A real-time back-calculation method for vehicle flow load on small and medium-span bridges

Abstract

The present invention discloses a real-time back-calculation method for vehicle flow loads of medium and small-span bridges. Based on the deep learning framework of YOLO v5 and DeepSort, secondary development of machine vision algorithms is carried out to accurately obtain information such as the lanes, vehicle speeds, vehicle types, and number of axles of vehicles running on the bridge in real time from traffic monitoring videos; an information storage stack is set up to store the spatio-temporal operation information of vehicles and the corresponding bridge dynamic response information between two stable states; by introducing the gradient descent backpropagation algorithm, the system control equation is improved based on the Moses algorithm to solve the ill-conditioned problem of the control equation of the traditional Moses algorithm, thereby realizing the real-time extraction of vehicle load flow. The method of the present invention realizes the real-time extraction of spatio-temporal information of vehicle flow, and at the same time can combine the statistical distribution law of vehicle weights in the actual local vehicle flow information, resist the interference of data noise, avoid solving the inverse matrix, and identify the information of vehicle flow load with higher accuracy.

Description

Technical Field

[0001] The present invention relates to the field of highway bridge safety monitoring, and particularly to a real-time inverse calculation method for vehicle flow loads of medium and small span bridges. Background Art

[0002] How to ensure the safety and durability of existing bridges is an urgent matter for the transportation infrastructure industry. As one of the basic dynamic loads acting on bridges, vehicle loads have a great impact on the durability of bridges. Accurately obtaining the vehicle load information experienced by the bridge structure and the time and space distribution information of vehicles plays a key role in reconstructing and predicting traffic flow, conducting structural response analysis and prediction, and evaluating the safety of bridge structures.

[0003] In the actual bridge traffic scenario, vehicles have characteristics such as large traffic flow, strong randomness, and diverse vehicle types. In order to obtain the vehicle weight information in the traffic scenario, American scholars introduced Bridge Weigh-In-Motion (abbreviated as BWIM) in the 1980s. The pavement type dynamic weighing system measures the dynamic pressure when the wheel passes through the sensor on the road surface to obtain information such as the axle weight of the vehicle. However, due to reasons such as high cost, easy damage, low efficiency in measuring vehicle weight, and easy traffic congestion, the vehicle dynamic weighing system has not been widely popularized and applied. The bridge dynamic weighing system is to perform inverse analysis of the vehicle weight through the dynamic response of the bridge when the vehicle passes over the bridge. This method does not require the vehicle to decelerate, has high weighing efficiency, and has advantages such as low installation cost, not easy to damage, and low cost, and has great advantages in calculating traffic load information. Currently, the bridge dynamic weighing system performs vehicle weight inverse calculation based on the Moses algorithm. The noise resistance performance of this algorithm is poor. Once the axle spacings of the axles are close, the system control equation is prone to a singular matrix, resulting in poor inverse calculation accuracy of the single axle weight. At the same time, most traditional bridge dynamic weighing systems are limited to the inverse calculation of the vehicle weight of a single vehicle and cannot be applied to the extraction of vehicle flow loads in bridge scenarios. Therefore, it is necessary to improve the traditional bridge dynamic weighing system. Summary of the Invention

[0004] The object of the present invention is to provide a real-time inverse calculation method for vehicle flow loads of medium and small span bridges. This method can realize the extraction of spatio-temporal information of vehicle flow based on the framework of deep learning. Introducing the BP-Moses algorithm can effectively resist the interference of noise. The error between the measured value and the predicted value of the dynamic response caused by the vehicle on the bridge is backpropagated through the gradient descent algorithm, and the vehicle axle weight is continuously corrected, avoiding the solution of the inverse matrix of the system control equation of the Moses algorithm and improving the recognition accuracy of the vehicle axle weight.

[0005] To achieve the above functions, the present invention designs a real-time back-calculation method for vehicle flow loads on medium and small-span bridges. For each vehicle running on the target bridge without branch roads, the following steps S1 - S4 are executed to achieve the real-time extraction of vehicle load flow:

[0006] Step S1: Use a camera to collect videos of each vehicle running on the target bridge. Through the YOLOv5 deep learning algorithm, identify each vehicle in the video, frame the positions of each vehicle in the video, and detect the vehicle position information of each vehicle in the video at a moment that spans a preset duration in the historical time direction starting from the current moment;

[0007] Step S2: Adopt the target tracking algorithm framework DeepSort to construct a vehicle tracking model for urban bridges. Based on the vehicle position information of each vehicle in the video at the preset time points obtained in Step S1, predict the positions of each vehicle in the video at the current moment;

[0008] Step S3: Adopt the direct linear transformation method. Based on the vehicle position information obtained in Step S1 and Step S2, convert the positions of each vehicle in the video based on the pixel coordinate system to the world coordinate system, and solve the axle spacing of each vehicle, the running speed of the vehicle, and the actual moving distance of the vehicle in the world coordinate system to obtain the vehicle running information of each vehicle on the target bridge. The vehicle running information includes vehicle lanes and vehicle speeds;

