Unmanned aerial vehicle self-adaptive exploration method and system for road traffic accident scene

Abstract

The invention provides a road traffic accident site unmanned aerial vehicle self-adaptive exploration method, and the method comprises the steps: calculating the accident site diameter D and the accident center position according to the accident site image data collected by an unmanned aerial vehicle, and further calculating the shooting height H of an unmanned aerial vehicle exploration image, and does not need to set a reference object on the site ground. The survey image acquisition height is obtained through dynamic calculation according to the actual site condition; when a traffic accident scene live-action aerial view record graph is drawn, a laser range finder carried by the unmanned aerial vehicle is utilized to obtain the accurate height from the camera to the ground, and the actual size of any target in the aerial view record graph is directly calculated by combining the focal length of the camera and the size parameter of the sensor without depending on a ground scale or carrying out additional camera calibration. The whole process is low in calculation amount, high in calculation efficiency and suitable for various scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of road traffic accident handling, in particular to a road traffic accident scene unmanned aerial vehicle self-adaptive investigation method and system. BACKGROUND

[0002] Traditional road traffic accident scene investigation mainly relies on manual measurement using tools such as a tape measure and a laser range finder, and draws a scene map and records data by hand. This method has the problems of low efficiency, strong subjectivity, easy to cause secondary traffic congestion and safety risks, etc. In recent years, with the development of unmanned aerial vehicle technology, it has been gradually introduced into the field of accident scene investigation, but the existing applications are mostly limited to aerial photography and video recording functions, and the flight height, flight route planning and data collection are still manually operated by the pilot, and the intelligent degree is limited.

[0003] In the prior art, the patent document with publication number CN110081860A discloses a traffic accident scene mapping method, which obtains a real scene map by aerial photography of an unmanned aerial vehicle, and calculates the position of accident elements based on image composition operation and a proportionality coefficient, but this method relies heavily on ground markers as a reference, and the measurement accuracy is difficult to guarantee in the absence of clear reference objects or in the presence of perspective distortion. The patent document with publication number CN105314122A discloses an unmanned aerial vehicle for emergency command and road occupation evidence collection, which proposes a scheme for distance measurement using binocular vision, which can obtain depth information, but the system structure is complex, the cost is high, and the calculation amount is large, making it difficult to realize real-time processing on a mobile platform. The patent document with publication number CN106767706A discloses an aerial image acquisition method and system for unmanned aerial vehicle investigation of a traffic accident scene, which provides a detailed aerial photography scheme design, but there is still a lack of effective technical means for real-time and accurate measurement of scene elements, especially size measurement under the condition of dynamic flight of the unmanned aerial vehicle and change of the camera posture. It is particularly important to note that the existing measurement method based on the unmanned aerial vehicle will produce significant errors due to the perspective effect when taking pictures at an angle, and the measurement work after taking pictures usually requires placing a scale at the scene and implementing size conversion through a complex camera calibration process, which is cumbersome and prone to human error. If the accident scene cannot place a reference object or there is no obvious reference object due to weather or other complex reasons, the measurement accuracy of the unmanned aerial vehicle cannot be guaranteed. SUMMARY

[0004] In order to solve the problem that the existing technology needs to rely on ground reference for reference or proportion reference when investigating a traffic accident scene based on an unmanned aerial vehicle, thereby limiting the use scenario, the present application provides a road traffic accident scene unmanned aerial vehicle self-adaptive investigation method, which can accurately measure without a ground scale and automatically generate a standard scene map. The present application also discloses a road traffic accident scene unmanned aerial vehicle self-adaptive investigation system.

