A method and device for monitoring vehicle violations based on edge and cloud collaboration

Abstract

This invention discloses a vehicle violation monitoring method based on edge and cloud collaboration, comprising: real-time acquisition of video streams of a designated area via cameras; real-time reception of the video streams by an edge computing device, performing target detection on vehicles within the streams, and determining whether a preset event has been triggered based on the vehicle's status; if a preset event is triggered, capturing the corresponding video slice and uploading it to a central server; the central server performing semantic analysis on the video slices to generate a structured sequence of vehicle behavior events, matching the sequence of vehicle behavior events with preset violation rules; if a match is successful, determining that a violation has occurred, and generating an alarm message. This achieves a leap from "target detection" to "behavior understanding," facilitating traffic management.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and urban public safety technology, and in particular to a method and device for monitoring vehicle violations based on edge and cloud collaboration. Background Technology

[0002] Currently, common regulatory methods for addressing violations by taxis or ride-hailing vehicles in transportation hub areas mainly include: manual patrols: traffic enforcement officers conduct on-site inspections and issue verbal warnings or penalties upon discovering violations. This method relies on manpower, has limited coverage, is inefficient, and is easily affected by subjective factors.

[0003] Alternatively, fixed cameras can be used in conjunction with manual review. Surveillance cameras are deployed in key areas, and staff can retrieve and review the footage afterward. While this method preserves evidence, it suffers from issues such as delayed response, high labor costs, and difficulty in achieving 24 / 7 monitoring.

[0004] Therefore, there is an urgent need to propose a new method. Summary of the Invention

[0005] To address at least one of the aforementioned technical problems, this disclosure proposes a vehicle violation monitoring method based on edge and cloud collaboration in its first aspect. The method includes: real-time acquisition of video streams from a designated area via a camera; real-time reception of the video streams by an edge computing device, performing target detection on vehicles within the streams, and determining whether a preset event is triggered based on the vehicle's status; if a preset event is triggered, capturing the corresponding video slice and uploading it to a central server; the central server performing semantic analysis on the video slices to generate a structured sequence of vehicle behavior events, matching the vehicle behavior event sequence with preset violation rules; if a match is successful, determining that a violation exists and generating an alarm message.

[0006] Preferably, the preset events include: detecting that a vehicle door is open, or detecting that a vehicle has been continuously parked in a designated area for a period of time exceeding a preset duration threshold.

[0007] Preferably, when the preset event is door opening, the video slice is extracted from consecutive preset time periods before and after the door opening time; when the preset event is continuous parking timeout, the video slice is extracted from the moment the vehicle is first determined to be parked to the current trigger time.

[0008] Preferably, the vehicle behavior event sequence is structured data arranged in chronological order, and each event record includes behavior type, timestamp, vehicle location coordinates and license plate recognition result.

[0009] Preferably, the preset violation rules include at least one of the following: passengers get on or off the vehicle in a non-designated drop-off area; the vehicle stops for more than a preset threshold and no passengers get on or off; the 0112 door is closed after the door is opened without any passengers approaching or getting on or off; the number of passengers getting on or off during a single stop exceeds a preset number.

[0010] Preferably, when performing semantic analysis on video slices, the central server calls the visual big model and inputs system prompt instructions to the visual big model before calling it. The system prompt instructions are used to constrain the visual big model to output a structured format that conforms to the vehicle behavior event sequence.

[0011] Preferably, the system prompt instruction includes output format constraints, requiring the visual large model to return results in JSON format, and the fields include: timestamp, behavior category, bounding box coordinates, and license plate number.

[0012] Preferably, the edge computing device uses the YOLOv8 object detection model to detect vehicles in the video stream and determines whether they are ride-hailing vehicles based on the vehicle's color, roof light markings, and vehicle type characteristics; the visual large model is the Qwen2.5-VL multimodal large model, deployed on the GPU-accelerated computing node of the central server; the alarm information includes evidence video links; the alarm information includes alarm ID, time of violation, geographical location, license plate number of the vehicle involved, violation type, and storage link of the original video slice.

[0013] This disclosure, in its second aspect, proposes a vehicle violation monitoring device based on edge and cloud collaboration, comprising: a data acquisition device for real-time acquisition of video streams from a designated area via a camera; a video slice extraction device for receiving the video stream in real-time via an edge computing device, performing target detection on vehicles within it, and determining whether a preset event is triggered based on the vehicle's status; if a preset event is triggered, the corresponding video slice is extracted and uploaded to a central server; and an alarm information generation device for the central server to perform semantic analysis on the video slices, generate a structured sequence of vehicle behavior events, match the sequence of vehicle behavior events with preset violation rules, and if a match is successful, determine that a violation has occurred and generate alarm information.

