An unmanned aerial vehicle intelligent agent operation method and system for traffic accident detection and evidence collection

Abstract

The application discloses a kind of unmanned plane intelligent body operation method and system of traffic accident detection and evidence, it is related to intelligent transportation and unmanned plane technical field.The method is based on the closed loop architecture of " perception-decision-control", including: S1 perception stage, unmanned plane intelligent body is based on multi-machine coordination strategy response accident task, utilizes the safe flight corridor generated by occupancy grid map constructed in real time and approaches site;S2 decision stage, utilize on-board edge computing to identify accident type, and accordingly retrieve corresponding adaptive three-dimensional fly-around trajectory (such as U-shaped trajectory of rear-end collision, spiral trajectory of side collision, etc.) from strategy library;S3 control stage, along with the trajectory generated, execute high-precision evidence, and based on multidimensional features, real-time determine accident grade, if meet slight accident and vehicle movable condition, automatically trigger on-board sound wave device to shout and guide.The application realizes the unmanned and refinement of accident evidence by differentiating trajectory planning and immediate site disposal for different accident forms, effectively improves traffic recovery efficiency.

Description

Technical Field

[0001] This invention relates to the fields of unmanned aerial vehicle (UAV) applications and intelligent transportation technology. Specifically, it relates to an autonomous operation method and system for traffic accident detection and evidence collection by an unmanned aerial vehicle (UAV) intelligent agent based on a closed-loop architecture of "perception-decision-control," capable of adaptive path planning, autonomous evidence collection, and on-site emergency response. Background Technology

[0002] With the increasing severity of urban traffic congestion, the rapid detection and handling of traffic accidents has become a major challenge for traffic management. Traditional accident handling methods rely on traffic police arriving at the scene, which is time-consuming and prone to causing secondary congestion.

[0003] In recent years, unmanned aerial vehicles (UAVs) have been introduced into the transportation sector due to their high mobility and wide field of view. However, existing applications of police drones mainly suffer from the following limitations:

[0004] Low level of automation: Most applications still rely on remote control by "pilots", making it difficult to achieve truly unmanned operation.

[0005] Rigid evidence collection methods: Existing automatic cruise control mostly uses preset fixed waypoints (GPS Waypoints), which cannot adjust the shooting angle according to the specific form of the accident (such as rear-end collision, rollover), resulting in the loss of key evidence (such as brake marks, contact points).

[0006] Lack of closed-loop management: Drones are usually only used as "flying cameras" to transmit video, and cannot intervene in minor accidents on-site in a timely manner (such as by making announcements to guide traffic), causing "minor accidents" to turn into "major traffic jams".

[0007] Therefore, there is an urgent need for a drone intelligent agent approach with environmental perception, autonomous decision-making, and on-site execution capabilities to solve the above problems. Summary of the Invention

[0008] The purpose of this invention is to provide an autonomous operation method and system for unmanned aerial vehicle (UAV) intelligent agents in traffic accident detection and evidence collection. Through a closed-loop intelligent agent architecture of "perception-decision-control", it can achieve adaptive evidence collection and autonomous traffic management for different accident types.

[0009] To achieve the above objectives, the present invention provides the following technical solution: An autonomous operation method for unmanned aerial vehicle (UAV) intelligent agents in traffic accident detection and evidence collection, the method is based on a closed-loop intelligent agent architecture of "perception-decision-control", and includes the following steps.

[0010] S1. Perception Phase: Task Response and Safe Approach: The UAV agent receives a trigger command containing the accident coordinates. Based on the Euclidean distance between its current position and the accident coordinates and its remaining energy state, it is identified as the execution unit through a multi-aircraft collaborative task allocation strategy. During its flight to the accident coordinates, it uses airborne LiDAR and visual sensors to construct a locally occupied grid map. Through a real-time replanning algorithm, it generates a safe flight corridor to avoid dynamic obstacles, prioritizing the safety of the UAV itself and quickly reaching the accident site.

[0011] S2. Decision-making stage: Scene understanding and strategy generation. After arriving at the accident scene, the drone is controlled to perform a panoramic scan. The onboard edge computing module is used to run a target detection algorithm to identify the accident type and vehicle distribution topology at the accident scene. Based on the identified accident type, the agent autonomously retrieves the corresponding evidence-gathering flight strategy from the pre-set strategy library and generates an adaptive three-dimensional flight trajectory that includes specific spatial location, gimbal angle and shooting sequence. The adaptive three-dimensional flight trajectory presents a differentiated spatial form according to different accident types to maximize the coverage of key evidence.

