Vehicle queuing length dynamic estimation method and system based on unmanned aerial vehicle cruise

Abstract

The invention discloses a vehicle queuing length dynamic estimation method and system based on unmanned aerial vehicle cruise, belongs to the technical field of traffic state monitoring, and is used for solving the technical problems of how to realize accurate identification and tracking of queuing vehicles and panoramic scene construction and solving the problem of vehicle position deviation caused by time difference under dynamic traffic flow. And the accuracy and stability of vehicle queuing length estimation are improved. The method comprises the following steps: acquiring cruise video data of a congested road section through an inspection unmanned aerial vehicle; performing image processing on the cruise video data, and extracting vehicle characteristic parameters of the congested road section; constructing a lane-level queuing vehicle discrimination model based on two indexes of vehicle speed and linear vehicle distance; acquiring an ordered queuing vehicle set of the congested road section through the model; determining the initial queuing length of each lane according to the ordered queuing vehicle set; and dynamically correcting the initial queuing length based on a space-time compensation correction algorithm to obtain the real-time queuing length of each lane in the congested road section.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traffic state monitoring, and particularly relates to a vehicle queue length dynamic estimation method and system based on unmanned aerial vehicle cruising. BACKGROUND

[0002] In modern urban traffic management and intelligent transportation system construction, accurate monitoring of road vehicle driving state and real-time analysis of vehicle queue condition on long-distance road section are key links to improve road traffic efficiency and optimize traffic scheduling decisions. However, the current mainstream traffic monitoring means still has many technical limitations, which is difficult to meet the fine management needs in complex traffic scenarios.

[0003] Traditional road traffic monitoring equipment has inherent defects such as limited coverage and single observation angle due to fixed installation position, and it is difficult to capture the whole scene of vehicle driving state on long-distance road section or complex intersection. For vehicle queue analysis, fixed monitoring equipment can only obtain vehicle information of local road section and cannot provide complete queue scene data, resulting in a significant reduction in the analysis accuracy of key parameters such as long-distance queue length and vehicle spatial distribution. Although unmanned aerial vehicle cruising technology has gradually been applied to the field of traffic monitoring due to its flexible movement advantage, the coverage range of single frame unmanned aerial vehicle image is still limited by the shooting height and angle, and in the dynamic cruising process, the traditional vehicle recognition and tracking algorithm is prone to inaccurate detection and tracking interruption due to the interference of factors such as changes in light conditions, various vehicle types and complex motion states. At the same time, the fusion processing of multiple images lacks effective algorithm support, and it is difficult to synthesize high-quality wide-angle panoramic image, which still cannot break through the range limitation of single frame image, restricting the realization of panoramic traffic scene analysis.

[0004] In the field of vehicle queue length estimation, the accuracy and stability of existing technologies face severe challenges when dealing with dynamic traffic flow. The accurate estimation of vehicle queue length requires accurate acquisition of vehicle position information, but in the analysis process based on multi-frame image fusion, the actual position of the vehicle will change over time due to the difference in key frame sampling time. The traditional method often uses static position mapping method without considering the position offset in time dimension, which leads to deviation in vehicle position estimation. Especially in road intersections, highways and complex traffic condition road sections, vehicles are often in slow driving, frequent start-stop state, and speed and acceleration are always in dynamic change. The limitation of static mapping method is more prominent, which easily causes misjudgment of the position of the head vehicle of the queue, and makes the estimated queue length result deviate significantly from the actual traffic flow state, which is difficult to meet the demand of traffic management for accurate state data.

[0005] In summary, how to break through the coverage range and perspective limit of the existing monitoring means, realize accurate identification, tracking and panoramic scene construction of vehicles, and on this basis solve the vehicle position deviation problem caused by time difference under dynamic traffic flow, and improve the accuracy and stability of vehicle queue length estimation, become the technical problems to be solved in the current intelligent traffic monitoring field. SUMMARY

[0006] The embodiment of the application provides a vehicle queue length dynamic estimation method and system based on unmanned aerial vehicle cruising, which is used for solving the following technical problems: how to realize accurate identification, tracking and panoramic scene construction of vehicles in a queue, and on this basis, solve the vehicle position deviation problem caused by time difference under dynamic traffic flow, and improve the accuracy and stability of vehicle queue length estimation.

[0007] The embodiment of the application adopts the following technical scheme:

[0008] On the one hand, the embodiment of the application provides a vehicle queue length dynamic estimation method based on unmanned aerial vehicle cruising, which comprises the following steps:

[0009] Image processing is performed on the cruising video data, and vehicle feature parameters of the congested road section are extracted; wherein the vehicle feature parameters at least include any one or more of the following: head coordinates, tail coordinates, vehicle type, vehicle speed, acceleration, vehicle color and straight-line vehicle spacing between adjacent vehicles in the same lane of each vehicle target;

[0010] Based on the vehicle feature parameters, a lane-level queuing vehicle discrimination model based on vehicle speed and straight-line vehicle spacing is constructed; and through the lane-level queuing vehicle discrimination model, an ordered set of queuing vehicles of the congested road section is obtained;

[0011] According to the ordered set of queuing vehicles, the initial queue length of each lane in the congested road section is determined;

[0012] Based on the space-time compensation correction algorithm, the initial queue length is dynamically corrected to obtain the real-time queue length of each lane in the congested road section.