[0014] In a third aspect, this disclosure provides a computer-readable medium storing a computer program that is loaded and executed by a processing module to implement the steps of any of the methods described above.

[0015] Some of the technical advantages of this disclosure are as follows: This invention provides a vehicle violation monitoring method based on edge and cloud collaboration. Through an architecture design of "lightweight edge detection and deep semantic parsing in the cloud," it achieves efficient, accurate, and automated identification of taxi violations. It significantly reduces the false positive rate and improves the accuracy of passenger pick-up and drop-off behavior recognition; shortens video analysis latency and supports near real-time response; achieves a leap from "target detection" to "behavior understanding," outputting structured event chains; reduces bandwidth and computing resource consumption, supporting large-scale deployment; and constructs a complete violation determination and evidence preservation mechanism to serve traffic law enforcement. Attached Figure Description

[0016] To better understand the technical solutions of this disclosure, the following accompanying drawings, which are used to assist in the illustration of the prior art or embodiments, can be referred to. These drawings selectively illustrate the products or methods involved in the prior art or some embodiments of this disclosure. The basic information of these drawings is as follows: Figure 1 This is a flowchart of an embodiment of a vehicle violation monitoring method based on edge and cloud collaboration according to this application. Detailed Implementation

[0017] The following will further describe the technical means or effects involved in this disclosure. Obviously, the provided embodiments (or implementation methods) are only some of the implementation methods covered by this disclosure, and not all of them. Based on the embodiments in this disclosure and the explicit or implicit descriptions in the figures and text, all other embodiments that can be obtained by those skilled in the art without creative effort will be within the scope of protection claimed in this disclosure.

[0018] With the acceleration of urbanization and the growth of travel demand, taxis and ride-hailing services, as important components of urban public transportation, undertake a large number of short-distance shuttle tasks in densely populated areas such as airports and train stations. However, due to the lack of efficient and real-time supervision, taxi drivers often engage in violations such as picking up and dropping off passengers in undesignated areas, parking illegally on sidewalks for extended periods, illegally carpooling, and refusing passengers, which seriously affect passenger safety and traffic order.

[0019] Traditional monitoring methods primarily rely on manual patrols or traditional video surveillance. While these methods have achieved some success, they suffer from high false alarm rates, significant processing delays, and weak behavioral understanding capabilities in complex and dynamic scenarios. In particular, when vehicle behavior is complex, traditional video surveillance often fails to determine whether the vehicle's actions are in violation of regulations, leading to inaccurate judgments and rendering the monitoring ineffective.

[0020] To address at least one of the above-mentioned problems, this disclosure provides a vehicle violation monitoring method based on edge and cloud collaboration. It achieves lightweight target detection and event triggering at the edge, utilizes a large visual model in the cloud for semantic analysis of key video segments, and combines a rule engine to accurately determine violations. This not only improves recognition accuracy and response speed but also achieves semantic-level understanding, providing a new technical path for intelligent traffic supervision. It is applicable to traffic order management in high-density pedestrian scenarios such as airports, train stations, and long-distance bus stations, enabling automatic identification, semantic analysis, and violation determination of behaviors such as taxi or ride-hailing vehicle stops, passenger pick-up and drop-off, and overstaying, thus improving the intelligence, precision, and automation of urban traffic supervision. It can efficiently, accurately, and automatically identify violations by taxis or ride-hailing vehicles, significantly reducing the false positive rate and improving the accuracy of passenger pick-up and drop-off behavior recognition; shortening video analysis latency to achieve near real-time response; not only detecting vehicle behavior but also understanding it and outputting structured event chains; reducing bandwidth and computing resource consumption, enabling large-scale deployment; and constructing a complete violation determination and evidence preservation mechanism to serve traffic enforcement.

[0021] like Figure 1 This disclosure provides a method for monitoring vehicle violations based on edge and cloud collaboration, characterized by comprising: S10: Real-time video stream capture of a designated area via camera; S20: The edge computing device receives video streams in real time, performs target detection on vehicles in the streams, and determines whether a preset event is triggered based on the vehicle's status. If a preset event is triggered, the corresponding video slice is captured and uploaded to the central server. S30: The central server performs semantic analysis on the video slices, generates a structured sequence of vehicle behavior events, matches the sequence of vehicle behavior events with preset violation rules, and if the match is successful, it determines that there is a violation and generates an alarm message.