[0012] S3. Control Phase: The drone is driven to fly along the adaptive 3D flight path for evidence collection and on-site handling, and the gimbal camera is controlled to perform fixed-point multi-exposure shooting at key nodes of the path to complete evidence collection. After evidence collection, the intelligent agent calculates the accident severity index and traffic congestion index in real time on the edge. It constructs a decision instruction containing the accident level and handling suggestions. In response to the decision instruction determining that the accident is minor and meets the safe movement conditions, it automatically generates and executes an audio-visual dispersal strategy containing traffic guidance voice. In response to the determination that the accident is serious, it automatically executes the on-site locking and alarm data uploading strategy.

[0013] Furthermore, in step S2, the generation of adaptive three-dimensional bypass trajectory based on accident type specifically includes: if a rear-end collision is identified, a "U-shaped side-rear bypass trajectory" is generated to cover the brake marks and contact surface of the rear vehicle; if a side collision at an intersection is identified, a "spiral ascending bypass trajectory" is generated to collect the relative positions of the two vehicles and the status of traffic lights; if a single vehicle rollover is identified, a "hemispherical coverage trajectory" is generated to perform gridded photography of the chassis and scattered objects.

[0014] Furthermore, the present invention also provides an unmanned aerial vehicle (UAV) intelligent agent system, including a perception and navigation module, an edge decision-making brain, and a collaborative control module, for executing the above-described methods.

[0015] The beneficial effects of the present invention are as follows:

[0016] Fully autonomous closed loop: No longer relying on ground control, it achieves unmanned operation of the entire process from receiving the alarm to handling the situation.

[0017] High-precision evidence collection: Specific geometric flight trajectories were designed for different accident scenarios, which significantly improved the completeness of evidence collection and the accuracy of liability determination.

[0018] Real-time traffic management: For the first time, a "self-announcement mechanism for minor accidents" based on visual analysis was introduced, which effectively shortened the time that accidents occupied the road. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Figure 1 This is a schematic diagram of the overall process of the method provided in the embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the principle of generating a safe flight corridor in phase S1 of this invention. Figure 3 This is a spatial geometry diagram of the adaptive three-dimensional flight trajectories (U-shaped, spiral, hemispherical) for different accident types in an embodiment of the present invention. Figure 4 This is a hardware architecture block diagram of the unmanned aerial vehicle (UAV) intelligent agent system in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0021] Example 1: The overall process is as follows Figure 1 As shown, the present invention provides an autonomous operation method for unmanned aerial vehicle (UAV) intelligent agents for traffic accident detection and evidence collection, which mainly includes three core stages.

[0022] Perception Phase (S1): When the traffic management center or roadside unit (RSU) detects an abnormal parking or collision signal, it generates a trigger command containing the GPS coordinates of the accident. The drone swarm receives the command. A multi-drone collaborative task allocation strategy (such as Contract Network Protocol (CNP) or a distributed auction algorithm) is employed here, comprehensively considering the current position (Euclidean distance), remaining battery power (SoC), and sensor health status of each drone. For example, the drone closest to the destination but with insufficient battery power to complete the return flight will abandon the mission, ultimately selecting the drone with the lowest overall cost as the execution unit. During the flight, the execution unit uses an onboard LiDAR and a binocular vision camera to perform SLAM mapping, constructing a locally occupied grid map with a resolution of 10cm-50cm. Based on this map, a gradient-based trajectory optimization algorithm (such as B-Spline Optimization) is used to generate a safe flight corridor that is not only collision-free but also meets the drone's dynamic constraints (maximum speed, maximum acceleration).

[0023] To ensure the smoothness and safety of the flight trajectory, this embodiment constructs the following safe flight corridor constraint model. The free space is divided into a set of convex polyhedra. C The position p(t) of the UAV at any time t must satisfy the linear inequality constraint:

[0024] Under this constraint, the Minimum Snap algorithm is used to optimize the path, with the objective function J as follows:

[0025] By solving the above optimization problem, the system generates a smooth flight trajectory that can both avoid dynamic obstacles and ensure stable camera shooting.

[0026] Decision-making phase (S2): The drone hovers at an altitude of approximately 30-50 meters above the accident site and performs a rapid panoramic scan. The onboard computing platform (such as the NVIDIA Jetson series) runs an improved YOLO or EfficientDet object detection algorithm to identify vehicles, lane lines, and debris. The core lies in adaptive strategy generation (such as... Figure 3 (As shown).

[0027] Scenario A: Rear-end collision. The model identifies two vehicles closely following each other. The agent calls the policy library to generate a "U-shaped rearward gliding trajectory." This trajectory is centered on the point of contact between the two vehicles, and within a 180-degree range behind them, it glides low from the left rear to the right rear, focusing on capturing the length of the brake marks on the following vehicle, the depth of the bumper damage, and the license plates of both vehicles.