[0005] The technical solution of the present application is as follows: a road traffic accident scene unmanned aerial vehicle adaptive investigation method, characterized in that it comprises the following steps: S1: after receiving a traffic accident report, send an unmanned aerial vehicle to the accident scene; based on the data collected by the unmanned aerial vehicle, analyze and obtain the accident area ROI; The unmanned aerial vehicle is equipped with a visible light camera, a laser range finder and an inertial measurement unit IMU; The accident area ROI includes the accident involved target, the accident center position and the accident scene diameter D; The accident involved target includes vehicles, human bodies, road surface traces and road facilities; S2: calculate the initial collection height H of the unmanned aerial vehicle according to the accident scene diameter D, and determine the initial hovering point of the unmanned aerial vehicle in combination with the longitude and latitude of the accident center position; S3: the unmanned aerial vehicle flies to the adaptively determined hovering point, and automatically takes pictures of the accident scene according to the preset shooting rules; the accident scene pictures include overhead shots, overview shots and azimuth shots; The shooting angle of the overhead shot is 90° in the vertical direction; The azimuth shot is a azimuth photo reflecting the geographical position of the accident scene and the mutual relationship between the scene and the surrounding environment, and the azimuth shot is taken on the left and right sides of the overhead shot shooting position at a preset inclination angle; The overview shot collects an overview photo of the scene environment and road alignment; the overview shot is taken according to the road direction of the accident scene, and the number of overview shots corresponds to the number of road directions; S4: real scene record drawing; Based on the taken images, the vehicle recognition model and the road element recognition model trained are used to identify the vehicles and lane lines in the accident area, automatically extract the vehicle contours and directions, and superimpose standard graph elements to draw a traffic accident scene real scene overhead record drawing; When drawing the traffic accident scene real scene overhead record drawing, the laser range finder carried by the unmanned aerial vehicle is used to obtain the accurate height of the camera from the ground, and in combination with the camera focal length and sensor size parameters, the actual size of each accident involved target in the overhead photo is directly calculated; S5: data management and uploading: upload the collected images, videos, drawn record drawings and generated documents to the backend management platform.

[0006] Further characterized in that: In step S1, the analysis method of the accident area ROI is as follows: S11: construct a YOLOv5-OBB rotating target detection model with a direction angle; S12: constructing a target detection model based on a YOLOv5-OBB rotating target detection model; The input of the target detection model is a video stream collected by a UAV, and the output is position information of each accident-involved target included in an accident area; The position information includes the width w and height h of the OBB bounding box of the accident-involved target, the center point (x, y) of the bounding box, and the direction angle a of the bounding box; S13: constructing an accident area identification model based on a clustering algorithm; The input of the accident area identification model is the position information of all the accident-involved targets, and the output is the center position of the accident and the diameter D of the accident site; When the target detection model is trained, the loss function is L_total; L_total=L_yolov5+λ*L_angle; In the formula, L_yolov5 is the YOLOv5 loss function, L_angle is the regression loss function of the direction angle a, and λ is the balance hyperparameter; In step S2, the calculation method of the initial collection height H is H=1.5×[D / tan(FOV / 2)]×1.2; In the formula, FOV is the field of view of the UAV; D=n*L*ρ*γ; In the formula, D is the diameter of the accident site, n is the number of vehicles, L is the reference length of a single vehicle, ρ is the collision type coefficient, and γ is the superposition compensation coefficient; In step S3, the shooting process also includes the need to perform dynamic adjustment of the height of the UAV, and the specific method is: According to the preset area proportion of the accident area in the shooting picture, the height required to reach the target area proportion is calculated by using the proportional relationship between the area and the square of the height; In step S4, it specifically includes the following steps: S41: using the laser slant distance l from the lens center point to the image plane center recorded in real time when shooting the photo, and the real-time pitch angle φ of the camera, to calculate the vertical height H of the camera from the ground in real time 相机 The calculation formula is: H 相机 =l*cosφ; In the formula, H 相机 is the vertical height of the camera from the ground; l is the laser slant distance from the lens center point to the image plane center; and φ is the real-time pitch angle of the camera; S42: based on the vertical height H 相机, camera focal length f and image sensor pixel size s, a dynamic mapping model of image pixel distance and actual ground distance P between image pixels is established, and the mapping relationship is as follows: P=(p*s*H 相机 ) / (f*cscθ); In the formula, P is the actual ground distance between image pixels; p is the image pixel distance; s is the image sensor pixel size; f is the camera focal length; and θ is the angle of the image point deviating from the camera optical axis. S43: Based on the dynamic mapping model of image pixel distance p and actual ground distance P between image pixels, the actual size and position of the target object such as vehicle, scattered object, road parameter and trace in the accident scene are dynamically calculated. In step S43, the following steps are specifically included: S431: Convert the image coordinates to the camera coordinate system; S432: Calculate the position of the camera in the world coordinate system; S433: According to the position and posture of the camera and the direction of the OBB vertex in the camera coordinate system, the three-dimensional coordinates of the four vertices of the OBB boundary box corresponding to each accident involved target in the world coordinate system are calculated by the space forward intersection method; X=X s +(Z-Z s )*(a1*x+a2*y-a3*f) / (c1*x+c2*y-c3*f); Y=Y s +(Z-Z s )*(b1*x+b2*y-b3*f) / (c1*x+c2*y-c3*f); In the formula, (X, Y, Z) is the coordinate of the ground point in the world coordinate system; (X s , Y s , Z s ) is the coordinate of the camera center in the world coordinate system; (x, y) is the coordinate of the image point in the image plane coordinate system; f is the camera focal length; a1, a2, a3, b1, b2, b3, c1, c2, c3 are the direction cosine matrix elements composed of the camera posture angles (φ, ω, κ); S434: According to the three-dimensional coordinates of the vertices of the OBB boundary box of each accident involved target in the world coordinate system, the actual length and width of each accident involved target are calculated. The accident scene picture also includes a local light, which reflects the details of the accident involved target in the traffic accident scene.

