CN115661749A - Vehicle axle load position monitoring method - Google Patents
Vehicle axle load position monitoring method Download PDFInfo
- Publication number
- CN115661749A CN115661749A CN202211342233.0A CN202211342233A CN115661749A CN 115661749 A CN115661749 A CN 115661749A CN 202211342233 A CN202211342233 A CN 202211342233A CN 115661749 A CN115661749 A CN 115661749A
- Authority
- CN
- China
- Prior art keywords
- vehicle
- axle load
- point
- monitoring
- coordinate system
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to a method for monitoring the axle load position of a vehicle, which comprises the following steps: detecting vehicles in the real-time monitoring picture based on the target detection model; extracting geometric characteristics of the vehicle, establishing an indirect inference method of axle load positions, and establishing a vehicle axle load key point data set and a key point detection model by taking the indirect inference method as a basis; training a key point detection model by using a vehicle axle load key point data set; based on the detection result of the key point detection model, the axle load coordinate in the image coordinate system is transformed into the world coordinate system through the homography matrix to obtain the axle load coordinate in the world coordinate system, the motion trail of the vehicle axle load is optimally estimated by using Kalman filtering and a kinematic bicycle model of the vehicle, and the real-time spatial distribution monitoring of the vehicle axle load is realized. Compared with the prior art, the method has the advantages that the positions of all axle load action points can be obtained according to the monocular image, the axle load real-time spatial distribution monitoring precision is high, and the like.
Description
Technical Field
The invention relates to the field of vehicle load monitoring, in particular to a method for monitoring a vehicle axle load position.
Background
For bridges in the operational phase, the most important load is the vehicle load. The vehicle load is different from other types of loads, has strong uncertainty and randomness, and therefore has a large influence on the safety and durability of the bridge operation period. With the continuous development of the logistics transportation industry, the phenomenon of vehicle overload on the bridge is more and more common, which accelerates the degradation of the performance of the bridge and even causes serious damage (such as the overturn of the main beam). Therefore, it is necessary to establish a vehicle load monitoring system for detecting the temporal and spatial distribution of the load of the vehicles passing through the bridge. Among them, the detection method based on computer vision is more and more widely applied because of its advantages of low cost, strong real-time performance, easy maintenance, etc.
Through the literature search of the prior art, research in vehicle load detection based on computer vision mostly focuses on detection of a vehicle so as to achieve the purpose of accurately detecting the real-time position of the vehicle load. CN112444311A discloses a bridge vehicle space-time load monitoring method, which is implemented by constructing a bridge deck vehicle panorama according to a vision system, obtaining a vehicle detection frame by using a vehicle detection model, and acquiring or updating vehicle track information through a preset improved Kalman filtering model to realize monitoring of vehicle load space-time distribution. However, the detection of the vehicle load is rough, the vehicle load is generally regarded as a concentrated load, and the real-time spatial distribution of the vehicle axle load cannot be effectively detected.
Disclosure of Invention
The invention aims to provide a method for monitoring the position of a vehicle axle load, which can effectively monitor the real-time spatial distribution of the vehicle axle load.
The purpose of the invention can be realized by the following technical scheme:
a vehicle axle load position monitoring method comprises the following steps:
detecting vehicles in the real-time monitoring picture based on the target detection model;
extracting geometric characteristics of the vehicle, establishing an indirect inference method of axle load positions, and establishing a vehicle axle load key point data set and a key point detection model by taking the indirect inference method as a basis;
training a key point detection model by using a vehicle axle load key point data set;
based on the detection result of the key point detection model, the axle load coordinate in the image coordinate system is transformed into the world coordinate system through the homography matrix to obtain the axle load coordinate in the world coordinate system, the motion trail of the vehicle axle load is optimally estimated by using Kalman filtering and a kinematic bicycle model of the vehicle, and the real-time spatial distribution monitoring of the vehicle axle load is realized.
According to the indirect estimation method of the axle load position, the positions of all axle load action points are obtained through monocular images by calculating the axle load position shielded by a vehicle body according to the detected axle load position and the window position on one side of the vehicle and the hypothesis that the distance between tires on two sides of the vehicle is equal to the width of the vehicle and the hypothesis that the contact points of the tires of the vehicle and the ground are surrounded into a rectangle.
The method comprises the steps that a key point detection model is built on the basis of HourglassNet, the HourglassNet conducts primary feature extraction on an image through a plurality of convolution layers, and then secondary feature extraction is conducted on features in the image through a plurality of encoding-decoding structures, wherein one encoding-decoding structure is a Hourglass module, thermodynamic diagrams are generated by stacking a plurality of Hourglass modules, so that key points in the image are detected, and the output of the previous Hourglass module is directly used as the input of the next Hourglass module.