[0009] Step S4: Through bridge dynamic sensors, monitor the bridge dynamic response information at the current moment, construct the bridge dynamic response information corresponding to the current moment and the vehicle running information of each passing vehicle. Based on the real-time execution of Steps S1 to S4, perform the following Steps A and B:

[0010] Step A: Based on the accumulation of the bridge dynamic response information at each moment and the vehicle running information of each passing vehicle, take the situation when no vehicle passes on the target bridge as a stable state, set up an information storage stack, and store the bridge dynamic response information at each moment and the vehicle running information of each passing vehicle between two adjacent stable states in the information storage stack;

[0011] Step B: Take the bridge dynamic response information and vehicle running information in the information storage stack obtained in Step A as inputs, take the axle weights of each vehicle as outputs, and based on the Moses algorithm, by introducing the gradient descent backpropagation algorithm, construct a vehicle weight back-calculation model. Through iterative correction for a preset number of times, obtain a corrected vehicle weight back-calculation model, and apply the vehicle weight back-calculation model to achieve the real-time extraction of vehicle load flow.

[0012] As a preferred technical solution of the present invention: The specific steps of Step S1 are as follows:

[0013] Step S11: Collect the videos of each vehicle running on the target bridge, convert the videos of each vehicle into image samples arranged in time sequence, frame the positions of each vehicle in the image samples, and select 5000 image samples to construct a dataset. Among them, each vehicle is divided into 8 vehicle types according to the vehicle category and the number of axles, which are: Sedan car, bus, Minibus, 2A truck, 3A truck, 4A truck, 5A truck, 6A truck. Among them, for 2A truck, 2A represents the number of axles is 2, and truck represents the vehicle category is truck. Sedan car represents the vehicle category is sedan, and bus represents the vehicle category is bus, and Mini bus represents the vehicle category is minibus;

[0014] Step S12: Use the image samples of each vehicle as input, the vehicle type as the training label, and as the output, where represents the vehicle type of the i-th vehicle, respectively represent the minimum pixel coordinate value and the maximum pixel coordinate value of the x-axis coordinate where the detection frame of the i-th vehicle is located, respectively represent the minimum pixel coordinate value and the maximum pixel coordinate value of the y-axis coordinate where the detection frame of the i-th vehicle is located, and train the YOLOv5 neural network, adjust the weight coefficients of the YOLOv5 neural network until the average precision of each vehicle type reaches more than 0.9;

[0015] Step S13: Determine the position of the detection frame of vehicle i in the pixel coordinate system through point and point

[0016] As a preferred technical solution of the present invention: The specific steps of Step S2 are as follows:

[0017] Step S21: Read the position of the detection frame of the vehicle in the current frame image sample and the depth features of each detection frame image block;

[0018] Step S22: Calculate the confidence of each detection frame, filter each detection frame according to the confidence, and delete the detection frames with a confidence lower than the preset value;

[0019] Step S23: If at least two detection frames are framed for the same vehicle, perform non-maximum suppression on the detection frames to eliminate redundant detection frames;

[0020] Step S24: Based on the Kalman filtering method, predict the position of the vehicle in the current frame image sample.

[0021] As a preferred technical solution of the present invention: The specific steps of Step S3 are as follows: ​

[0022] Step S31: Convert the vehicle position in the pixel coordinate system in the image sample to the world coordinate system. The specific method is as follows:

[0023] Convert the coordinates in the pixel coordinate system xO 1 y to the coordinates in the camera coordinate system O 2 X c Y c Z c where O 1 represents the origin of the image coordinate system, and O 2 represents the origin of the camera coordinate system;

[0024]

[0025] where f is the camera focal length, x and y are the x-axis coordinate and y-axis coordinate in the pixel coordinate system respectively, and X c , Y c , Z c are the x-axis coordinate, y-axis coordinate, and z-axis coordinate in the camera coordinate system respectively;

[0026] It is represented in matrix form as:

[0027]

[0028] Convert the coordinates in the camera coordinate system O 2 X c Y c Z c to the coordinates in the world coordinate system O 3 X w Y w Z w where O 3 represents the origin of the world coordinate system;

[0029]

[0030] In the formula: R is a 3×3 rotation matrix; t is a 3×1 translation vector; X w , Y w , Z w are the x-axis coordinate, y-axis coordinate, and z-axis coordinate in the world coordinate system respectively;

[0031] Obtain the transformation formula from the world coordinate system (X w , Y w , Z w ) to the pixel coordinate system xO 1 y as follows:

[0032]

[0033] The bridge deck height is fixed, and Z w is a constant. Let Z w = 0, and the matrix M is as follows:

[0034]

[0035] Through the inverse transformation of the matrix, let respectively obtain the coordinates in the world coordinate system corresponding to