[0007] A road traffic accident scene unmanned aerial vehicle adaptive investigation system, comprising: an unmanned aerial vehicle and a backend management platform, characterized in that; The unmanned aerial vehicle is provided with a visible light camera, a laser range finder and an inertial measurement unit (IMU). The unmanned aerial vehicle is provided with a visible light camera, a laser range finder and an inertial measurement unit (IMU). The scene investigation module receives the output of the visible light camera, the laser range finder and the inertial measurement unit (IMU), and realizes adaptive identification and adaptive flight path planning of the accident scene, and transmits the scene data to the backend management platform. The backend management platform realizes traffic accident scene real scene overhead recording drawing, data processing, model operation, data storage and case management.

[0008] The present application provides a road traffic accident scene unmanned aerial vehicle adaptive investigation method, which calculates the accident scene diameter D and the accident center position according to the accident scene picture data collected by the unmanned aerial vehicle, and then calculates the height H of the unmanned aerial vehicle investigation map shooting. Without setting a reference object on the ground, the investigation image acquisition height is dynamically calculated according to the actual situation on the scene. When drawing the traffic accident scene real scene overhead recording drawing, the laser range finder carried by the unmanned aerial vehicle is used to obtain the accurate height of the camera from the ground, and the actual size of any target in the overhead photo is directly calculated by combining the camera focal length and sensor size parameters, without relying on the ground scale or performing additional camera calibration. The whole process has low calculation amount and high calculation efficiency, and can be applied in various scenes. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 The flowchart for accident scene identification; Figure 2 The dynamic visual measurement principle diagram; Figure 3 The adaptive flight model diagram for road section; Figure 4 The adaptive flight model diagram for intersection. DETAILED DESCRIPTION

[0010] As Figure 1 shown, the present application includes a road traffic accident scene unmanned aerial vehicle adaptive investigation method, which comprises the following steps.

[0011] S1: After receiving the traffic accident report, the unmanned aerial vehicle is sent to the accident scene; based on the accident scene data collected by the unmanned aerial vehicle, the accident area ROI is analyzed and obtained.

[0012] The unmanned aerial vehicle is provided with a visible light camera, a laser range finder and an inertial measurement unit (IMU); the inertial measurement unit (IMU) is integrated with an accelerometer, a gyroscope and a magnetometer to measure the motion state of an object in a three-dimensional space.

[0013] The accident area ROI includes an accident-involved target, an accident center position and an accident site diameter D. The accident-involved target includes a vehicle, a human body, a road trace and a road facility.

[0014] In view of the problem that the direction of the photo taken by the unmanned aerial vehicle is uncertain, the application is improved based on a YOLOv5-OBB rotating target detection model, and a YOLOv5-OBB rotating target detection model with a direction angle is proposed; the accident site vehicle and the direction perception are detected based on the improved YOLOv5-OBB rotating target detection model; the improved YOLOv5-OBB rotating target detection model is used to process real-time images; the model additionally predicts the direction angle a of the boundary box center point (x, y) and the width and height (w, h) while predicting the boundary box center point (x, y), the width and height (w, h); and the rotating boundary box (OBB) information of the vehicle is output. The loss function increases the regression loss L_angle of the direction angle a on the basis of the original YOLOv5.

[0015] In step S1, the analysis method of the accident area ROI is as follows.

[0016] S11: constructing a target detection model based on the YOLOv5-OBB rotating target detection model with a direction angle.

[0017] S12: the input of the target detection model is the video stream collected by the unmanned aerial vehicle, and the output is the position information of each accident-involved target included in the accident area; The position information includes the width w and the height h of the OBB boundary box of the accident-involved target, the center point (x, y) of the boundary box and the direction angle a of the boundary box.

[0018] When the target detection model is trained, the loss function is L_total. L_total=L_yolov5+λ*L_angle; In the formula, L_yolov5 is the original loss function of YOLOv5, L_angle is the regression loss function of the direction angle a, and λ is a balance hyperparameter.