[0013] In a feasible implementation manner, cruising video data of a congested road section is collected through a patrol unmanned aerial vehicle, specifically comprising:

[0014] Real-time monitoring of road traffic running state information is performed based on an electronic map;

[0015] When it is monitored that congestion occurs in a certain road section and the congestion duration reaches a preset determination condition, an unmanned aerial vehicle cruising task is automatically triggered;

[0016] The head position of the congested road section is located, and after the inspection unmanned aerial vehicle reaches above the head position, a reverse cruise mode is started, the inspection unmanned aerial vehicle flies in the reverse direction of the traffic flow, cruise video data is collected, and a flight trajectory is calibrated in real time; wherein the flight parameter of the inspection unmanned aerial vehicle is set as a vertical ground low-angle view angle.

[0017] In an available implementation, the cruise video data is image-processed to extract vehicle feature parameters of the congested road section, specifically including:

[0018] In the cruise video data, continuous image frames are extracted, and a YOLOv8 vehicle detection model is used to identify vehicle targets in each image frame, so as to determine a plurality of vehicle targets;

[0019] A DeepSORT motion detection model is used to continuously track the vehicle targets, so as to obtain vehicle numbers and vehicle recognition frame coordinates of each vehicle target;

[0020] Based on the vehicle recognition frame coordinates, head coordinates, tail coordinates, vehicle types and straight-line inter-vehicle distances between adjacent vehicles in the same lane of each vehicle target are obtained;

[0021] According to the flight speed of the inspection unmanned aerial vehicle, the detected vehicle speed in the image is corrected to obtain actual vehicle speed and vehicle acceleration of the vehicle target;

[0022] RGB values of all pixels in the vehicle recognition frame are extracted, and an average value is taken as a vehicle color of the vehicle target.

[0023] In an available implementation, based on the vehicle recognition frame coordinates, head coordinates, tail coordinates, vehicle types and straight-line inter-vehicle distances between adjacent vehicles in the same lane of each vehicle target are obtained, specifically including:

[0024] The traffic flow direction of each lane in the congested road section is determined; wherein the traffic flow direction includes west-to-east, east-to-west, south-to-north and north-to-south;

[0025] According to the vehicle recognition frame coordinates of the vehicle template and the traffic flow direction, position coordinates of a head line center point are determined to obtain the head coordinates, and position coordinates of a tail line center point are determined to obtain the tail coordinates;

[0026] The length of the vehicle recognition frame is determined as a vehicle length, and the vehicle length is matched in a preset vehicle type determination table to obtain the vehicle type of each vehicle target;

[0027] The distance between the front vehicle tail coordinates and the rear vehicle head coordinates of the adjacent vehicles is calculated, and is converted by a conversion coefficient between a pixel distance and an actual distance to obtain the straight-line inter-vehicle distance between the adjacent vehicles in the same lane.

[0028] In a feasible implementation, based on the vehicle characteristic parameters, a lane-level queued vehicle discrimination model based on two indexes of vehicle speed and straight-line inter-vehicle distance is constructed, specifically including:

[0029] According to the vehicle speed and a first preset speed threshold, a queued vehicle preliminary discrimination model based on the vehicle speed index is constructed, which is used for preliminary discrimination of queued vehicles; wherein the first preset speed threshold is related to the lane free flow speed;

[0030] According to the preliminary discrimination of the queued vehicles, the tail coordinates of each queued vehicle are taken as the reference, and a set of upstream vehicles located in the same lane neighborhood range is searched by a neighborhood search algorithm; wherein when the head coordinates of an upstream vehicle are located in the range, the upstream vehicle is counted into the set of upstream vehicles;

[0031] According to the straight-line inter-vehicle distance between the queued vehicles and each upstream vehicle in the set of upstream vehicles and a preset inter-vehicle distance threshold, a queued vehicle re-discrimination model based on the straight-line inter-vehicle distance is constructed; wherein the preset inter-vehicle distance threshold is related to the vehicle speed and deceleration of the upstream vehicle;

[0032] The queued vehicle preliminary discrimination model and the queued vehicle re-discrimination model are constructed into the lane-level queued vehicle discrimination model.

[0033] In a feasible implementation, by the lane-level queued vehicle discrimination model, an ordered set of queued vehicles of the congestion road section is obtained, specifically including:

[0034] The vehicle speed of each vehicle target in the same frame image is input into the queued vehicle preliminary discrimination model to obtain a preliminary queued vehicle sequence in each lane;

[0035] According to the preliminary queued vehicle sequence, a corresponding set of upstream vehicles is obtained;

[0036] The straight-line inter-vehicle distance between each queued vehicle in the preliminary queued vehicle sequence and each corresponding upstream vehicle, and the vehicle speed and deceleration of the upstream vehicle are input into the queued vehicle re-discrimination model to obtain a complete queued vehicle sequence of each lane;

[0037] The complete queued vehicle sequences of all lanes are combined into a queued vehicle set;

[0038] All queued vehicle sets in each frame image in the cruise video data are integrated into the ordered set of queued vehicles of the congestion road section.

[0039] In an embodiment, the initial queue length of each lane in the congestion road section is determined according to the ordered set of queued vehicles, specifically including:

[0040] The key frames of the cruising video data are extracted based on the frame difference method, and the key frames are fused and spliced into a panoramic image of the congestion road section by using the scale invariant feature transform (SIFT) algorithm and the fade-in and fade-out weighted average fusion algorithm;

[0041] In the ordered set of queued vehicles, a vehicle sequence chain in each lane in the panoramic image of the congestion road section is obtained;

[0042] The initial queue length of each lane is determined according to the head coordinate of the head vehicle and the tail coordinate of the tail vehicle in the vehicle sequence chain.