The basic structure of the Hourglass module comprises two branches, wherein one branch extracts deep features of an input array through multiple pooling operations, the resolution is gradually restored to the original size through upsampling, and the other branch only extracts the features without changing the size of the array; the Hourglass module adds arrays of two branches when outputting, so that the neural network can combine information of multiple scales in the image.
And training the key point detection model in an intermediate supervision mode, wherein a loss function of the model is set as a sigmoid cross entropy between the feature graph output by each Hourglass module and the labeling data.
Assuming a homography matrix
The coordinates of point i in the image coordinate system are
The coordinates in the real coordinate system are
n is the number of reference points, then:
and obtaining a least square solution of H by solving the equation so as to determine the conversion relation between the image coordinate system and the world coordinate system as follows:
the kinematic bicycle model is used for tracking and correcting the track of the axle load action point, and the kinematic bicycle model adopts four variables to describe the current state of the bicycle: coordinates (x, y) and speed (v) of the center point, and attitude (θ) of the vehicle body; variables describing the real-time position of the axle load point of action of a vehicle include: length (h) and width (w) of the rectangle, coordinates (x, y) of the center of the rectangle, and attitude (θ) of the vehicle body;
according to the principle of rigid motion, the kinematic equation of the bicycle model is as follows:
wherein, the first and the second end of the pipe are connected with each other,
is the included angle between the speed direction of the central point and the x axis;
the projection of the velocity in the direction of the coordinate axis is defined as v x And v y Then there is
The kinematic equation for the bicycle model is equivalent to:
according to an equivalent kinematics equation of a bicycle model, five parameters (h, w, x, y, theta) are needed for determining the real-time state of each axle load, and since the size parameters (h, w) of a rectangle cannot change along with time, the calculation of the posture parameters theta of the bicycle body depends on the value of h, the optimal estimation of the motion trail of the axle load of the bicycle specifically comprises the following steps:
kalman filtering based determination of central point coordinates and optimal estimation value (x) of corresponding movement speed opt ,y opt ,v x,opt ,v y,opt ) And realize the tracking of the vehicle;
calculating an optimum estimated value (h) of the rectangular size based on the result of the tracking detection opt ,w opt )；
Calculating an optimal estimate of body attitude θ from an equivalent kinematics equation for a bicycle model opt 。
System state S at time t t As the position and velocity of the centre point M, i.e. S t ＝(x t ,y t ,v x,t ,v y,t ) T The measurement state being the position of the centre point M, i.e. M t ＝(x t ,y t ) T ；
According to a system state equation in Kalman filtering, a state transition and measurement equation of a vehicle axle load rectangular central point is as follows:
wherein u is t Is process noise.
According to the principle of Kalman filtering, the coordinates of the central point M and the optimal estimation value of the corresponding movement speed are obtained through three steps of initialization, prediction and updating, and the tracking of the vehicle is realized:
a) Initialization:
when the central point M is first detected, the system state S is initialized, and S is determined 1 Wherein, the position is initialized to a measured value, and the speed is initialized according to the designed speed of the urban road;
b) A prediction step:
wherein Q is u t Covariance matrix of (2), P t A covariance matrix that is a predicted value;
according to the predicting step, the predicted value of the coordinate of the M point of the current frame is calculated by utilizing the state information of the central point M of each vehicle obtained from the previous frame, the distance between the predicted value and the measured value of the coordinate of each M point is calculated, and if the distance is smaller than a pre-configured threshold value, the vehicles are matched into the same vehicle;
c) And (3) updating:
wherein R is the variance of the observation noise;
and according to the updating step, after the prediction of each prediction step is completed, updating the vehicle state so as to realize the tracking of the vehicle.
Compared with the prior art, the invention has the following beneficial effects:
(1) The invention carries out reasonable abstraction on the geometric shape of the vehicle, makes the axle load space-time distribution detection based on monocular vision possible, and can obtain key point information by only adopting one camera for shooting, thus having more practicability.
(2) The method and the device perform optimal estimation on the motion trail of the vehicle axle load by using Kalman filtering and a kinematic bicycle model of the vehicle, realize the monitoring of the real-time spatial distribution of the vehicle axle load, have high monitoring precision and have reference monitoring results.
Drawings
FIG. 1 is a flow chart of a method of the present invention;
FIG. 2 is a diagram showing the results of the detection of the target detection model;
FIG. 3 is a schematic diagram of an indirect method of inferring axle load position;
FIG. 4 is a schematic diagram of the normalization and labeling of a key point data set;
FIG. 5 is a block diagram of the Hourglass module;
FIG. 6 is a block diagram of HourgalaNet;
FIG. 7 is a test result of HourgalaNet;
FIG. 8 is a schematic representation of a kinematic bicycle model;
FIG. 9 is a schematic flow chart of the optimal estimation of the motion trajectory of the vehicle axle load using Kalman filtering and the kinematic bicycle model of the vehicle;
FIG. 10 is a detection zone in an embodiment;
FIG. 11 is a diagram illustrating the trajectory estimation of the center point of an axle-borne rectangle in one embodiment;
FIG. 12 shows the results of the spatial distribution of the vehicle axle loads in one embodiment;
FIG. 13 is a schematic diagram of a trace of the axle load matrix and its center point in the monitored area according to an embodiment.