[0019] S13: Constructing an accident area recognition model based on a clustering algorithm; the clustering algorithm can be based on a clustering algorithm in the prior art, and in this embodiment, the accident area recognition model is constructed based on DBSCAN. The potential accident area (ROI) is generated by clustering according to the spatial distance of the targets. The center position and the approximate range (diameter D) of the accident are comprehensively judged by analyzing the clustering degree and distribution of the targets.

[0020] The input of the accident area recognition model is the position information of all the accident-involved targets, and the output is the center position of the accident and the diameter D of the accident site.

[0021] S2: Calculating the initial collection height H of the unmanned aerial vehicle according to the diameter of the accident site, and determining the initial hovering point of the unmanned aerial vehicle in combination with the longitude and latitude of the center position of the accident.

[0022] The calculation method of the initial collection height H is: H = 1.5 × [D / tan(FOV / 2)] × 1.2; In the formula, FOV is the field of view angle of the unmanned aerial vehicle; D = n * L * ρ * γ; In the formula, D is the diameter of the accident site, n is the number of vehicles, L is the reference length of a single vehicle; ρ is the collision type coefficient; and γ is the superposition compensation coefficient.

[0023] The reference length L of a single vehicle is valued according to different values preset in the system according to the identified vehicle type.

[0024] The collision type coefficient ρ and the superposition compensation coefficient γ are empirical coefficients established based on statistical analysis of historical accident data and in close combination with the classification of accident forms in relevant industry standards. In this embodiment, the collision type coefficient ρ takes the following values: When the collision type is a single vehicle accident or a double vehicle head-on collision, ρ = 1; When the collision type is a double vehicle rear-end collision, ρ = 1.1; When the collision type is a double vehicle side collision, ρ = 1.8; When the collision type is a single vehicle rollover, ρ = 2.5; When the collision type is a multi-vehicle chain collision or a rollover of two or more vehicles, ρ = 3.0.

[0025] The collision type coefficient γ takes the following values: For an accident involving two or more vehicles, γ = 0.8 as long as there is a superposition phenomenon, otherwise, when there is no superposition between all the vehicles, γ = 1.0.

[0026] To ensure that the collected images can be used for subsequent investigation, the system will be pre-specified accident area and the picture collected by the camera ratio is not less than 2 / 3. In the flight process, the system continuously identifies the accident, if the current picture is found in the accident area area ratio is less than 2 / 3, according to the inverse square relationship between area and height, automatically adjust the flight height to the best observation position. Flight model diagram as shown in Figure 2

[0027] S3: the unmanned aerial vehicle flies to the adaptively determined hovering point, and automatically photographs the accident scene pictures according to the preset photographing rules after reaching the hovering point; the accident scene pictures include: bird's-eye view, overview, azimuth view and local view.

[0028] The photographing angle of the bird's-eye view is 90° in the vertical direction.

[0029] The azimuth view reflects the geographical position of the accident scene and the mutual relationship between the scene and the surrounding environment. The azimuth view is photographed on the left and right sides of the bird's-eye view photographing position according to the preset inclination angle. In the embodiment, the included angle between the photographing angle of the azimuth view and the horizontal plane is set to 60°.

[0030] The overview view collects the overview view of the scene environment and the road alignment; the overview view is photographed according to the road direction of the accident scene, and the number of the overview view corresponds to the number of the road direction.

[0031] The local view reflects the details of the target involved in the traffic accident scene.

[0032] When photographing the bird's-eye view, if the accident range calculated by the system exceeds the coverage capability of a single photo, a two-dimensional orthographic flight path is automatically planned, multiple photos are collected and automatically spliced into a two-dimensional orthographic map.

[0033] In step S3, the photographing process further includes the need to perform dynamic adjustment of the height of the unmanned aerial vehicle, and the specific method is: According to the preset area ratio of the accident area in the photographing picture, the height required to reach the target area ratio is calculated by using the proportional relationship between the area and the square of the height.

[0034] S4: real scene record drawing; Based on the photographed images, the vehicle detection, vehicle head detection and vehicle segmentation model trained are used to identify the vehicles on the scene, the road element recognition model trained is used to identify the lane lines included in the accident area, the vehicle contour and direction are automatically extracted, and the standard graph elements are superimposed to draw the traffic accident scene real scene bird's-eye view record drawing; ​When drawing the real scene overhead view record of traffic accident scene, the precise height of camera from the ground is obtained by using the laser range finder carried by the unmanned aerial vehicle, and the actual size of any target in the overhead photo is directly calculated by combining the camera focal length and sensor size parameters, without relying on the ground scale or performing additional camera calibration.