[0043] In an embodiment, the initial queue length of each lane is determined according to the head coordinate of the head vehicle and the tail coordinate of the tail vehicle in the vehicle sequence chain, specifically including:

[0044] If the congestion road section is a straight road, a straight-line distance between the head coordinate of the head vehicle and the tail coordinate of the tail vehicle is directly calculated to determine the initial queue length;

[0045] If the congestion road section is a curved road, road curve fitting is performed according to the head coordinate of the front vehicle, the left tail of the front vehicle, and the left head of the rear vehicle in each adjacent two queued vehicles, and the positions of the three points are taken as curve fitting reference points;

[0046] The curve distance between each adjacent two queued vehicles is calculated according to the characteristic parameters of the fitted road curve;

[0047] The initial queue length of each lane in the congestion road section is obtained by accumulating the curve distance between each adjacent two queued vehicles and the length of each vehicle by using the accumulation method.

[0048] In an embodiment, the initial queue length is dynamically corrected based on a space-time compensation correction algorithm to obtain the real-time queue length of each lane in the congestion road section, specifically including:

[0049] The key frame sequence of the appearance of the head vehicle, the departure of the head vehicle, and the appearance of the tail vehicle is identified in the cruising video data, and the running parameters of the head vehicle and the tail vehicle are extracted in the key sequence frames; wherein the running parameters of the head vehicle at least include the appearance time of the head vehicle, the departure time of the head vehicle, the instantaneous speed of the head vehicle in the key frame image, and the acceleration; and the running parameters of the tail vehicle at least include the appearance time of the tail vehicle;

[0050] According to the instantaneous vehicle speed and acceleration of the head vehicle in the key frame image and the time difference between adjacent key frame sampling time, the estimated speed of the head vehicle is calculated;

[0051] According to the estimated speed and the second preset speed threshold, it is judged whether the head vehicle has left the queue vehicle queue;

[0052] Based on the moving distance of the head vehicle and the tail vehicle, a dynamic correction model of real-time queue length is constructed;

[0053] The head vehicle operating parameters and the tail vehicle operating parameters are input into the dynamic correction model to obtain the real-time queue length of each lane in the congestion road section.

[0054] In another aspect, the embodiments of the present application also provide a vehicle queue length dynamic estimation system based on unmanned aerial vehicle cruising, which comprises:

[0055] A parameter extraction module is configured to collect cruising video data of the congestion road section by the inspection unmanned aerial vehicle, and extract vehicle feature parameters of the congestion road section by image processing on the cruising video data, wherein the vehicle feature parameters at least include any one or more of the following: head coordinates, tail coordinates, vehicle type, vehicle speed, acceleration, vehicle color and straight-line inter-vehicle distance between adjacent vehicles in the same lane;

[0056] A queue vehicle identification module is configured to construct a lane-level queue vehicle discrimination model based on vehicle speed and straight-line inter-vehicle distance based on the vehicle feature parameters, and obtain an ordered set of queue vehicles of the congestion road section through the lane-level queue vehicle discrimination model;

[0057] A queue length estimation module is configured to determine the initial queue length of each lane in the congestion road section according to the ordered set of queue vehicles, and dynamically correct the initial queue length based on a space-time compensation correction algorithm to obtain the real-time queue length of each lane in the congestion road section.

[0058] Compared with the prior art, the vehicle queue length dynamic estimation method and system based on unmanned aerial vehicle cruising provided by the embodiments of the present application have the following beneficial effects:

[0059] (1) The present application adopts double indicators of vehicle speed and straight-line inter-vehicle distance to cyclically determine queue vehicles, proposes an automatic search algorithm for queue vehicles in the same lane by combining neighborhood search and angular deviation constraint, and can accurately obtain an ordered set of queue vehicles. At the same time, the three-point positioning method of adjacent two vehicles is used to fit the road alignment, the straight-line distance between vehicles is corrected, the error problem of traditional straight-line ranging under complex alignment road conditions is effectively solved, and the estimation accuracy of queue length under complex road conditions is improved.

[0060] (2) The application aims at the vehicle position deviation problem caused by different key frame sampling time in image fusion, and proposes a queuing length dynamic correction method based on space-time compensation, which effectively improves the adaptability of the method to the vehicle slow driving or frequent start-stop scene, significantly improves the real-time performance and reliability of the queuing length estimation in the congestion dynamic evolution process, and is especially suitable for scenes such as interlaced area of urban expressway and long-distance queuing on expressway, thereby providing accurate situation information for traffic control. BRIEF DESCRIPTION OF DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. In the drawings:

[0062] Figure 1 A flow chart of a vehicle queuing length dynamic estimation method based on unmanned aerial vehicle cruising is provided for the embodiments of the present application;

[0063] Figure 2 A schematic diagram of an unmanned aerial vehicle cruising scene is provided for the embodiments of the present application;

[0064] Figure 3 A road curve fitting schematic diagram is provided for the embodiments of the present application;

[0065] Figure 4 A structure schematic diagram of a vehicle queuing length dynamic estimation system based on unmanned aerial vehicle cruising is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0066] In order to make the person skilled in the art better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only some 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 should be within the scope of protection of the present application.

[0067] The embodiments of the present application provide a vehicle queuing length dynamic estimation method based on unmanned aerial vehicle cruising, as shown in Figure 1 The vehicle queuing length dynamic estimation method based on unmanned aerial vehicle cruising specifically includes steps S101-S105:

[0068] S101, collecting cruising video data of the congestion road section through the inspection unmanned aerial vehicle.