[0022] In one embodiment, the present invention provides a vehicle violation monitoring device based on edge and cloud collaboration, comprising: a data acquisition device for real-time acquisition of video streams of a designated area via a camera; a video slice extraction device for real-time reception of the video stream by an edge computing device, target detection of vehicles within the stream, and determination of whether a preset event is triggered based on the vehicle's status; if a preset event is triggered, the corresponding video slice is extracted and uploaded to a central server; and an alarm information generation device for semantic analysis of the video slices by the central server to generate a structured sequence of vehicle behavior events, matching the sequence of vehicle behavior events with preset violation rules; if a match is successful, a violation is determined to exist, and an alarm information is generated.

[0023] The aforementioned cameras are deployed in transportation hub areas to continuously collect video streams from designated areas. The edge computing devices are located in edge computing boxes near the cameras or within a server room. These devices incorporate lightweight target detection models, facilitating real-time detection and analysis of the video streams received from the cameras and determining whether preset events have been triggered. The central server is deployed in the cloud or a local data center, and it houses a large-scale visual model analysis engine, a violation determination engine, and a violation event database. The central server can also send violation alerts to enforcement terminals, which can be PCs or mobile devices used by traffic management personnel. These terminals can view alert information, retrieve video evidence, and perform enforcement actions. The cameras, edge computing devices, central server, and enforcement terminals are connected via a local area network or dedicated network, forming a collaborative distributed system.

[0024] S10: Real-time video stream capture of a designated area via camera; In one embodiment, video capture and transmission are performed first. High-definition cameras deployed in the transportation hub area continuously capture video streams of a designated area and transmit the video streams to edge computing devices in real time. Multiple cameras cover the entire designated area, ensuring no blind spots. The designated area can be a passenger pick-up and drop-off area.

[0025] S20: The edge computing device receives video streams in real time, performs target detection on vehicles in the streams, and determines whether a preset event is triggered based on the vehicle's status. If a preset event is triggered, the corresponding video slice is captured and uploaded to the central server. Edge computing devices are also installed in transportation hub areas, and can be integrated into light poles within the hub area. Multiple edge computing devices can be installed. At least one edge computing device receives the video stream from the aforementioned camera, which outputs an H.264 / H.265 encoded video stream. The edge computing device decodes the video stream using GStreamer, FFmpeg, or a dedicated SDK. A video resolution greater than 1280×720 (720P) is recommended to ensure accurate vehicle and door recognition. The edge computing device performs target detection on the vehicles in the video stream and determines whether a preset event is triggered based on the vehicle's status. In one embodiment, the preset events include: detecting that a vehicle door is open, or detecting that a vehicle has been continuously parked in a designated area for more than a preset duration threshold. The preset events include two typical scenarios: one is the door opening event: when the edge computing device detects a change in the shape of the vehicle's side door area, such as widening of the door gap or displacement of the door, and this change lasts for more than 2 frames, it is determined as "door opening"; the other is the timed-out stopping event: when a vehicle remains stationary in a designated passenger drop-off area for more than a preset duration threshold (e.g., 300 seconds), a timed-out stopping event is triggered. A stationary state can be determined by the displacement between consecutive frames being less than the threshold. These events serve as trigger conditions for video slice generation, effectively filtering out key video segments requiring in-depth analysis.

[0026] The vehicle's status includes the shape of the vehicle's side door area, the vehicle's position information, displacement, and other information. Specifically, the edge computing device detects the vehicle in the video stream in real time, including whether the vehicle has entered the designated passenger drop-off area and whether the door is open or closed. In other words, the video stream reception process and the subsequent two detections are executed in parallel, forming a pipelined process to ensure low latency.

