual language model to generate image description information.

3. The autonomous evolved intelligent agent method for traffic risk early warning and handling according to claim 2, characterized in that, The normalization result includes at least: rescue and response, suspected accidents, abnormal passage, and normal passage. The normalization result is determined based on the target identification result, including: When the fire indication information and smoke indication information indicate the presence of fire smoke and the emergency rescue indication information indicates that rescue has been carried out, or when the accident visibility indication information indicates the presence of an accident and the emergency rescue indication information indicates that rescue has been carried out, the normalized result is determined as rescue and disposal. When the fire indication information and smoke indication information indicate the presence of fire smoke and the emergency rescue indication information indicates that no rescue has been carried out, or when the accident visibility indication information indicates the presence of an accident and the emergency rescue indication information indicates that no rescue has been carried out, the normalization result is determined to be a suspected accident. When the abnormal parking indication information indicates abnormal parking, the lane blockage indication information indicates lane blockage, or the vehicle blockage indication information indicates vehicle blockage, the normalization result is determined to be abnormal passage. When the abnormal parking indication information indicates that there is no abnormal parking, the lane blockage indication information indicates that the lane is not blocked, and the vehicle blockage indication information indicates that the vehicle is not blocked, the normalization result is determined to be normal passage.

4. The autonomous evolved intelligent agent method for traffic risk early warning and handling according to claim 2, characterized in that, Based on the target identification results, a risk score is determined, including: The specific method for determining the risk score based on the target identification results is as follows: in, Indicates overall risk. Indicates secondary risks. Indicates fire warning information. This indicates a smoke warning message. This indicates that the accident is visible. This indicates lane congestion information. Indicates emergency rescue instructions. This indicates a traffic congestion warning. This indicates information indicating where passengers should disembark. This indicates the exposure information in the rescue operation area. This indicates that traffic has resumed. , , , , 5, 6, , 2, 3, 4 represents the weighting coefficients.

5. The autonomous evolved intelligent agent method for traffic risk early warning and handling according to claim 1, characterized in that, The step of deduplicating the first candidate event object to obtain the second candidate event object includes: Based on the video stream data time, camera number, road segment information, event type information, keyframe similarity in the video stream data, video feature similarity, screen description similarity, and event stage continuity indication information, it is determined whether there is a common candidate event object among the first candidate event objects. If there is a common candidate event object among the first candidate event objects, then one candidate event object is retained as the second candidate event object.

6. The autonomous evolved intelligent agent method for traffic risk early warning and handling according to claim 1, characterized in that, The step of sending the second candidate event object to an instant messaging group for review to obtain the target event object includes: Based on the risk score, the second candidate event object is marked to obtain the third candidate event object; The third candidate event object is sent to an instant messaging group for review, thereby obtaining the target event object.

7. An autonomous automated intelligent agent system for traffic risk early warning and handling, characterized in that, include: The acquisition module is used to acquire video stream data, which includes video information; The target recognition module is used to perform target recognition on the video stream data using a pre-trained traffic event recognition model to obtain the target recognition result; The first generation module is used to determine the normalization result and risk score based on the target identification result, and generate risk warning items and handling suggestions; The second generation module is used to generate a first candidate event object based on the video information, target recognition results, normalization results, risk score, risk warning items and handling suggestions; The deduplication module is used to deduplicat the first candidate event object to obtain the second candidate event object; The review module is used to send the second candidate event object to the instant messaging group to review the second candidate event object and obtain the target event object; The optimization module is used to optimize the pre-trained traffic event recognition model based on the target event object.

8. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the autonomous automated agent method for traffic risk warning and handling as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the autonomous automated agent method for traffic risk early warning and handling as described in any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the autonomous automated agent method for traffic risk early warning and handling as described in any one of claims 1-6.

Images