[0036] Step S32: According to the position information of the vehicle detection frame obtained in Step S31 in the world coordinate system Judge the lane information to which the vehicle belongs through the center coordinates of the detection frame <000026>indicates that the i-th vehicle is in the k-th lane;

[0037] Step S33: Detect the vehicle speed. The specific method is as follows:

[0038] Based on the center coordinates of the detection frame According to the following formula, let calculate the detection line L 1 and the detection line L 2 :

[0039] L 1 : a 1 X w + b 1 Y w + c 1 = 0

[0040] L 2 : a 2 X w + b 2 Y w + c 2 = 0

[0041] And calculate the judgment flag as follows:

[0042]

[0043] where a 1 b 1 c 1 a 2 b 2 c 2 are all parameters, which are set according to the preset bridge scenario and the preset camera angle;

[0044] When flag ≤ 0, calculate the vehicle i on the detection line L1 and L 2 The actual time t experienced between i is as follows:

[0045]

[0046] In the formula, A i m is the initial number of frames of the image sample, A i n is the final number of frames of the image sample, and FPS is the frame rate of the camera;

[0047] Step 3.4: Detect the line L 1 and L 2 The actual distance between is L, and calculate the driving speed v of vehicle i i as follows:

[0048]

[0049] In the formula, t i is the actual time experienced by vehicle i between the detection lines L 1 and L 2 and.

[0050] As a preferred technical solution of the present invention: The specific steps of step A in step S4 are as follows:

[0051] Step S41: Use a bridge dynamic sensor to monitor the bridge dynamic response information, and set up an information storage stack. When the vehicle starts to couple with the target bridge, the information storage stack starts to record the vehicle's running information and bridge dynamic response information;

[0052] Step S42: When the vehicle stops coupling with the target bridge, the information storage stack stops recording, and the vehicle's running information and bridge dynamic response information in the information storage stack are imported into the vehicle weight inverse calculation model in step S5 for vehicle weight inverse calculation;

[0053] Step S43: Empty the information storage stack. When the subsequent vehicle gets on the bridge, repeat steps S41 - S42, and the information storage stack continues to store the vehicle's running information and bridge dynamic response information.

[0054] As a preferred technical solution of the present invention: The vehicle's running information includes the initial number of frames A of the image sample i m , the final number of frames A of the image sample i n , lane the driving speed v of the vehicle i ; The bridge dynamic response information includes the mid-span deflection of the bridge, the mid-span main girder strain of the bridge, and the reaction force of the bridge bearing.

[0055] As a preferred technical solution of the present invention, the specific steps of step B in step S4 are as follows:

[0056] Step S51: Based on the Moses algorithm, calculate the error E between the measured value and the predicted value of the dynamic response signal caused by the vehicle on the target bridge as follows:

[0057] E = (ε m - Aw) T (ε m - Aw)

[0058] Where ε m represents the vector matrix of the measured values of the bridge dynamic response signal, with a dimension of N t × 1; A represents the vector matrix of the influence line coefficient at the mid-span of the bridge for N a axles at N t moments, with a dimension of N t × N a ; w represents the vector matrix of axle weights, with a dimension of N a × 1;

[0059] Step S52: Through the gradient descent backpropagation algorithm, adjust the axle weights to reduce the error between the measured value and the predicted value of the bridge dynamic response signal, specifically as follows:

[0060]

[0061] Step S53: Set the partial derivative of the reduction of the error between the measured value and the predicted value of the bridge dynamic response signal with respect to the vehicle axle weight to 0 to obtain the axle weight matrix w as follows:

[0062] w = (A T A) -1 A T ε m

[0063] Step S54: According to the probability distribution relationship of the axle weights of the vehicle model to which the vehicle belongs, assign an initial value w to the axle weights 0 ;

[0064] Step S55: Solve the error E between the predicted value and the measured value of the dynamic response signal caused by the vehicle with axle weight w j on the target bridge in the 0th round of iteration, and perform iterative correction on the axle weights according to the following formula: 0 where η is the learning step size, w

[0065]

[0066] where η is the learning step size, w jThe axle weight vector matrix is corrected for the j-th round of iteration; E j represents the error between the predicted value and the measured value of the dynamic response signal caused by a vehicle with an axle weight of w j on the target bridge in the j-th round of iteration;

[0067] Substituting the formula in step S52 into the above formula and simplifying, we can obtain:

[0068] w j+1 = w j - 2ηA T (ε m - Aw j )

[0069] Step S56: Set the error limit E p , repeat step S55 to iteratively correct the axle weight. If E j+1 < E p , then end the iterative correction and obtain the axle weight w = w j+1 .

[0070] Advantageous effects: Compared with the prior art, the advantages of the present invention include:

[0071] 1. The method of the present invention provides a basis for initializing the axle weight by inverse calculation of vehicle weight in the subsequent process by secondary development of the deep learning framework, selecting monitoring video photos in the bridge traffic scenario to formulate a labeled data set, and classifying vehicles according to two dimensions of vehicle type and axle.