[0028] The specific trajectory generation model is as follows: Let the coordinates of the accident center be... U-shaped trajectory Defined as:

[0029] In the formula, a and b are the lengths of the major and minor semi-axles of the U-shaped trajectory, respectively, which are determined by the size of the enclosure box of the accident vehicle; θ is the main direction angle of the road, θ(t) is a parameterized angle with a value range of [π, 2π]; and θ is the optimal overhead shooting height.

[0030] Scenario B: Side collision (intersection). The model identifies the two vehicles as being arranged in a T-shape or X-shape. The agent generates a "spiral upward orbital trajectory." The drone, with the accident center as its axis, gradually increases in radius and altitude (e.g., spiraling upward from 5 meters to 20 meters). The lower altitude is used to clearly see the impact point, while the upper altitude is used to capture the status of traffic lights, road stop lines, and the overall view of the intersection to determine if there was any red-light running.

[0031] Spiral Ascent Trajectory Equation Defined as:

[0032] In the formula, R is the helical radius, and w is the angular velocity. This refers to the vertical ascent speed. Simultaneously, the camera gimbal tilt angle... Dynamically adjusts with altitude: Scenario C: Rollover accident. The model identifies an abnormal aspect ratio of the vehicle or detects the chassis structure. The agent generates a "hemispherical coverage trajectory." Using the rolled-over vehicle as the center, high-density waypoint planning is performed on the hemispherical surface to ensure no blind spots in recording traces of chassis mechanical failures and the distribution of surrounding debris, ruling out vehicle-related malfunctions.

[0033] Control and Handling Phase (S3): The UAV executes the above trajectory and uses the gimbal's PID control algorithm to keep the camera constantly pointed at the region of interest (ROI). Autonomous Handling Logic for Minor Accidents: After evidence collection, the agent immediately analyzes the acquired images.

[0034] Step A: Detect the integrity of the vehicle outline (e.g., calculate the edge irregularity of the Bounding Box). If it is higher than the preset threshold, it means that the structure is fine.

[0035] Step B: Use a semantic segmentation algorithm to detect whether there are large areas of oil (black / colored liquid areas) on the road surface.

[0036] Step C: Detect if anyone has fallen to the ground (based on human key point detection).

[0037] To quantify the severity of an accident, the system introduces a Vehicle Damage Index (VDI) calculation model:

[0038] In the formula: The area of ​​the vehicle outline identified on-site; This represents the complete outline area of ​​the corresponding vehicle in the standard vehicle model library. The variable represents liquid leakage as a binary variable. W is the vehicle tilt angle. 1， W 2， W3 is the weighting coefficient.

[0039] Based on the above calculation results, the system executes the following judgment logic: if the conditions of "complete outline" + "no oil leakage" + "no casualties" are met, the agent determines it to be a minor accident (S_low). At this time, the agent does not need to wait for manual instructions and directly activates the onboard directional acoustic wave detector (LRAD) to synthesize the voice: "The current accident has been documented and is a minor accident. Please take photos immediately and drive into the emergency lane to avoid causing congestion."

[0040] Example 2: Self-Check of Evidence Quality To prevent blurry images from rendering evidence invalid, this invention also introduces a quality feedback mechanism. During the shooting process, a no-reference image quality evaluation algorithm (such as Brisque score) is used for real-time scoring. If a key frame (such as a close-up of a license plate) is found to be blurry or overexposed, the intelligent system automatically generates a reshoot waypoint within the current small area, adjusts the exposure parameters or hovers the image for secondary acquisition, ensuring "one-time attendance, effective evidence collection".

[0041] Example 3: Evidence Chain Security To ensure the legal validity of evidence collection, all data is processed by the onboard security chip (SE) before being written to the storage card. The system packages the image data, the flight log at the time, and the agent's decision logic log (why it was determined to be a minor accident), adds a timestamp and GPS watermark, and finally performs SHA-256 hash encryption and digital signature. This generates a tamper-proof, structured accident file, ensuring the authenticity of the data from the collection point to the court.

[0042] Example 4: System architecture as follows Figure 4 As shown, the system of the present invention includes...

[0043] Perception and Navigation Module: Integrates RTK-GPS, IMU, and LiDAR, and is responsible for obstacle avoidance and positioning in the S1 phase.

[0044] Edge Decision Brain: The core processing unit, which deploys deep learning models and is responsible for accident identification and trajectory simulation in the S2 phase.

[0045] Cooperative control module: connects the flight control interface (API) and payload (camera, loudspeaker), and is responsible for the precise execution of the S3 phase.