[0035] In step S4, the following steps are specifically included. S41: Real-time calculate the vertical height H of the camera from the ground by using the laser slant distance l of the lens center point to the image plane center recorded in real time when shooting the photo, and the real-time pitch angle φ of the camera 相机 , and the calculation formula is: H 相机 =l*cosφ; In the formula, H 相机 is the vertical height of the camera from the ground; l is the laser slant distance of the lens center point to the image plane center; and φ is the real-time pitch angle of the camera. S42: Based on the vertical height H of the camera from the ground 相机 , the camera focal length f and the image sensor pixel size s, a dynamic mapping model of the image pixel distance and the actual ground distance P between image pixels is established, and the mapping relationship is as follows: P=(p*s*H 相机 ) / (f*cscθ); In the formula, P is the actual ground distance between image pixels; p is the image pixel distance; s is the image sensor pixel size; f is the camera focal length; and θ is the angle of the image point deviating from the camera optical axis.

[0036] The pixel is the smallest light-sensing unit of the image sensor, and the pixel size s is the physical size of a single light-sensing element, which is usually marked in microns (μm), and the specific value is taken according to the technical index of each camera.

[0037] S43: Based on the dynamic mapping model of the image pixel distance p and the actual ground distance P between image pixels, the actual size and position of the target objects such as vehicles, scattered objects, road parameters and traces in the accident scene are dynamically solved.

[0038] In step S43, the actual size and position of the target objects such as vehicles, scattered objects, road parameters and traces in the accident scene are dynamically solved, and for the perspective distortion problem caused by the non-vertical shooting of the unmanned aerial vehicle, the dynamic mapping model obtained in step S42 and the oriented bounding box (OBB) information are combined, and the actual length, width and position coordinates of the vehicle in the real world coordinate system are solved in real time by using the perspective projection correction algorithm based on the collinear equation. Specifically, step S43 includes the following steps.

[0039] S431: For the four vertices of the OBB box, the image coordinates are converted to the camera coordinate system by using the collinear condition equation; S432: Combine the slant range l obtained by laser ranging and the camera pose angle, and calculate the position of the camera in the world coordinate system by the space resection principle; S433: According to the camera position, pose, and the direction of the OBB vertex in the camera coordinate system, solve the three-dimensional coordinates of the four vertices of the OBB bounding box corresponding to each target involved in the accident in the world coordinate system by the space forward intersection method; X=X s +(Z-Z s )*(a1*x+a2*y-a3*f) / (c1*x+c2*y-c3*f); Y=Y s +(Z-Z s )*(b1*x+b2*y-b3*f) / (c1*x+c2*y-c3*f); In the formula, (X, Y, Z) is the coordinates of the ground point in the world coordinate system; (X s , Y s , Z s ) is the coordinates of the camera center in the world coordinate system; (x, y) is the coordinates of the image point in the image plane coordinate system; f is the focal length of the camera; a1, a2, a3, b1, b2, b3, c1, c2, c3 are elements of a 3x3 direction cosine matrix R composed of camera pose angles (φ, ω, κ).

[0040] The direction cosine matrix R composed of the camera pose angles (φ, ω, κ) is as follows: ; Wherein, the relationship between the elements of the matrix R and the camera pose angles (φ, ω, κ) is as follows: a1=cos(φ)*cos(κ); a2=cos(φ)*sin(κ); a3=-sin(φ); b1=sin(ω)*sin(φ)*cos(κ)-cos(ω)*sin(κ); b2=sin(ω)*sin(φ)*sin(κ)+cos(ω)*cos(κ); b3=sin(ω)*cos(φ); c1=cos(ω)*sin(φ)*cos(κ)+sin(ω)*sin(κ); c2 = cos(ω) * sin(φ) * sin(k) - sin(ω) * cos(k); c3 = cos(ω) * cos(φ).

[0041] S434: According to the three-dimensional coordinates of the vertices of the OBB bounding box of each accident-involved target in the world coordinate system, the actual length and width of each accident-involved target are calculated.

[0042] The detailed calculation process involved in step S43 is realized based on the prior art.

[0043] S5: Data management and uploading: upload the collected images, videos, drawn records and generated documents to the backend management platform.

[0044] The method realizes the full-process automation from on-site identification to intelligent shooting through the closed loop of "perception-decision-execution".

[0045] In order to realize the above-mentioned one kind road traffic accident scene unmanned aerial vehicle self-adapting investigation method, the present application also constructs the investigation system, it includes: unmanned aerial vehicle and backend management platform;Unmanned aerial vehicle is carried visible light camera, laser range finder and inertial measurement unit IMU.