[0069] Specifically, the road traffic running state information is monitored in real time based on an electronic map. When congestion occurs on a road section and the congestion duration reaches a preset determination condition, an unmanned aerial vehicle (UAV) patrol task is triggered automatically.

[0070] Further, the head position of the congested road section is located, and after the patrol UAV reaches above the head position, a reverse patrol mode is started, the patrol UAV flies reversely to the traffic direction, collects patrol video data and calibrates the flight trajectory in real time, wherein the flight parameter of the patrol UAV is set as a vertical ground low-angle view.

[0071] As a feasible implementation manner, Figure 2 A UAV patrol scene schematic diagram provided by the embodiment of the application is shown in FIG. 1. Figure 2 As shown in FIG. 1, the road traffic running state information is monitored in real time based on an electronic map, when the congestion duration of a road section reaches a preset determination condition, a UAV patrol task is triggered automatically. The head position of the congested road section is located accurately, a reverse patrol mode of the UAV is started, the UAV flies reversely to the traffic direction, collects and calibrates the flight trajectory in real time. The flight parameter of the UAV is set as a vertical ground low-angle view (the angle between the lens optical axis and the ground is 90°), the flight height is kept at 80-120 meters from the ground, and the flight speed is controlled at 10-15 kilometers per hour on a normal road section.

[0072] S102, image processing is performed on the patrol video data to extract vehicle feature parameters of the congested road section.

[0073] The vehicle feature parameters at least include any one or more of the following: the head coordinates, the tail coordinates, the vehicle type, the vehicle speed, the acceleration, the vehicle color and the straight-line vehicle spacing between adjacent vehicles in the same lane of each vehicle target.

[0074] Specifically, continuous image frames are extracted from the patrol video data, a vehicle target recognition is performed on each image frame by a YOLOv8 vehicle detection model, and a plurality of vehicle targets are determined. A continuous tracking is performed on the vehicle targets by a DeepSORT motion detection model, and the vehicle number and the vehicle recognition frame coordinates of each vehicle target are obtained.

[0075] In one embodiment, the continuous tracking of the vehicle is realized by a motion detection model in the DeepSORT algorithm, and the vehicle number i (i∈Z + ) and the vehicle recognition frame coordinates (x c ,y c ,w,h) are obtained, wherein x c is the coordinate of the center point of the boundary frame in the width direction of the image; y cis the coordinate of the center point of the bounding box in the height direction of the image; w is the width of the bounding box; h is the height of the bounding box; (0, 0) is the coordinate of the top left corner of the image; and (1, 1) is the coordinate of the bottom right corner of the image.

[0076] Further, the head coordinates and tail coordinates of each vehicle target, the vehicle type, and the straight-line inter-vehicle distance between adjacent vehicles in the same lane are obtained based on the vehicle bounding box coordinates.

[0077] As a feasible implementation, first, the traffic direction of each lane in the congested section is determined; the traffic direction includes west to east, east to west, south to north, and north to south. According to the vehicle bounding box coordinates of the vehicle template and the traffic direction, the position coordinates of the head line center point are determined to obtain the head coordinates, and the position coordinates of the tail line center point are determined to obtain the tail coordinates. Then the length of the vehicle bounding box is determined as the vehicle length, and the vehicle length is matched in the preset vehicle type determination table to obtain the vehicle type of each vehicle target. The distance between the front vehicle tail coordinates and the rear vehicle head coordinates of adjacent vehicles is calculated, and the distance is converted through a conversion coefficient between the pixel distance and the actual distance to obtain the straight-line inter-vehicle distance between adjacent vehicles in the same lane.

[0078] In one embodiment, the position of the head line center point A in the vehicle bounding box is set as (x i ,y i ), and the position of the tail line center point B is set as (X i ,Y i ). The expression of A and B points is different under different traffic directions, and the specific expression is shown in Table 1. According to Table 1, the calculation and expression of the head coordinates and the tail coordinates of the vehicle in the vehicle flow in different directions are different, and the congested vehicle recognition of the vehicle flow in different directions can be realized in the same coordinate system.

[0079] Table 1

[0080]

[0081] The length w of the vehicle bounding box is taken as the vehicle length, the value is matched with the preset vehicle type determination table to obtain the vehicle type determination result, and the vehicle type determination is shown in Table 2. According to Table 2, according to the different length ranges of the vehicle length, the vehicle type can be divided into small cars (vehicle length less than 4 meters), medium cars (vehicle length between 4 meters and 6 meters), large cars (vehicle length between 6 meters and 12 meters), and super large cars (vehicle length greater than 12 meters).

[0082] Table 2

[0083] Vehicle type Length (m) Small car <4 Mid-size car [4,6] Large car [6,12] Full-size car >12

[0084] Then, the straight-line inter-vehicle distance d i= λdist((X i ,Y i ),(x i+1 ,y i+1 )),i∈Z + , calculate the straight-line vehicle spacing d i between adjacent vehicles i and i+1 in the same lane. Wherein, λ is the conversion factor between pixel distance and actual distance.

[0085] Further, according to the flight speed of the inspection unmanned aerial vehicle, the detected speed of the vehicle identified in the image is corrected to obtain the actual vehicle speed and vehicle acceleration of the vehicle target. Then the RGB values of all pixels in the vehicle identification frame are extracted, and the average value is taken as the vehicle color of the vehicle target.