[0027] A lightweight object detection model, such as YOLOv8n or YOLOv8s from the YOLOv8 series, runs on an edge computing device. This model performs frame-by-frame detection and analysis on the video stream received from the camera, processing 10 to 15 frames per second. Within each received frame or every few frames, it first identifies whether a taxi or other monitored vehicle appears in a designated area of ​​the video, obtaining information such as the vehicle's category, location, and confidence score representing the object detection model's assessment of the detection result. Specifically, the object detection model identifies objects in the video stream's image frames. When a vehicle is detected and its bounding box's center point is within a designated area, the detected vehicle is determined to have entered that area. Vehicle category information can be pre-trained using features such as color, roof lights, and vehicle type. When the confidence score of the category model detecting a taxi is greater than a preset threshold, such as 0.8, the detected vehicle is determined to be a taxi. Whether a taxi has entered the designated drop-off area: This is determined by using a pre-defined Region of Interest (ROI). Entry is considered valid only if the detection confidence level is ≥ 0.8.

[0028] Once a vehicle is confirmed to have entered a designated area, such as a drop-off area, the edge computing device initiates continuous status monitoring of the vehicle. Vehicle status includes whether the vehicle is stationary (parked), whether the doors are open, and the duration of stationary or parked status. Preset events are key behaviors constituting violations; for example, a preset event could be a door opening or exceeding the permitted parking time. The system determines whether a preset event is triggered based on the vehicle's status, i.e., it tracks and monitors the opening and closing status of vehicle doors and the parking time within the designated area, using a model to perform fine-grained detection on the vehicle's side door area to determine whether the door is open. It can process approximately 15 frames per second, balancing accuracy and efficiency. A preset event is triggered when two or more consecutive frames detect a "door open" status for the vehicle, and the confidence level is greater than a preset threshold.

[0029] Alternatively, when a vehicle is detected to be stationary (parked) within a specified area, the parking time of the vehicle is tracked and recorded. The time when the vehicle is first identified as parked is recorded, and the vehicle is continuously monitored for movement through inter-frame displacement or speed. If the parking time exceeds a preset threshold, it is also determined that a preset event has been triggered.

[0030] When any preset event is triggered, the edge computing device immediately captures a video segment containing the complete vehicle behavior process, known as a "video slice," and uploads it to the central server via the network for in-depth analysis. When the preset event is a door opening, the video slice captures continuous video over preset time periods before and after the door opening moment. When the preset event is a continuous parking timeout, the video slice captures continuous video from the moment the vehicle is first identified as parked to the current trigger moment. When the trigger event is a door opening, upon detection of the door opening, a video slice is captured over preset time periods before and after the door opening moment and uploaded to the central server. For example, a 30-second video slice is captured, centered on the door opening moment, capturing 15 seconds before and 15 seconds after the opening moment. This video slice includes the entire process of the vehicle parking, door opening, passenger boarding and alighting, door closing, and vehicle departure. When the triggering event is a timed-out stop, a continuous video segment is captured from the initial stop to the current time of the timeout determination and uploaded to the central server. This video segment completely records the entire process of the vehicle's timed-out stop, including the process from the vehicle's initial stop in the drop-off area to its continued stop until the timeout period expires. This ensures a complete record of the lane occupancy duration. The video segment is uploaded to the central server immediately after generation to facilitate subsequent video processing. This method ensures the integrity of the evidence video while minimizing the upload of invalid data. All video segments include metadata, including: trigger timestamp, camera ID, license plate number (if identifiable), and preliminary detection results such as the triggering event "door opened." Uploading only the generated video segments saves bandwidth consumption and facilitates subsequent processing. The generated video segments and their metadata are uploaded to the central server via an encrypted channel. Because only key segments are uploaded, bandwidth consumption is reduced by more than 80% compared to uploading the entire video.

[0031] S30: The central server performs semantic analysis on the video slices, generates a structured sequence of vehicle behavior events, matches the sequence of vehicle behavior events with preset violation rules, and if the match is successful, it determines that there is a violation and generates an alarm message.

[0032] The central server performs semantic analysis on video slices, specifically calling upon a large-scale visual model deployed on it for deep semantic understanding to accurately determine whether traffic violations exist. Upon confirmation of a violation, it automatically generates structured alert information for user use. The central server deploys the Qwen2.5-VL multimodal large-scale model, which possesses powerful video-language joint understanding capabilities, supporting input of short video clips and outputting behavioral descriptions. The model runs on compute nodes equipped with NVIDIA A100 GPUs, accelerated by the vLLM inference framework. The analysis time for a single 30-second video slice is ≤ 8 seconds, meeting near real-time monitoring requirements.