[0046] The on-site investigation module is built-in on the unmanned aerial vehicle, and the on-site investigation module is in communication connection with the backend management platform, realizing data transmission.

[0047] The on-site investigation module receives the output of the visible light camera, the laser range finder and the inertial measurement unit IMU, realizes the adaptive identification and adaptive route planning of the accident scene, and simultaneously uploads all the collected photos, videos, drawn scene maps and generated investigation records, inquiry records and other documents according to the standard template to the backend management platform through the 4G / 5G network, realizes the full-process information management of data.

[0048] The backend management platform is set in the backend management platform, realizing the traffic accident scene real scene panoramic record drawing, data processing, model operation, data storage and case management.The backend management platform also includes a vehicle model library, a size calculation model and a road marking model, which are used to assist the real scene record drawing in step S4.

[0049] To further illustrate the technical means and effects taken by the present application to achieve the predetermined inventive purpose, the following describes the specific implementation, structure, features and effects of a road traffic accident scene unmanned aerial vehicle adaptive investigation method and system according to the present application, based on a 3-car rear-end accident scene, and details the unmanned aerial vehicle adaptive investigation steps. Obviously, the described embodiments are part of the embodiments of the present application, not all. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0050] In this embodiment, a complete road traffic accident scene unmanned aerial vehicle adaptive investigation system is provided, Hardware composition: the front end includes a multi-rotor unmanned aerial vehicle, a visible light camera, a laser range finder, an IMU, a control handle, a scene investigation module, a portable printer and a storage box are mounted on the unmanned aerial vehicle, and a unmanned aerial vehicle fast investigation system is arranged in the back-end management platform.

[0051] The work flow is as follows.

[0052] 1. After the investigator arrives at the scene, the scene investigation module on the control handle is opened, and the basic information of the case is entered.

[0053] 2. Select the "automatic flight" function. The unmanned aerial vehicle takes off, and the system starts the accident scene identification model (corresponding to Figure 1 ) through the unmanned aerial vehicle video stream picture. The model identifies that there are 3 cars gathered at the scene, and there is no obvious overlap, which is preliminarily judged as a rear-end accident. In this embodiment, n=3, L=4.5m, p=1.1, g=1.0, and FOV=84° According to the formula D=n*L*P*G, it is calculated that D≈14.85m.

[0054] According to the unmanned aerial vehicle FOV, the initial collection height H is calculated by the formula H=1.5×[D / tan(FOV / 2)]×1.2, and the initial collection height H≈55 meters is obtained. The system plans the route and controls the unmanned aerial vehicle to automatically fly to the accident center point calculated.

[0055] 3. During the flight of the unmanned aerial vehicle, the real-time picture is returned. The system continuously identifies that the accident area area ratio at the current height is less than 2 / 3, and then automatically adjusts the height to 45 meters to obtain a better panoramic view angle.

[0056] 4. After the unmanned aerial vehicle hovers, the gimbal automatically adjusts to 90° vertically downward. Since the estimated height does not exceed 70 meters, the system automatically adopts the "single shot mode" to shoot a high-resolution "bird's eye view".

[0057] 5. After taking the "bird's eye view", the system automatically performs the "orientation photo" task, controls the UAV to fly to the side of the accident point 30 meters away, the gimbal angle is 60°, and the orientation photo that can reflect the geographical position of the accident scene and the mutual relationship between the scene and the surrounding environment is taken. If the scene is located on the road section, it is taken along the road direction; if the scene is located at the intersection, it is taken along the road direction of all roads intersecting the intersection.

[0058] 6. After taking the "orientation photo", the system automatically performs the "overview photo" task, controls the UAV to fly to the side of the accident point 50 meters away, the gimbal angle is 45°, and the overview photo that can reflect the scene environment and road alignment is taken. If the accident is located on the road section, it is taken along the road direction or the vehicle driving track; if the scene is located at the intersection, it should be taken from three or more different directions along the road direction.

[0059] 7. After taking the photo, the data is returned to the back-end management platform through the control handle. The back-end management platform starts the AI drawing algorithm, based on the accurate height obtained by the laser range finder and the camera parameters, through the dynamic mapping model and the perspective projection correction algorithm, the accurate size of the target object is automatically calculated. The standard target object vector graph element is superimposed on the photo according to the recognized position and direction, and the traffic accident scene real scene bird's eye view record graph with size annotation is automatically generated.