[0046] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for operating unmanned aerial vehicle (UAV) intelligent agents for traffic accident detection and evidence collection, characterized in that, The method is based on a closed-loop intelligent agent architecture of "perception-decision-control" and includes the following steps: S1. Perception Phase: Task Response and Safe Approach The drone agent receives a trigger command containing the accident coordinates. Based on the Euclidean distance between its current position and the accident coordinates and its remaining energy state, it is identified as the execution unit through a multi-drone collaborative task allocation strategy. During its flight to the accident coordinates, it uses airborne lidar and visual sensors to construct a local occupation grid map and generates a safe flight corridor to avoid dynamic obstacles through a real-time replanning algorithm, prioritizing the safety of the drone itself and quickly reaching the area above the accident site. S2, Decision-making stage: Scene understanding and strategy generation Upon arrival at the accident scene, the drone was controlled to perform a panoramic scan, and the onboard edge computing module was used to run a target detection algorithm to identify the accident type and vehicle distribution topology at the accident scene. Based on the identified accident type, the intelligent agent autonomously generates flight paths based on the learned spatial topology relationships, generating adaptive three-dimensional flight paths that include specific spatial locations, gimbal angles, and shooting sequences. The adaptive three-dimensional flight trajectory exhibits a differentiated spatial form depending on the type of accident, in order to maximize the coverage of key evidence; S3, Control Phase: Execution of Evidence Collection and On-site Handling The drone is driven to fly along the adaptive three-dimensional flight path, and the gimbal camera is controlled in conjunction with the gimbal camera to perform fixed-point multiple exposure shooting at key nodes of the path to complete the evidence collection. After the evidence collection is completed, the intelligent agent calculates the accident severity index and traffic congestion index in real time on the edge. If the calculation results indicate a minor accident and the vehicle is mobile, the intelligent agent directly activates the onboard directional sound wave device to synthesize and play specific traffic control voice commands to guide the vehicle away from the scene; if it is determined to be a serious accident, the scene is locked and the alarm data is uploaded.

2. The method according to claim 1, characterized in that, In step S2, the specific method for generating an adaptive three-dimensional flight trajectory based on the accident type includes: If the accident type is identified as a rear-end collision, the decision-making level generates a "U-shaped side-rear flight trajectory" and controls the drone to focus on covering the starting point of the brake marks of the rear vehicle, the collision contact surface, and the license plate areas of both vehicles. If the accident type is identified as a side collision at an intersection, the decision-making layer generates a "spiral ascending circular trajectory" and controls the drone to collect information on the relative positions of the two vehicles, the status of traffic lights at the intersection, and road markings from multiple angles from low to high altitudes. If the accident type is identified as a single vehicle rollover, the decision-making layer generates a "hemispherical coverage trajectory" and controls the drone to conduct high-density grid-based photography of the undercarriage of the rollover vehicle and the area where the debris is distributed.

3. The method according to claim 1, characterized in that, In step S3, the specific logic for the intelligent agent to determine a minor accident is as follows: the intelligent agent uses an airborne vision algorithm to detect the integrity of the vehicle's outline and segment the oil leakage area on the road; if the vehicle's outline integrity is higher than a preset threshold, there is no obvious oil leakage, and the accident area does not contain any features of people lying on the ground, then a "minor accident" label is generated, and a "shouting to drive away" control command is triggered. The content of the traffic guidance voice instructions is dynamically synthesized by the intelligent agent based on the congestion index of the current time period, including the standard prompt "The accident has been verified, please immediately enter the emergency lane".

4. The method according to claim 1, characterized in that, The method also includes a "self-inspection and feedback of evidence collection quality" step: During evidence collection and photography, the intelligent agent uses a no-reference image quality assessment algorithm to detect the sharpness and integrity of key elements of the acquired images in real time. If a blurry license plate or a key collision point is detected, the intelligent agent automatically generates a reshoot decision and performs secondary enhanced acquisition of the unqualified area by adjusting the hovering position or changing the gimbal tilt angle.

5. The method according to claim 1, characterized in that, In step S1, the safe flight corridor is generated using a gradient-based trajectory optimization algorithm (B-Spline Optimization), which incorporates obstacle distance as a penalty term into the cost function to ensure that the generated path maintains a dynamic safe distance from obstacles while satisfying dynamic constraints.

6. The method according to claim 1, characterized in that, In step S3, after the evidence collection is completed, the agent automatically performs the "evidence chain encapsulation" operation: The collected image data, flight trajectory logs, and intelligent agent decision logs (including accident type determination results and handling instruction records) are packaged; the data packets are digitally signed and spatiotemporally hashed using an onboard security chip to generate a tamper-proof structured accident file, which serves as valid evidence for subsequent legal liability determination.

7. A drone intelligent agent system for traffic accident detection and evidence collection, characterized in that, include: The perception and navigation module is used to perform step S1, which uses multi-sensor fusion to perceive environmental obstacles and plan a safe path; The edge decision-making brain, deployed on the drone's onboard computing platform, is used to execute step S2, identify accident types, and dynamically generate differentiated evidence-gathering flight strategies. The collaborative control module is used to execute step S3, driving the flight attitude, controlling the gimbal to capture images, and triggering the acoustic device to autonomously communicate based on the decision results.

Images