[0033] First, the central server receives and preprocesses video slices uploaded from edge computing devices. These video slices can be 30-second MP4 files along with accompanying metadata such as the trigger event type, camera ID, and timestamp. The central server decodes the received video slices into image frame sequences, for example, at 25fps; it also enhances low-light and blurry frames using techniques like CLAHE and noise reduction; and extracts keyframes at 2-5 frames per second to reduce the load on the large model. Second, the central server calls the large visual model for semantic analysis. This multimodal visual model deployed on the central server can be Qwen2.5-VL, InternVL, or LLaVA, etc. Then, a structured event chain is constructed, and the output of the large visual model is format-validated and standardized, checking the integrity of fields such as time, behavior, and coordinates; a unified time base is established using the video start time as 00:00:00; multiple "door opening" events are merged into a single continuous identical behavior; and license plates are associated with vehicle trajectories. A standardized "vehicle behavior event sequence" object is output as the sole basis for violation determination. Semantic analysis of video clips here means understanding what actions the vehicle has taken, such as parking, opening doors, passengers getting on, and the vehicle driving away. All actions with semantic meaning are called "vehicle behavior events".

[0034] When performing semantic analysis on video segments, the central server invokes a large-scale visual model. Before invoking, it inputs system prompts into the large-scale visual model, constraining its output to conform to a structured format that follows a vehicle behavior event sequence. The system prompts include output format constraints, requiring the large-scale visual model to return results in JSON format, with fields including: timestamp, behavior category, bounding box coordinates, and license plate number. By injecting preset system prompts before invoking the large-scale visual model, the central server forces it to output structured results that conform to the specifications. Through this mechanism, the free text output of large models such as Qwen2.5-VL is constrained to a machine-readable JSON format, avoiding the problem of uncontrollable AI output.

[0035] The vehicle behavior event sequence is structured data arranged in chronological order. Each event record includes the behavior type, timestamp, vehicle location coordinates, and license plate recognition result. The vehicle behavior event sequence is a list of structured data, arranged in chronological order, and each record contains the following fields: Behavior type: An enumerated value, including "vehicle stop", "vehicle departure", "door open", "door close", "passenger boarding", "passenger alighting"; Timestamp: In the format of HH:MM:SS or HH:MM:SS–HH:MM:SS (for continuous behaviors); Vehicle location coordinates: Represented as a bounding box [x1, y1, x2, y2], based on a 1280×720 resolution image; License plate recognition result: If recognizable, the standard license plate number is output (such as "粤B12345"), otherwise it is an empty string. This structure ensures that the subsequent rule engine can be directly parsed without natural language processing. Finally, after obtaining the vehicle behavior event sequence, the central server inputs it into the violation determination module; this violation determination module performs logical matching based on the preset violation rule library; once the match is successful, it is determined as a violation and an alarm information record is generated and pushed to the law enforcement terminal. The preset violation rule library includes, but is not limited to, the following situations: Boarding or alighting passengers in a non-designated area: There is a "passenger boarding" or "alighting" behavior in the event sequence, but the vehicle coordinates are not within the preset alighting area ROI; False boarding or alighting: There is a "door open" behavior, but there is no subsequent "passenger boarding / alighting" record; Overtime stop: The stop duration > 300 seconds and there is no valid boarding or alighting behavior; Illegal carpooling: During a single stop, the number of "passenger boarding" or "alighting" times ≥ 2, and there is no record of filing. The rules are stored in the rule engine in the form of logical expressions and support dynamic updates.

[0036] Execute the violation rule matching. The built-in violation rule engine of the central server compares the above-structured vehicle behavior event sequence with the preset rules one by one to determine the violation behavior of the vehicle, and generates an alarm message when it is determined as a violation. The preset violation rules include at least one of the following: Passengers board or alight in a non-designated alighting area; The vehicle stops for more than the preset threshold and no passengers board or alight; The door closes without passengers approaching or boarding after it is opened; The number of passengers boarding or alighting during a single stop exceeds the preset quantity. Typical violation rules include boarding or alighting in a non-designated area, where there is a "passenger boarding" or "alighting" behavior, but the vehicle coordinates are not within the preset alighting area ROI; Overtime stop, where the time difference between "vehicle stop" and "vehicle departure" > the preset threshold, such as 300 seconds; False boarding or alighting, where there is a "door open", but there is no "passenger boarding or alighting" behavior, perhaps the driver gets off to smoke; Illegal carpooling, where during a single stop, the number of "passenger boarding" or "alighting" times is not less than 2, and there is no carpooling record; Refusal to carry passengers, where after the door is opened, a passenger approaches but does not board, and the vehicle quickly closes the door and drives away.