[0072] 2. The present invention sets the concept of a vehicle information storage stack, which can make full use of the whole process information of vehicle driving and the bridge dynamic response signal, has high universality, and can be migrated to the vehicle weight inverse calculation algorithm of other sample bridges.

[0073] 3. The BP-Moses algorithm introduced in the present invention can avoid the ill-conditioned problem of the system control state equation caused by the too close distance between axles, and at the same time can combine the actual vehicle weight distribution law of the local area, resist the interference of data noise, avoid solving the inverse matrix, and identify the vehicle flow load information with higher accuracy. Brief description of the drawings

[0074] Figure 1 is a flowchart of the real-time inverse calculation method of vehicle flow load for medium and small span bridges provided by an embodiment of the present invention;

[0075] Figure 2 is a vehicle type classification diagram provided by an embodiment of the present invention;

[0076] Figure 3 is a machine vision target detection and recognition diagram provided by an embodiment of the present invention;

[0077] Figure 4(a) is a top view of a bridge with sensors arranged according to an embodiment of the present invention;

[0078] Figure 4(b) is a cross-sectional view of the mid-span of a bridge with sensors arranged according to an embodiment of the present invention;

[0079] Figure 5 is a diagram showing the axle and axle distance of a three-axle vehicle provided according to an embodiment of the present invention;

[0080] Figure 6 is a diagram showing the division of the test lanes on the bridge deck provided according to an embodiment of the present invention;

[0081] Figure 7 is an influence surface diagram for the strain of measuring point 1 at the mid-span of main girder G1 provided according to an embodiment of the present invention;

[0082] Figure 8(a) is a diagram showing the results of back-calculating the vehicle weight and axle weight by the Moses algorithm provided according to an embodiment of the present invention;

[0083] Figure 8(b) is a diagram showing the results of back-calculating the vehicle weight and axle weight by the BP-Moses algorithm provided according to an embodiment of the present invention. Detailed implementation manners

[0084] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and cannot be used to limit the protection scope of the present invention.

[0085] Refer to Figure 1 An on-line vehicle load real-time back-calculation method for small and medium-span bridges provided by an embodiment of the present invention performs the following steps S1-S4 for each vehicle running on a target bridge without branch roads to realize the real-time extraction of vehicle load flow:

[0086] Step S1: Use a camera to collect videos of each vehicle running on the target bridge, identify each vehicle in the video through the YOLOv5 deep learning algorithm, frame the positions of each vehicle in the video, and detect the vehicle position information of each vehicle in the video at a moment that spans a preset time period in the historical time direction starting from the current moment;

[0087] The specific application steps of the YOLO v5 algorithm are as follows: First, make a training set, extract vehicle photos from traffic monitoring videos, and divide the vehicles into 8 categories according to axle weight and axle distance, such as Figure 2As shown in sequence are Sedan car, Mini bus, bus, 2A truck, 3A truck, 4A truck, 5A truck, 6A truck. Then, train the neural network, load the training data into the network for training, and adjust the weight coefficients of the neural network. Finally, test the network, read the traffic video information into the neural network, and check the recognition accuracy and efficiency. The recognition results are as Figure 3 shown. In this traffic scenario, the YOLO v5 algorithm can identify and classify vehicles. The vehicle recognition confidence is above 0.95, and the axle recognition confidence is above 0.7.

[0088] Step S2: Adopt the target tracking algorithm framework DeepSort to construct a vehicle tracking model for urban bridges. Based on the vehicle position information of each vehicle in the video at the preset time point obtained in Step S1, predict the position of each vehicle in the video at the current moment.

[0089] The DeepSort model mainly combines the Hungarian algorithm, recursive Kalman filtering, and single hypothesis of frame-by-frame data association. Among them, the Hungarian algorithm realizes the matching function of the targets in the current frame image and the targets in the next frame image, and the Kalman filtering can predict the position at the current moment based on the position of the target at the previous moment.

[0090] Step S3: Adopt the direct linear transformation method. Based on the vehicle position information obtained in Step S1 and Step S2, convert the position of each vehicle in the video based on the pixel coordinate system to the world coordinate system, and solve the axle spacing, running speed of the vehicle, and the actual moving distance of the vehicle in the world coordinate system to obtain the vehicle running information of each vehicle on the target bridge. The vehicle running information includes vehicle lanes and vehicle speed.

[0091] Based on the deep learning framework of YOLO v5 and DeepSort, secondary development is carried out. It is obtained from the vehicle type recognition by the YOLO v5 algorithm. The lane information can be determined by the projection of the detection box on the bridge deck. The vehicle speed information can be obtained by setting two detection lines horizontally on the road surface, using the DeepSort algorithm to associate the running trajectories of the vehicles, and then obtaining the time when the vehicle passes between the two detection lines. Then, divide the distance between the detection lines by the time when the vehicle passes between the two detection lines to obtain the running speed of the vehicle, thereby extracting the spatio-temporal signal of the traffic flow in the bridge scene.