[0086] In one embodiment, in the traffic video captured by the unmanned aerial vehicle cruising, the directly detected vehicle speed is not the real speed because the unmanned aerial vehicle itself is in motion. In the present application, the vehicle speed is estimated according to v n,i = V n,i -v, and the vehicle acceleration is estimated according to . Wherein, v is the cruising flight speed of the unmanned aerial vehicle; v n,i is the instantaneous speed of vehicle i at the nth frame of image; V n,i is the vehicle detection speed; Δt is the unit time length between each frame of image.

[0087] S103, based on the vehicle feature parameters, a lane-level queuing vehicle discrimination model based on vehicle speed and straight-line vehicle spacing is constructed; and through the lane-level queuing vehicle discrimination model, an ordered set of queuing vehicles of the congested road section is obtained.

[0088] Specifically, according to the vehicle speed and the first preset speed threshold, a queuing vehicle preliminary discrimination model based on the vehicle speed index is constructed, which is used for preliminary determination of queuing vehicles; wherein the first preset speed threshold is related to the lane free flow speed.

[0089] As a feasible implementation, based on the vehicle speed parameters extracted from each frame of image of the video segment, a queuing vehicle preliminary discrimination model is constructed: Wherein, v th1 = αV N , when S1(i) = 1, it is determined that the vehicle i is in a low-speed or stalled state; when S1(i) = 0, it is determined that the vehicle i is in a normal driving state. Wherein, v i is the instantaneous speed of vehicle i; v th1 is the first preset speed threshold; α is the adjustment coefficient; V N is the free flow speed of the Nth lane.

[0090] Further, according to the preliminary determination of the queuing vehicles, taking the tail coordinates of each queuing vehicle as the reference, the upstream vehicle set located in the same lane neighborhood range is searched by a neighborhood search algorithm; wherein, when the head coordinates of the upstream vehicle is located in the neighborhood range, the upstream vehicle is counted into the upstream vehicle set.

[0091] Further, according to the straight-line inter-vehicle distance between the queuing vehicle and each upstream vehicle in the upstream vehicle set and the preset inter-vehicle distance threshold, a queuing vehicle re-determination model based on the straight-line inter-vehicle distance is constructed; wherein, the preset inter-vehicle distance threshold is related to the vehicle speed and deceleration of the upstream vehicle.

[0092] Finally, the queuing vehicle preliminary determination model and the queuing vehicle re-determination model are constructed as a lane-level queuing vehicle determination model.

[0093] Further, the vehicle speed of each vehicle target in the same frame image is input into the queuing vehicle preliminary determination model to obtain the preliminary queuing vehicle sequence in each lane. According to the preliminary queuing vehicle sequence, the corresponding upstream vehicle set is obtained. The straight-line inter-vehicle distance between each queuing vehicle in the preliminary queuing vehicle sequence and each upstream vehicle, and the vehicle speed and deceleration of the upstream vehicle are input into the queuing vehicle re-determination model to obtain the complete queuing vehicle sequence of each lane. The complete queuing vehicle sequences of all lanes are combined into a queuing vehicle set. All queuing vehicle sets in each frame image in the cruise video data are integrated into an ordered queuing vehicle set of the congested road section.

[0094] As a feasible implementation, for the vehicle preliminarily determined to be queued, taking the tail line center point B as the reference, the upstream vehicle set R(i) located in the same lane neighborhood range is searched, and only when the head line center point A of the upstream vehicle is in the neighborhood range, it is recorded into the set, and the search algorithm is as follows: R(i) = {i | r i ≤ 3L veh , -15° ≤ θ(i, i+1) ≤ 15°}; wherein, r is the search neighborhood radius (m); L veh is the standard vehicle length; θ(i, i+1) is the angular deviation between the vehicle i and the adjacent vehicle i+1.

[0095] Then the speed of the upstream vehicle adjacent to i in R(i) is obtained, and whether the speed is less than or equal to v th1 , if so, and the straight-line inter-vehicle distance is less than or equal to the preset inter-vehicle distance threshold, the adjacent upstream vehicle is recorded into the queuing vehicle, and the ordered queuing vehicle set of the same lane is obtained in turn. The queuing vehicle re-determination formula based on the straight-line inter-vehicle distance is as follows: wherein, When S2(i+1) = 1, it is determined that the vehicle i+1 is spatially adjacent to the vehicle i; when S2(i+1) = 0, it is determined that the vehicle i+1 does not belong to the same queue as the vehicle i.d i+1 is the distance between the vehicle i+1 and the vehicle i;d th is the distance threshold between vehicles; β is an adjustment coefficient; v i+1 is the speed of the vehicle i+1; t r is the driver reaction time; a d is the deceleration.

[0096] Based on the process in the above embodiment, the queue vehicle set q in the same frame image is determined, where q is composed of multiple lane queue vehicle subsets in the same frame image, and each subset represents a queue vehicle sequence chain of lane N, which is denoted as P N . Through the time sequence correlation and integration of the consecutive frame queue vehicle sets q1, q2… q n , the ordered sequence set Q n composed of all queue vehicles is finally formed, and the calculation formula is: Q n = Q n-1 ∪q n , n∈Z + and n≥1; wherein, In the formula, when n = 1, Q n-1 is Q0, which represents the initial queue vehicle set, n is the number of UAV cruising video frames; Q n is the ordered set composed of all queue vehicles in n UAV cruising video frame images; q n is the ordered queue vehicle set of the nth frame image.

[0097] S104, according to the ordered queue vehicle set, determine the initial queue length of each lane in the congestion section.