[0037] When the central server performs semantic analysis on video slices, it generates a structured event chain, namely the "vehicle behavior event sequence". Based on this "vehicle behavior event sequence", it performs violation judgment, and after confirming a violation, it saves the relevant video evidence to the database.

[0038] For vehicle behavior events determined to be in violation, the system automatically performs the following operations: storing event chain data, original video slices, camera information, etc. in the violation event database; generating an alarm record, including time, location, license plate number, and violation type; pushing an alarm notification to the law enforcement terminal through a message queue or API interface. Staff can view the alarm details through the law enforcement terminal, retrieve video evidence, and make corresponding handling after verification, such as warning, fine, scoring, etc.

[0039] When a violation is determined, the system automatically generates structured alarm information, including the following fields: alarm_id: unique alarm number (such as ALM20251203001); timestamp: violation occurrence time (UTC+8); location: geographical description (such as "Taxi Drop-off Area A, Terminal 3, XX Airport"); license_plate: license plate of the involved vehicle; violation_type: violation type (such as "Overtime Parking"); evidence_url: HTTPS link of the original video slice in object storage. This alarm information is pushed to the law enforcement terminal APP through the MQTT protocol and stored in the violation database for traceability.

[0040] In an embodiment, taking the taxi drop-off area of an airport as an example: A taxi with a license plate of "粤B67890" drives into the designated area and stops. The edge computing device detects the opening of the car door, triggers the video slices in the range of [-15s, +15s], and uploads them. Qwen2.5-VL analyzes the video and outputs the "vehicle behavior event sequence": "00:02:10 Vehicle stops → 00:02:12 Car door opens → 00:02:15-00:02:22 Passenger gets on". The system determines it as a compliant behavior and does not give an alarm. Another vehicle stops and the car door opens, but there is no passenger getting on or off, and it stays for 6 minutes. The edge computing device triggers the overtime slice, and Qwen2.5-VL confirms that there is no getting on or off behavior. The system determines it as "False getting on or off + Overtime parking", generates an alarm, and saves the evidence.

[0041] By leveraging a large-scale visual model to deeply understand behavioral semantics, this system avoids misjudgments caused by dense crowds, achieving a false judgment rate of less than 10% in actual testing, a reduction of over 60% compared to traditional methods. Employing an "edge-triggered + video slicing" mechanism, only key segments are uploaded, reducing video transmission volume by 80% and cloud analysis latency from minutes to seconds. It outputs structured event chains, supporting complex behavioral reasoning and rule matching, achieving a leap from "seeing" to "understanding." Over 90% of invalid videos are processed at the edge, with only key segments uploaded to the cloud, saving bandwidth and server resources and supporting large-scale deployment. The system architecture supports multi-camera and multi-scene access and can be adapted to other monitoring scenarios such as bus stops and school bus stops by changing prompts. It automatically generates evidence chains, supporting traceability and review, improving law enforcement credibility and efficiency. Non-critical videos are not uploaded to the cloud, reducing the risk of privacy leaks.

[0042] Edge computing devices use the YOLOv8 object detection model to detect vehicles in the video stream and determine whether they are ride-hailing vehicles based on the vehicle's color, roof light markings, and vehicle type characteristics. The visual big model is the Qwen2.5-VL multimodal big model, and the alarm information of the GPU-accelerated computing nodes deployed on the central server includes evidence video links. The alarm information includes alarm ID, time of violation, geographical location, license plate number of the vehicle involved, violation type, and storage link of the original video slice.

[0043] This disclosure, in its second aspect, proposes a vehicle violation monitoring device based on edge and cloud collaboration. The device comprises: a data acquisition unit for real-time acquisition of video streams from a designated area via a camera; a video slice generation unit for receiving the video streams in real-time via an edge computing device, performing target detection on vehicles within the streams, and determining whether a preset event is triggered based on the vehicle's status. If a preset event is triggered, the corresponding video slice is captured and uploaded to a central server; and an alarm generation unit for the central server to perform semantic analysis on the video slices, generate a structured sequence of vehicle behavior events, match the sequence of vehicle behavior events with preset violation rules, and if a match is successful, determine that a violation has occurred and generate an alarm message.