[0092] The specific steps of Step S3 are as follows:

[0093] Step S31: Convert the vehicle position based on the pixel coordinate system in the image sample to the world coordinate system. The specific method is as follows:

[0094] Convert the pixel coordinate system xO 1The coordinates in y are transformed into the camera coordinate system O through the following formula 2 X c Y c Z c where O 1 represents the origin of the image coordinate system, and O 2 represents the origin of the camera coordinate system;

[0095]

[0096] where f is the camera focal length, x and y are the x-axis coordinate and y-axis coordinate in the pixel coordinate system respectively, and X c 、 Y c 、Z c are the x-axis coordinate, y-axis coordinate, and z-axis coordinate in the camera coordinate system respectively;

[0097] It is represented in matrix form as:

[0098]

[0099] The coordinates in the camera coordinate system O 2 X c Y c Z c are transformed into the world coordinate system O 3 X w Y w Z w through the following formula, where O 3 represents the origin of the world coordinate system;

[0100]

[0101] In the formula: R is a 3×3 rotation matrix; t is a 3×1 translation vector; X w 、Y w 、Z w are the x-axis coordinate, y-axis coordinate, and z-axis coordinate in the world coordinate system respectively;

[0102] The transformation formula from the world coordinate system (X w ,Y w ,Z w ) to the pixel coordinate system xO 1 y is as follows:

[0103]

[0104] M is a 3×4 matrix, and this matrix has 11 unknown parameters, which need to be determined by at least the coordinates of 6 reference points. Since the bridge deck height remains unchanged, Z w can be regarded as a constant. Let Z w= 0, so there can be only 8 independent unknowns, and 4 reference points are needed to determine the matrix M. The matrix M is as follows:

[0105]

[0106] Through the inverse transformation of the matrix, let Obtain respectively The coordinates in the world coordinate system corresponding to

[0107] Step S32: According to the position information of the vehicle detection frame obtained in step S31 in the world coordinate system Based on the center coordinates of the detection frame Judge the lane information to which the vehicle belongs Indicates that the i-th vehicle is in the k-th lane;

[0108] Step S33: Detect the vehicle speed. The specific method is as follows:

[0109] Based on the center coordinates of the detection frame According to the following formula, let Calculate the detection line L 1 、Detection line L 2 :

[0110] L 1 :a 1 X w +b 1 Y w +c 1 = 0

[0111] L 2 :a 2 X w +b 2 Y w +c 2 = 0

[0112] And calculate the judgment flag flag as follows:

[0113]

[0114] Where a 1 、b 1 、c 1 、a 2 、b 2 、c 2 Are all parameters, which are set according to the preset bridge scene and the preset camera angle;

[0115] In one embodiment, the detection line L 1 (a1 X w +b 1 Y w +c 1 =0, where a 1 =0.0358,b 1 = 1,c 1 =-827.684) and detection line L 2 (a 2 X w +b 2 Y w +c 2 =0, where a 2 =0.0841,b 2 = 1,c 2 =-649.308) is shown in FIG4(a), and after coordinate transformation, it is shown in FIG4(b).

[0116] That is, when the center coordinate of the detection box Located at detection line L 1 and L 2 On both sides, flag>0, when the center coordinate of the detection box Located at detection line L 1 and L 2 When flag is less than 0, flag≤0. Therefore, by recording the camera's start and end frame numbers A when flag≤0 i m and A i n Then we can get the vehicle detection L 1 and L 2 The number of frames running between the detection line L can be obtained. 1 and L 2 The actual time t elapsed between i for:

[0117]

[0118] Among them, FPS is the frame rate of the camera.

[0119] Detection line L 1 and L 2 The actual distance between them is L, and the speed v of vehicle i is calculated. i As follows:

[0120]

[0121] Among them, L is obtained by actual measurement or through the virtual and real distribution of lane lines in the scene;

[0122] Step S4: Monitor the bridge dynamic response information at the current moment through bridge dynamic sensors, construct the bridge dynamic response information corresponding to the current moment and the vehicle operation information of each vehicle passing by. Based on the real-time execution of Steps S1 to S4, perform the following Steps A and B:

[0123] Step A: Based on the accumulation of the bridge dynamic response information at each moment and the vehicle operation information of each vehicle passing by, take the situation when no vehicle passes through the target bridge as a stable state, set up an information storage stack, and store the bridge dynamic response information at each moment and the vehicle operation information of each vehicle passing by between two adjacent stable states in the information storage stack;

[0124] In one embodiment, a typical box girder bridge is selected for simulation experiments. Figures 4(a) and 4(b) respectively show the top view and the mid-span cross-sectional view of the experimental bridge. The bridge is a simply supported box girder bridge with a longitudinal span of 30 m and a transverse width of 13 m, which is composed of four hollow box girder beams with a width of 3.25 m and a height of 1.6 m spliced together. This BWIM system collects the time-history strain response signals of measuring points 1, 2, 3, and 4 along the longitudinal bridge direction at the mid-span main girders (G1, G2, G3, G4) by arranging strain sensors, and at the same time, arranges cameras at the bridge tail BB' end to collect information such as the time and space of running vehicles, vehicle types, and axles. Figure 5 Shows that the present invention selects a typical three-axle vehicle for simulation experiments, and the axle weights of the three axles are 12 t, 8 t, and 8 t in sequence, and the axle spacings are 3.8 m and 1.35 m in sequence. Figure 6 Shows that in this example, the bridge is horizontally divided into six lane lines, and the test vehicle moves the vehicle center along the lane line. Figure 7 Shows the influence surface based on the strain of measuring point 1 at the mid-span of main girder G1 in this example.