[0098] Specifically, the key frames of the cruising video data are extracted based on the frame difference method, and the key frames are fused and spliced into a panoramic image of the congestion section by using the scale invariant feature transform algorithm SIFT and the fade-in and fade-out weighted average fusion algorithm.

[0099] Further, in the ordered queue vehicle set, the vehicle sequence chain in each lane of the panoramic image of the congestion section is obtained. According to the head coordinates of the head vehicle and the tail coordinates of the tail vehicle in the vehicle sequence chain, the initial queue length of each lane is determined.

[0100] As a feasible implementation method, if the congested road segment is a straight road, the straight-line distance between the coordinates of the front of the first vehicle and the rear of the last vehicle is directly calculated to determine the initial queue length. If the congested road segment is a curved road, the coordinates of the front of the vehicle, the left rear of the vehicle, and the left front of the vehicle behind the vehicle in each adjacent queue are used as curve fitting reference points to fit the road curve. Based on the characteristic parameters of the fitted road curve, the curve spacing between each adjacent queue of vehicles is calculated. Finally, the curve spacing and the length of each vehicle are summed using an accumulation method to obtain the initial queue length for each lane in the congested road segment.

[0101] In one embodiment, Figure 3 A schematic diagram of road curve fitting provided for an embodiment of the present invention, such as... Figure 3 As shown, the positions of points ABA (two adjacent queued vehicles) are used as the reference points for curve fitting of the road linearity, and linear fitting is performed on the curved road. The curve spacing is... The calculation is as follows: In the formula, r is the radius of the fitted circle for the road alignment; (a,b) are the coordinates of the circle's center; (X... i ,Y i ),(x i+1 ,y i+1 ( ) are the coordinates of point B, downstream of two adjacent queued vehicles, and point A, upstream of two adjacent queued vehicles.

[0102] Based on the curve spacing between each pair of adjacent vehicles, and combined with the target vehicle length, the initial queue length L0 is calculated using the following formula: In the formula, I is the number of the last car in the queue.

[0103] S105. Based on the spatiotemporal compensation correction algorithm, the initial queue length is dynamically corrected to obtain the real-time queue length of each lane in the congested road section.

[0104] Specifically, key frame sequences of the appearance of the lead vehicle, the lead vehicle's impending departure, and the appearance of the tail vehicle are identified in the cruise video data; and the operating parameters of the lead vehicle and the tail vehicle are extracted from the key sequence frames; wherein, the operating parameters of the lead vehicle include at least the time of its appearance, the time of its impending departure, and the instantaneous speed and acceleration of the lead vehicle in the key frame image; the operating parameters of the tail vehicle include at least the time of its appearance.

[0105] Furthermore, based on the instantaneous speed and acceleration of the lead vehicle in the keyframe image, and the time difference between sampling times of adjacent keyframes, the estimated speed of the lead vehicle is calculated. Based on the estimated speed and a second preset speed threshold, it is determined whether the lead vehicle has left the queue of vehicles.

[0106] Further, based on the moving distance of the head vehicle and the tail vehicle, a dynamic correction model of real-time queue length is constructed. The head vehicle running parameters and the tail vehicle running parameters are input into the dynamic correction model to obtain the real-time queue length of each lane in the congestion road section.

[0107] As a feasible implementation, in the queue vehicle queue determination process, whether the head vehicle leaves is a key influencing factor. The speed threshold for the head vehicle to leave the queue is set as v th2 = γV N , and the determination formula of whether the head vehicle is in the queue is: wherein, v i′ = v n′,i′ + a n′,i′ T, when S3(i') = 1, it is determined that the head vehicle i is in low speed and still in the queue vehicle queue; when S3(i') = 0, it is determined that the head vehicle i' has normally driven and left the queue vehicle queue. Wherein, v i′ is the estimated speed of the head vehicle i'; T is the time difference of adjacent key frame sampling time; v th2 is the second preset speed threshold required to leave the queue; γ is the adjustment coefficient; V N is the free flow speed of the Nth lane.

[0108] The motion distance of the vehicle in the time interval T1 = t1-t0 from the appearance of the head vehicle to the departure of the head vehicle is recorded as L1. The time difference between the departure of the head vehicle and the appearance of the tail vehicle is recorded as T2 = t2-t1, the time difference between the image of the departure of the head vehicle and the departure of the head vehicle from the queue is recorded as T3, and the dynamic correction model of the real-time queue length L s is constructed as follows:

[0109] wherein, n' = 1, n' = 1, 2, 3, i, j ∈ Z + and i ≤ j ≤ I. In the formula, L s is the corrected real-time queue length; when T3 ≥ T2, it indicates that the head vehicle has not left the queue when the tail vehicle appears, and the head vehicle displacement is L2 at this time; when T3 < T2, the head vehicle has left the queue at the time of determining the tail vehicle, and the position of the new tail vehicle j and the head vehicle displacement L3 after T2 need to be determined.

[0110] Through the queue length dynamic correction method based on space-time compensation, the adaptability of the method to the vehicle slow running or frequent start-stop scene is effectively improved, the real-time performance and reliability of the queue length estimation in the congestion dynamic evolution process are significantly improved, and the method is especially suitable for scenes such as interlaced area of urban expressway and long-distance queue of expressway, thereby providing accurate situation information for traffic control.