[0044] In a third aspect, this disclosure provides a computer-readable medium storing a computer program, which is loaded and executed by a processing module to implement the steps of the acquisition method. Those skilled in the art will understand that all or part of the steps in the embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable medium, which may include various media capable of storing program code, such as flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0045] Within the scope of knowledge and ability of those skilled in the art, the various embodiments or technical features mentioned herein can be combined with each other as other optional embodiments without conflict. These finite number of optional embodiments, which are not listed one by one and are formed by combining a finite number of technical features, still fall within the scope of the technology disclosed herein and are also derived by those skilled in the art from the accompanying drawings and the foregoing.

[0046] In addition, the descriptions of most embodiments are based on different focuses. For further understanding of the parts not described in detail, reasonable inference can be made by referring to the relevant content of the prior art, other relevant descriptions in this document, or the inventive intent.

[0047] To reiterate, the embodiments listed above are typical and preferred embodiments of this disclosure, and are only used to describe and explain the technical solutions of this disclosure in detail to facilitate the reader's understanding. They are not intended to limit the scope or application of the protection claimed in this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure to obtain technical solutions should be covered within the scope of protection claimed in this disclosure.

Claims

1. A method for monitoring vehicle violations based on edge and cloud collaboration, characterized in that, include: Real-time video stream capture of a designated area using a camera; The edge computing device receives video streams in real time, performs target detection on vehicles in the streams, and determines whether a preset event is triggered based on the vehicle's status. If the preset event is triggered, the corresponding video slice is captured and uploaded to the central server. The central server performs semantic analysis on the video slices to generate a structured sequence of vehicle behavior events. It then matches the sequence of vehicle behavior events with preset violation rules. If a match is found, a violation is determined and an alarm message is generated.

2. The method according to claim 1, characterized in that, Preset events include: detecting that a vehicle door is open, or detecting that a vehicle has been continuously parked in a designated area for a period of time exceeding a preset time threshold.

3. The method according to claim 2, characterized in that, When the preset event is the opening of the car door, the video slice is extracted from the continuous video of the preset time periods before and after the opening of the car door. When the preset event is continuous parking timeout, the video slice is captured from the moment the vehicle is first determined to be parked to the current trigger moment, which is a continuous video clip.

4. The method according to claim 1, characterized in that, The vehicle behavior event sequence is structured data arranged in chronological order. Each event record includes the behavior type, timestamp, vehicle location coordinates, and license plate recognition result.

5. The method according to claim 1, characterized in that, Pre-defined violation rules include at least one of the following: passengers getting on or off the vehicle in an undesignated drop-off area; The vehicle's parking time exceeds a preset threshold and no passengers have boarded or alighted; The doors will be closed once no passengers approach or board / alight after the doors are opened. The number of passengers getting on and off the bus during a single stop exceeds the preset limit.

6. The method according to claim 1, characterized in that, When performing semantic analysis on video slices, the central server calls the visual big model and inputs system prompts into the visual big model before calling it. The system prompts are used to constrain the visual big model to output a structured format that conforms to the vehicle behavior event sequence.

7. The method according to claim 6, characterized in that, The system prompt includes output format constraints, requiring the visual model to return results in JSON format, with fields including: timestamp, behavior category, bounding box coordinates, and license plate number.

8. The method according to claim 1, characterized in that, Edge computing devices use the YOLOv8 object detection model to detect vehicles in the video stream and determine whether they are ride-hailing vehicles based on the vehicle's color, roof light markings, and vehicle type characteristics. The visual large model is a Qwen2.5-VL multimodal large model, and the alarm information of the GPU-accelerated computing node deployed on the central server includes evidence video links; the alarm information includes alarm ID, time of violation, geographical location, license plate number of the vehicle involved, type of violation and storage link of the original video slice.

9. A vehicle violation monitoring device based on edge and cloud collaboration, characterized in that, include: The acquisition device is used to acquire video streams of a designated area in real time via a camera; The video slicing and capturing device is used to receive video streams in real time from edge computing devices, perform target detection on vehicles in the streams, and determine whether a preset event is triggered based on the status of the vehicles. If the preset event is triggered, the corresponding video slice is captured and uploaded to the central server. The alarm information generation device is used by the central server to perform semantic analysis on video slices, generate structured vehicle behavior event sequences, match the vehicle behavior event sequences with preset violation rules, and if the match is successful, determine that there is a violation and generate alarm information.

10. A computer-readable medium, characterized in that: The computer-readable medium stores a computer program, which is loaded and executed by a processing module to implement the steps of the method of any one of claims 1 to 8.

Images