[0125] Step B: Take the bridge dynamic response information and vehicle operation information in the information storage stack obtained in Step A as inputs, take the axle weights of each vehicle as outputs, based on the Moses algorithm, by introducing the gradient descent backpropagation algorithm, construct a vehicle weight back-calculation model, and through a preset number of iterative corrections, obtain a corrected vehicle weight back-calculation model, and apply the vehicle weight back-calculation model to realize the real-time extraction of vehicle load flow.

[0126] The specific steps of Step B are as follows:

[0127] Step S51: The algorithm for the vehicle weight of the traditional Bridge Weigh-In-Motion (BWIM) system is basically the Moses algorithm, which calculates the vehicle axle weight by minimizing the error between the measured value and the predicted value of the dynamic response caused by the vehicle on the bridge;

[0128] Based on the Moses algorithm, the error E between the measured value and the predicted value of the dynamic response signal caused by the vehicle on the target bridge is calculated as follows:

[0129] E = (ε m - Aw) T (ε m - Aw)

[0130] where ε m represents the vector matrix of the measured values of the bridge dynamic response signal, with a dimension of N t × 1; A represents the vector matrix of the influence line coefficient at the mid-span of the bridge for N a axles at N t moments, with a dimension of N t × N a ; w represents the vector matrix of the axle weights, with a dimension of N a × 1;

[0131] Step S52: Adjust the axle weights through the gradient descent backpropagation algorithm to reduce the error between the measured value and the predicted value of the bridge dynamic response signal, as follows:

[0132]

[0133] Step S53: Set the partial derivative of the reduction of the error between the measured value and the predicted value of the bridge dynamic response signal with respect to the vehicle axle weight to 0, and obtain the axle weight matrix w as follows:

[0134] w = (A T A) -1 A T ε m

[0135] Step S54: To solve the ill-posed inverse problem caused by matrix inversion in the system control equation of the Moses algorithm, the present invention proposes the BP-Moses algorithm, that is, by introducing the gradient descent backpropagation algorithm to improve its system control equation on the basis of the Moses algorithm to solve the ill-conditioned problem of the control equation. Since the type of the vehicle and the number of axles of the vehicle are already known, the initial value w 0 of the axle weight can be assigned according to the probability distribution relationship of the axle weights of the vehicle model;

[0136] Step S55: Solve the error E j between the predicted value and the measured value of the dynamic response signal caused by the vehicle with axle weight w 0 in the 0th round of iteration, and iteratively correct the axle weight according to the following formula:

[0137]

[0138] where η is the learning step size, and w j is the axle weight vector matrix corrected in the j-th iteration; E j represents the error between the predicted value and the measured value of the dynamic response signal caused by the vehicle with axle weight w j on the target bridge in the j-th iteration;

[0139] Substituting the formula in step S52 into the above formula and simplifying, we can obtain:

[0140] w j+1 = w j - 2ηA T (ε m - Aw j )

[0141] Step S56: Set the error limit E p , repeat step S55 to iteratively correct the axle weight. If E j+1 < E p , then end the iterative correction and obtain the axle weight w = w j+1 .

[0142] In one embodiment, by simulating the vehicle passing through lanes 1 to 6 at a speed of 60 km / h, the six working conditions are named working condition 1, working condition 2,..., working condition 6 in sequence. Figures 8(a) and 8(b) show the percentage errors of identifying the axle weight and vehicle weight by the Moses algorithm and the BP-Moses algorithm during the driving process of the upper three-axle test vehicle under different working conditions. Due to the fact that the second axle and the third axle of the test three-axle vehicle are too close, the system control equation of the Moses algorithm will have a pathological property, resulting in a large back-calculation error. However, the BP-Moses algorithm proposed by the present invention can effectively separate the axle weights of axles that are relatively close by introducing the gradient descent method to correct the axle weights, thereby improving the identification results of the second and third axle weights.

[0143] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those of ordinary skill in the art.