[0111] In addition, the embodiment of the present application further provides a vehicle queuing length dynamic estimation system based on unmanned aerial vehicle cruising, as shown in the figure Figure 4 The vehicle queuing length dynamic estimation system 400 based on unmanned aerial vehicle cruising specifically comprises:

[0112] The parameter extraction module 410 is configured to collect cruising video data of the congestion road section by the inspection unmanned aerial vehicle, and extract vehicle feature parameters of the congestion road section by image processing on the cruising video data, wherein the vehicle feature parameters at least include any one or more of the following: head coordinates, tail coordinates, vehicle type, vehicle speed, acceleration, vehicle color and straight-line inter-vehicle distance of each vehicle target and adjacent vehicles in the same lane.

[0113] The queuing vehicle identification module 420 is configured to construct a lane-level queuing vehicle discrimination model based on vehicle speed and straight-line inter-vehicle distance based on the vehicle feature parameters, and obtain an ordered set of queuing vehicles of the congestion road section through the lane-level queuing vehicle discrimination model.

[0114] The queuing length estimation module 430 is configured to determine the initial queuing length of each lane in the congestion road section according to the ordered set of queuing vehicles, and dynamically correct the initial queuing length based on a space-time compensation correction algorithm to obtain the real-time queuing length of each lane in the congestion road section.

[0115] Each of the embodiments in the present application is described in a progressive manner, and the same or similar parts between each of the embodiments can be referred to each other. Each of the embodiments mainly describes the difference from other embodiments. Especially, for the device, equipment and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the related parts can be referred to the part of the method embodiment.

[0116] The above describes the specific embodiments of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

[0117] The above only describes the embodiments of the present application and is not used to limit the present application. The embodiments of the present application can be variously changed and modified by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of the present application shall be included in the protection scope of the present application.

Claims

1. A method for dynamic estimation of vehicle queue length based on UAV cruising, characterized in that, The method comprises: collecting cruise video data of a congested road section by a patrol unmanned aerial vehicle; performing image processing on the cruise video data to extract vehicle feature parameters of the congested road section; wherein the vehicle feature parameters at least include any one or more of the following: head coordinates, tail coordinates, vehicle type, vehicle speed, acceleration, vehicle color, and straight-line inter-vehicle distance between adjacent vehicles in the same lane; based on the vehicle feature parameters, constructing a lane-level queued vehicle discrimination model based on vehicle speed and straight-line inter-vehicle distance two indicators; and obtaining an ordered set of queued vehicles of the congested road section through the lane-level queued vehicle discrimination model; determining the initial queue length of each lane in the congested road section according to the ordered set of queued vehicles; based on a space-time compensation correction algorithm, dynamically correcting the initial queue length to obtain the real-time queue length of each lane in the congested road section. 2.The method of claim 1, wherein, Through the patrol unmanned aerial vehicle, the cruise video data of the congested road section is collected, which specifically includes: real-time monitoring of road traffic running state information based on an electronic map; when congestion occurs in a certain road section and the congestion duration reaches a preset determination condition, automatically triggering an unmanned aerial vehicle cruise task; positioning the head position of the congested road section, and controlling the patrol unmanned aerial vehicle to reach above the head position, starting a reverse cruise mode, flying against the traffic direction, collecting cruise video data and real-time correcting flight trajectory; wherein the flight parameters of the patrol unmanned aerial vehicle are set to a vertical ground low-angle view angle.

3. The method of claim 1, wherein, Performing image processing on the cruise video data to extract vehicle feature parameters of the congested road section, specifically including: extracting continuous image frames in the cruise video data, and identifying vehicle targets in each image frame through a YOLOv8 vehicle detection model to determine a plurality of vehicle targets; tracking the vehicle targets continuously through a DeepSORT motion detection model to obtain vehicle number and vehicle bounding box coordinates of each vehicle target; based on the vehicle bounding box coordinates, obtaining the head coordinates, tail coordinates, vehicle type, and straight-line inter-vehicle distance between adjacent vehicles in the same lane of each vehicle target; correcting the vehicle detection speed of the identified vehicle in the image according to the flight speed of the patrol unmanned aerial vehicle to obtain the actual vehicle speed and vehicle acceleration of the vehicle target; extracting the RGB values of all pixels in the vehicle bounding box and taking the average value as the vehicle color of the vehicle target.

4. The method of claim 3, wherein, Based on the vehicle bounding box coordinates, the head coordinates, tail coordinates, vehicle type, and straight-line inter-vehicle distance between adjacent vehicles in the same lane of each vehicle target are obtained, specifically including: determining the traffic direction of each lane in the congested road section; wherein the traffic direction includes west to east, east to west, south to north, and north to south; determining the position coordinates of the head line center point to obtain the head coordinates, and determining the position coordinates of the tail line center point to obtain the tail coordinates according to the vehicle bounding box coordinates of the vehicle template and the traffic direction. The length of the vehicle recognition frame is determined as a vehicle length, and a vehicle type of each vehicle target is obtained by matching in a preset vehicle type determination table according to the vehicle length; The distance between the front vehicle tail coordinates and the rear vehicle head coordinates of the adjacent vehicles is calculated, and the distance is converted by a conversion coefficient between a pixel distance and an actual distance to obtain a straight-line inter-vehicle distance between the adjacent vehicles in the same lane.