[0060] 8. After the back-end management platform completes the drawing, the traffic accident scene real scene bird's eye view record graph is pushed to the front-end control handle on-site investigation module, and the on-site investigation personnel can preview and fine-tune the traffic accident scene real scene bird's eye view record graph on the on-site investigation module, and then print it on the spot through the connected portable printer. At the same time, all data (photos, videos, scene graphs) are automatically uploaded to the back-end platform through 4G / 5G, and a standardized investigation record is generated.

[0061] The application technical scheme is characterized in that: the visible light camera and the laser range finder carried by the unmanned aerial vehicle collect field data; the rotating target detection model and the clustering analysis technology are used to automatically identify the accident area, the vehicle and the personnel, and determine the accident center point and the range; according to the identification result, the preset algorithm is combined to adaptively calculate the initial collection height H and the flight route, and the position, the height and the attitude of the unmanned aerial vehicle are dynamically adjusted to ensure that the accident scene is fully covered; the actual size of the target in the overhead photograph is directly calculated by using the laser ranging data and the camera parameters without the ground scale or the complex calibration; the overhead photograph, the overview photograph, the direction photograph and the local photograph are automatically taken; the field elements such as the vehicle and the lane line are identified based on the deep learning model, and the traffic accident scene real scene overhead record drawing is automatically drawn. The application realizes the automation, the intelligentization and the precision of the traffic accident scene investigation, effectively overcomes the perspective distortion problem of the traditional method under the dynamic flight condition, and realizes the high-precision real-time measurement without the ground scale. Through the technical scheme of the application, the traffic accident scene investigation is realized to be fast, accurate, automatic and information-based, the burden of the scene investigation personnel is greatly reduced, and the accident handling quality is improved.

Claims

1. An adaptive unmanned aerial vehicle (UAV) reconnaissance method for road traffic accident scenes, characterized in that, It includes the following steps: S1: After receiving a traffic accident report, dispatch a drone to the area above the accident scene; based on the data collected by the drone at the accident scene, analyze and determine the ROI of the accident area; The drone is equipped with a visible light camera, a laser rangefinder, and an inertial measurement unit (IMU). The ROI of the accident area includes: the target involved in the accident, the location of the accident center, and the diameter D of the accident site; The targets involved in the accident included: vehicles, human bodies, road marks, and road facilities; S2: Calculate the initial data collection height H of the UAV based on the diameter D of the accident site, and determine the initial hovering point of the UAV by combining the latitude and longitude of the accident center location; S3: The drone flies to the adaptively determined hovering point and automatically takes pictures of the accident scene according to the preset shooting rules; the accident scene pictures include: overhead view, overview view and azimuth view; The overhead photo was taken from a 90° vertical angle. The azimuth photo is a azimuth photo that reflects the geographical location of the accident site and the relationship between the site and the surrounding environment. The azimuth photo is taken from the left and right sides of the position where the overhead photo was taken, at a preset tilt angle. The overview photos are taken as overview photos of the scene environment and road alignment; the overview photos are taken according to the road direction of the accident scene, and the number of overview photos corresponds to the number of road directions; S4: Real-scene recording drawing; Based on the captured images, the trained vehicle recognition model and road element recognition model are used to identify vehicles and lane lines in the accident area, automatically extract vehicle outlines and directions, and overlay standard primitives to draw a real-time aerial view of the traffic accident scene. When drawing an aerial view of a traffic accident scene, the precise height of the camera above the ground is obtained using a laser rangefinder mounted on a drone. Combined with the camera's focal length and sensor size parameters, the actual size of each target involved in the accident in the aerial photo is directly calculated. S5: Data Management and Upload: Upload the collected images, videos, drawn record diagrams, and generated documents to the backend management platform.

2. The adaptive UAV reconnaissance method for road traffic accident scenes according to claim 1, characterized in that: In step S1, the analysis method for the ROI of the accident area is as follows: S11: Construct a YOLOv5-OBB rotating target detection model with orientation angle; S12: Construct a target detection model based on the YOLOv5-OBB rotating target detection model with orientation angle; The target detection model takes a video stream captured by a drone as input and outputs the location information of each target involved in the accident within the accident area. The location information includes: the width w and height h of the OBB bounding box of the target involved in the accident, as well as the center point (x,y) of the bounding box and the orientation angle α of the bounding box; S13: Constructing an accident area identification model based on clustering algorithm; The input to the accident area identification model is the location information of all the targets involved in the accident, and the output is the location of the accident center and the diameter D of the accident site.