Claims

1. A real-time inverse calculation method for vehicle flow load on medium and small-span bridges, characterized in that, for each vehicle running on the target bridge without branch roads, the following steps S1 - S4 are executed to realize the real-time extraction of vehicle load flow: Step S1: Use a camera to collect videos of each vehicle running on the target bridge. Through the YOLOv5 deep learning algorithm, identify each vehicle in the video, frame the positions of each vehicle in the video, and detect the vehicle position information of each vehicle in the video at a moment that spans a preset duration in the historical time direction starting from the current moment; Step S2: Adopt the target tracking algorithm framework DeepSort to construct a vehicle tracking model for urban bridges. Based on the vehicle position information of each vehicle in the video at the preset time point obtained in Step S1, predict the position of each vehicle in the video at the current moment; Step S3: Adopt the direct linear transformation method. Based on the vehicle position information obtained in Step S1 and Step S2, convert the positions of each vehicle in the video based on the pixel coordinate system to the world coordinate system, and solve the axle spacing of each vehicle, the running speed of the vehicle, and the actual moving distance of the vehicle in the world coordinate system to obtain the vehicle running information of each vehicle on the target bridge. The vehicle running information includes vehicle lanes and vehicle speeds; Step S4: Through bridge dynamic sensors, monitor the bridge dynamic response information at the current moment, construct the bridge dynamic response information corresponding to the current moment and the vehicle running information of each vehicle passing by. Based on the real-time execution of Steps S1 to S4, perform the following Steps A and B: Step A: Based on the accumulation of the bridge dynamic response information at each moment and the vehicle running information of each vehicle passing by, take the situation when no vehicle passes on the target bridge as a stable state, and set up an information storage stack. Store the bridge dynamic response information at each moment and the vehicle running information of each vehicle passing by between two adjacent stable states in the information storage stack; Step B: Use the bridge dynamic response information and vehicle running information in the information storage stack obtained in Step A as inputs, and the axle weight of each vehicle as the output. Based on the Moses algorithm, by introducing the gradient descent backpropagation algorithm, construct a vehicle weight inverse calculation model. Through iterative correction for a preset number of times, obtain a corrected vehicle weight inverse calculation model, and apply the vehicle weight inverse calculation model to realize the real-time extraction of vehicle load flow.

2. The real-time inverse calculation method for vehicle flow load on medium and small-span bridges according to claim 1, characterized in that, The specific steps of Step S1 are as follows: Step S11: Collect videos of each vehicle running on the target bridge, convert the videos of each vehicle into image samples arranged in time sequence, frame the positions of each vehicle in the image samples, and select 5000 image samples to construct a dataset. Among them, each vehicle is divided into 8 vehicle types according to vehicle category and the number of axles, which are: Sedan car, bus, Minibus, 2A truck, 3A truck, 4A truck, 5A truck, 6A truck. Among them, for 2A truck, 2A represents the number of axles is 2, and truck represents the vehicle category is truck. Sedan car represents the vehicle category is sedan, and bus represents the vehicle category is bus, and Mini bus represents the vehicle category is minibus; Step S12: Using the image samples of each vehicle as input, the vehicle type as the training label, and as the output, where represents the vehicle type of the i-th vehicle, respectively represent the minimum pixel coordinate value and the maximum pixel coordinate value of the x-axis coordinate where the detection frame of the i-th vehicle is located, respectively represent the minimum pixel coordinate value and the maximum pixel coordinate value of the y-axis coordinate where the detection frame of the i-th vehicle is located, train the YOLOv5 neural network, and adjust the weight coefficients of the YOLOv5 neural network until the average precision of each vehicle type reaches more than 0.9; Step S13: Determine the position of the detection box of vehicle i in the pixel coordinate system through point and point .

3. A real-time back-calculation method for vehicle flow load on small and medium-span bridges according to claim 2, characterized in that, The specific steps of step S2 are as follows: Step S21: Read the position of the detection frame of the vehicle in the current frame image sample and the depth features of each detection frame image block; Step S22: Calculate the confidence of each detection frame, filter each detection frame according to the confidence, and delete the detection frames with a confidence lower than the preset value; Step S23: If at least two detection frames are framed for the same vehicle, perform non-maximum suppression on the detection frames to eliminate redundant detection frames; Step S24: Based on the Kalman filtering method, predict the position of the vehicle in the current frame image sample.