5. The method of claim 1, wherein, Based on the vehicle characteristic parameters, a lane-level queuing vehicle discrimination model based on two indexes of vehicle speed and straight-line inter-vehicle distance is constructed, and the model specifically includes: According to the vehicle speed and a first preset speed threshold, a queuing vehicle preliminary discrimination model based on the vehicle speed index is constructed, which is used for preliminarily determining the queuing vehicle; wherein the first preset speed threshold is related to the lane free flow speed; According to the preliminarily determined queuing vehicle, the tail coordinates of each queuing vehicle are taken as a reference, and a set of upstream vehicles located in the same lane neighborhood range is searched by a neighborhood search algorithm; wherein when the head coordinates of the upstream vehicle are located in the range, the upstream vehicle is counted into the set of upstream vehicles; According to the straight-line inter-vehicle distance between the queuing vehicle and each upstream vehicle in the set of upstream vehicles and a preset inter-vehicle distance threshold, a queuing vehicle re-discrimination model based on the straight-line inter-vehicle distance is constructed; wherein the preset inter-vehicle distance threshold is related to the vehicle speed and deceleration of the upstream vehicle; The queuing vehicle preliminary discrimination model and the queuing vehicle re-discrimination model are constructed as the lane-level queuing vehicle discrimination model.

6. The method of claim 5, wherein, Through the lane-level queuing vehicle discrimination model, an ordered set of queuing vehicles of the congestion road section is obtained, and the ordered set specifically includes: The vehicle speed of each vehicle target in the same frame image is input into the queuing vehicle preliminary discrimination model to obtain a preliminary queuing vehicle sequence in each lane; According to the preliminary queuing vehicle sequence, a corresponding set of upstream vehicles is obtained; The straight-line inter-vehicle distance between each queuing vehicle in the preliminary queuing vehicle sequence and each corresponding upstream vehicle, and the vehicle speed and deceleration of the upstream vehicle are input into the queuing vehicle re-discrimination model to obtain a complete queuing vehicle sequence of each lane; The complete queuing vehicle sequences of all lanes are combined into a queuing vehicle set; All queuing vehicle sets in each frame image in the cruise video data are integrated into the ordered set of queuing vehicles of the congestion road section.

7. The method of claim 1, wherein, According to the ordered set of queuing vehicles, the initial queuing length of each lane in the congestion road section is determined, and the determination specifically includes: Based on a frame difference method, key frames of the cruise video data are extracted, and a scale-invariant feature transformation algorithm SIFT and a fade-in and fade-out weighted average fusion algorithm are used to fuse and splice the key frames into a congestion road section panoramic image; In the ordered set of queuing vehicles, a vehicle sequence chain in each lane in the congestion road section panoramic image is obtained; According to the head coordinates of the head vehicle and the tail coordinates of the tail vehicle in the vehicle sequence chain, the initial queuing length of each lane is determined.

8. The method of claim 7, wherein, According to the head coordinates of the head vehicle and the tail coordinates of the tail vehicle in the vehicle sequence chain, the initial queuing length of each lane is determined, and the determination specifically includes: If the congested road section is a straight road, a straight-line distance between a head coordinate of the head vehicle and a tail coordinate of the tail vehicle is directly calculated to determine the initial queue length; If the congested road section is a curved road, a head coordinate, a tail left side of a front vehicle, and a head left side of a rear vehicle of each adjacent two queued vehicles are taken as curve fitting reference points to perform road curve fitting; According to characteristic parameters of the fitted road curve, a curve distance between each adjacent two queued vehicles is calculated; The curve distance and a vehicle length of each vehicle are accumulated by an accumulation method to obtain the initial queue length of each lane in the congested road section.

9. The method of claim 1, wherein, Based on a space-time compensation correction algorithm, the initial queue length is dynamically corrected to obtain a real-time queue length of each lane in the congested road section, specifically including: Key frame sequences in which a head vehicle appears, the head vehicle is about to leave, and a tail vehicle appears are identified in the cruise video data, and head vehicle running parameters and tail vehicle running parameters are extracted in the key sequence frames; wherein the head vehicle running parameters at least include a head vehicle appearance time, a head vehicle about to leave time, a head vehicle instantaneous speed in a key frame image, and an acceleration; the tail vehicle running parameters at least include a tail vehicle appearance time; An estimated speed of the head vehicle is calculated according to the head vehicle instantaneous speed in the key frame image, the acceleration, and a time difference between adjacent key frame sampling time points; Whether the head vehicle has left the queued vehicle queue is judged according to the estimated speed and a second preset speed threshold; A dynamic correction model of the real-time queue length is constructed based on a moving distance of the head vehicle and the tail vehicle; The head vehicle running parameters and the tail vehicle running parameters are input into the dynamic correction model to obtain the real-time queue length of each lane in the congested road section. 10.A system for dynamic estimation of vehicle queue length based on UAV cruising, characterized in that, The system includes: A parameter extraction module is configured to collect cruise video data of a congested road section by a patrol unmanned aerial vehicle, and extract vehicle feature parameters of the congested road section by image processing on the cruise video data; wherein the vehicle feature parameters at least include any one or more of the following: a head coordinate, a tail coordinate, a vehicle type, a vehicle speed, an acceleration, a vehicle color, and a straight-line distance between adjacent vehicles in the same lane of each vehicle target; A queued vehicle identification module is configured to construct a lane-level queued vehicle discrimination model based on vehicle speed and straight-line distance between vehicles based on the vehicle feature parameters, and obtain an ordered set of queued vehicles of the congested road section by the lane-level queued vehicle discrimination model; A queue length estimation module is configured to determine an initial queue length of each lane in the congested road section according to the ordered set of queued vehicles, and dynamically correct the initial queue length based on a space-time compensation correction algorithm to obtain a real-time queue length of each lane in the congested road section.

Images