3. The adaptive UAV reconnaissance method for road traffic accident scenes according to claim 2, characterized in that: The loss function for training the target detection model is L_total; L_total=L_yolov5+λ*L_angle; In the formula, L_yolov5 is the YOLOv5 loss function, L_angle is the regression loss function for the orientation angle α, and λ is the balancing hyperparameter.

4. The adaptive UAV reconnaissance method for road traffic accident scenes according to claim 1, characterized in that: In step S2, the initial acquisition height H is calculated as follows: H = 1.5 × [D / tan(FOV / 2)] × 1.2; In the formula, FOV is the field of view of the UAV; D=n*L*ρ*γ; In the formula, D is the diameter of the accident scene, n is the number of vehicles, L is the reference length of a single vehicle; ρ is the collision type coefficient; and γ is the superposition compensation coefficient.

5. The adaptive UAV reconnaissance method for road traffic accident scenes according to claim 1, characterized in that: Step S3 also includes dynamically adjusting the drone's altitude during the shooting process. The specific method is as follows: Based on the area proportion of the accident zone in the preset shooting frame, the height required to achieve the target area proportion is calculated using the ratio of area to the square of height.

6. The adaptive UAV reconnaissance method for road traffic accident scenes according to claim 1, characterized in that: Step S4 specifically includes the following steps; S41: Using the laser slant distance l from the lens center point to the image plane center recorded in real time during photo taking, and the camera's real-time pitch angle φ, the vertical height H of the camera above the ground is calculated in real time. 相机 The calculation formula is: H 相机 =l*cosφ; In the formula, H 相机 φ is the vertical height of the camera above the ground; l is the laser slant distance from the center of the lens to the center of the image plane; φ is the real-time pitch angle of the camera. S42: Based on the vertical height H of the camera above the ground 相机 Given the camera focal length f and image sensor pixel size s, a dynamic mapping model is established between the distance between image pixels and the actual ground distance P between image pixels. The mapping relationship is as follows: P=(p*s*H 相机 ) / (f*cscθ); In the formula, P is the actual ground distance between image pixels; p is the image pixel distance; s is the image sensor pixel size; and f is the camera focal length. θ is the angle at which the image point deviates from the camera's optical axis; S43: Based on the dynamic mapping model of image pixel distance p and actual ground distance P between image pixels, dynamically calculate the actual size and position of target objects such as vehicles, debris, road parameters, and traces at the accident scene.

7. The adaptive UAV reconnaissance method for road traffic accident scenes according to claim 6, characterized in that: Step S43 specifically includes the following steps: S431: Transform the image coordinates to the camera coordinate system; S432: Calculate the position of the camera in the world coordinate system; S433: Based on the camera position, attitude, and the orientation of the OBB vertices in the camera coordinate system, the three-dimensional coordinates of the four vertices of the OBB bounding box of each target involved in the incident are calculated in the world coordinate system using the spatial forward intersection method. X=X s +(Z-Z s )*(a1*x+a2*y-a3*f) / (c1*x+c2*y-c3*f); Y=Y s +(Z-Z s )*(b1*x+b2*y-b3*f) / (c1*x+c2*y-c3*f); In the formula, (X,Y,Z) are the coordinates of the ground point in the world coordinate system; (X s Y s Z s () represents the coordinates of the camera center in the world coordinate system; (x, y) are the coordinates of the image point in the image plane coordinate system; f is the camera focal length; a1, a2, a3, b1,b2, b3, c1, c2, c3 are the elements of the direction cosine matrix composed of camera attitude angles (φ,ω,κ); S434: Calculate the actual length and width of each target involved in an accident based on the three-dimensional coordinates of the vertices of the OBB bounding box of each target involved in the accident in the world coordinate system.

8. The adaptive UAV reconnaissance method for road traffic accident scenes according to claim 1, characterized in that: The accident scene photos also include close-up shots, which show details of the targets involved in the traffic accident.

9. An adaptive unmanned aerial vehicle (UAV) investigation system for road traffic accident scenes, comprising: The drone and backend management platform are characterized by: The drone is equipped with a visible light camera, a laser rangefinder, and an inertial measurement unit (IMU). The drone has a built-in field survey module, which is communicatively connected to the backend management platform. The on-site reconnaissance module receives the outputs from the visible light camera, the laser rangefinder, and the inertial measurement unit (IMU), and enables adaptive identification and adaptive route planning of the accident site, while transmitting the on-site data to the backend management platform. The backend management platform enables the creation of real-time aerial view recordings of traffic accident scenes, data processing, model calculations, data storage, and case management.

Images