4. A real-time back-calculation method for vehicle flow load on small and medium-span bridges according to claim 3, characterized in that, The specific steps of step S3 are as follows: Step S31: Convert the vehicle position based on the pixel coordinate system in the image sample into the world coordinate system. The specific method is as follows: Convert the coordinates in the pixel coordinate system xO 1 y to the coordinates in the camera coordinate system O 2 X c Y c Z c , where O 1 represents the origin of the image coordinate system, and O 2 represents the origin of the camera coordinate system; where f is the camera focal length, x and y are the x-axis coordinate and y-axis coordinate in the pixel coordinate system, and X c , Y c , Z c are the x-axis coordinate, y-axis coordinate, and z-axis coordinate in the camera coordinate system, respectively; It is represented in matrix form as: Convert the coordinates in the camera coordinate system O 2 X c Y c Z c to the coordinates in the world coordinate system O 3 X w Y w Z w where O 3 represents the origin of the world coordinate system; where: R is a 3×3 rotation matrix; t is a 3×1 translation vector; X w , Y w , Z w are the x-axis coordinate, y-axis coordinate, and z-axis coordinate in the world coordinate system, respectively; Obtain the transformation formula from the world coordinate system (X w , Y w , Z w ) to the pixel coordinate system xO 1 y as follows: The bridge deck height is fixed, and Z w is a constant. Let Z w = 0, and the matrix M is as follows: Through the inverse transformation of the matrix, let obtain respectively the coordinates in the world coordinate system corresponding to Step S32: Based on the position information of the vehicle detection box obtained in step S31 Through the center coordinates of the detection box Judge the lane information to which the vehicle belongs Indicates that the i-th vehicle is in the k-th lane; Step S33: Detect the vehicle speed. The specific method is as follows: Based on the center coordinates of the detection box According to the following formula, let Calculate the detection line L 1 and the detection line L 2 : L 1 : a 1 X w + b 1 Y w + c 1 = 0 L 2 : a 2 X w + b 2 Y w + c 2 = 0 And calculate the judgment flag flag as follows: where a 1 , b 1 , c 1 , a 2 , b 2 , c 2 are all parameters, which are set according to the preset bridge scenario and the preset camera angle; When flag ≤ 0, calculate the actual time t that vehicle i experiences between detection lines L 1 and L 2 as follows: i The following equation: Where A i m is the initial number of frames of the image sample, and A i n is the final number of frames of the image sample, and FPS is the frame rate of the camera; Step 3.4: Detect the detection line L 1 and L 2 The actual distance between them is L, and calculate the driving speed v of vehicle i i as follows: where t i is the actual time elapsed by vehicle i between 1 detection lines L 2 and L 5. A real-time back-calculation method for vehicle flow load on small and medium-span bridges according to claim 4, characterized in that, The specific steps of step A of step S4 are as follows: Step S41: Use a bridge dynamic sensor to monitor the bridge dynamic response information, and set up an information storage stack. When the vehicle starts to couple with the target bridge, the information storage stack starts to record the vehicle operation information and the bridge dynamic response information; Step S42: When the vehicle stops coupling with the target bridge, the information storage stack stops recording, and imports the vehicle operation information and the bridge dynamic response information in the information storage stack into the vehicle weight back-calculation model in step S5 for vehicle weight back-calculation; Step S43: Empty the information storage stack. When the subsequent vehicle gets on the bridge, repeat steps S41 - S42, and the information storage stack continues to store the vehicle operation information and the bridge dynamic response information.

6. A real-time back-calculation method for vehicle flow load on small and medium-span bridges according to claim 5, characterized in that, The operating information of the vehicle includes the initial number of frames A of the image sample i m , the final number of frames A of the image sample i n , the lane the vehicle driving speed v i ; The dynamic response information of the bridge includes the mid-span deflection of the bridge, the strain of the main girder at the mid-span of the bridge, and the reaction force of the bridge bearing.

7. A real-time back-calculation method for vehicle flow load on small and medium-span bridges according to claim 5, characterized in that, The specific steps of step B of step S4 are as follows: Step S51: Based on the Moses algorithm, calculate the error E between the measured value and the predicted value of the dynamic response signal caused by the vehicle on the target bridge as follows: E = (ε m - Aw) T (ε m - Aw) Among them, ε m represents the measurement value vector matrix of the bridge dynamic response signal, with a dimension of N t ×1; A represents N a axles' influence line coefficient vector matrix at the mid-span of the bridge at N t moments, with a dimension of N t ×N a ; w represents the vector matrix of axle weights, with a dimension of N a ×1; Step S52: Through the gradient descent backpropagation algorithm, adjust the axle weight to reduce the error between the measured value and the predicted value of the bridge dynamic response signal, specifically as follows: Step S53: Set the partial derivative of the error between the measured value and the predicted value of the bridge dynamic response signal with respect to the vehicle axle weight to 0 to obtain the axle weight matrix w as follows: w = (A T A) -1 A T ε m Step S54: Assign an initial value w to the axle weight according to the probability distribution relationship of the axle weight of the vehicle model to which the vehicle belongs 0 ; Step S55: Solve for the error E between the predicted value and the measured value of the dynamic response signal caused by a vehicle with an axle load of w in the 0th round of iteration j on the target bridge 0 , and iteratively correct the axle load according to the following formula: where η is the learning step size, and w j is the axle weight vector matrix corrected by the j-th round of iteration; E j represents the error between the predicted value and the measured value of the dynamic response signal caused by the vehicle with axle weight w j on the target bridge in the j-th round of iteration; Substitute the formula in Step S52 into the above formula and simplify to get: w j+1 = w j - 2ηA T (ε m - Aw j ) Step S56: Set the error limit E p , repeat step S55 to iteratively correct the axle weight. If E j+1 < E p , then end the iterative correction and obtain the axle weight w = w j+1 .

Images