Detailed Description
The invention is described in detail below with reference to the figures and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the scope of the present invention is not limited to the following embodiments.
The embodiment provides a method for monitoring a vehicle axle load position, as shown in fig. 1, comprising the following steps:
1) And detecting the vehicle in the real-time monitoring picture based on the target detection model.
In this embodiment, a YOLO-v5 network is adopted, and a target detection model of a vehicle is obtained by training with a public data set detac, as shown in fig. 2.
2) Extracting the geometric characteristics of the vehicle, establishing an indirect inference method of the axle load position, and establishing a vehicle axle load key point data set and a key point detection model by taking the indirect inference method as a basis.
21 Method for indirect inference of vehicle geometry and establishing an on-axle position and a set of key point data
Due to the shielding of the vehicle body, the position of the tire on one side of the vehicle can only be directly obtained in a single picture, and the position of the tire on the other side cannot be directly obtained from the image. The shape of the vehicle is highly standardized, however, the distance between the tires on either side of the vehicle can be considered to be equal to the width of the vehicle, and the tires of the vehicle enclose a rectangle with the ground contact point. Therefore, only the axle load acting point position and the vehicle width on one side of the vehicle need to be obtained, all the wheel load acting point positions of the vehicle can be obtained through calculation, and the width of the vehicle can be considered to be equal to the distance between the rearview mirrors of the vehicle. In other words, the indirect estimation method of the axle load position calculates the axle load position shielded by the vehicle body according to the rectangular characteristic on the basis of the assumption that the distance between the tires on the two sides of the vehicle is equal to the vehicle width and the assumption that the contact point between the vehicle tire and the ground is enclosed into a rectangle according to the detected axle load position on one side of the vehicle and the window position, and realizes the position of all the axle load action points through the monocular image.
As shown in fig. 3, the contact points of the vehicle and the ground are a, B, C, D, the quadrangle ABCD is rectangular, and the line segment BC is parallel to and equal to the line segment W1W 2. The points at which the coordinates can be directly obtained from the graph are A (x) A ,y A )，B:(x B ,y B )，W1:(x W1 ,y W1 )，W2:(x W2 ,y W2 ) From the above geometric relationship, the coordinates of C and D can be obtained: c (x) A +x W2 -x W1 ,y A +y W2 -y W1 )，D:(x B +x W2 -x W1 ,y B +y W2 -y W1 ). Therefore, the positions of all the axle load acting points can be obtained through the monocular image only by detecting the axle load position and the window position on the vehicle side.
In the embodiment, videos in the public data set detac data set are used, and then the YOLO-v5 model obtained by training in step 1) is used for detection, vehicle images detected in the vehicle videos are cut out, the positions of points a, B, W1 and W2 in the images are labeled, and 1638 key points of different vehicles are labeled. Since the size of the picture in the original data set is different from each other, the image needs to be normalized, and assuming that the size of the original image is (m, n), the image is first scaled to an image with a width of 256, i.e. the size becomes (m/n × 256,256), and then gray is filled under the picture, so that the image is finally converted into a square with a size of 256 × 256, and at the same time, the annotation data is correspondingly transformed, as shown in fig. 4.
By labeling, a set of 1638 size-normalized images and their corresponding keypoint data was obtained, where 60% (983) was used for training the keypoint detection model, 20% (328) for model verification, and 20% (327) for testing the keypoint detection model.
22 ) establishing a keypoint detection model
The method comprises the steps that a key point detection model is built on the basis of HourglassNet, the HourglassNet conducts primary feature extraction on an image through a plurality of convolution layers, and then secondary feature extraction is conducted on features in the image through a plurality of encoding-decoding structures, wherein one encoding-decoding structure is a Hourglass module, thermodynamic diagrams are generated by stacking a plurality of Hourglass modules, so that key points in the image are detected, and the output of the previous Hourglass module is directly used as the input of the next Hourglass module.
The basic structure of the Hourglass module comprises two branches, wherein one branch extracts deep features of an input array through multiple pooling operations, the resolution is gradually restored to the original size through upsampling, and the other branch only extracts the features without changing the size of the array; the Hourglass module adds arrays of two branches when outputting, so that the neural network can combine information of multiple scales in the image.
The Hourglass module in this embodiment is formed by recursion of four basic structures, and 1/2,1/4,1/8,1/16 of input resolution of multi-scale information is fused, as shown in FIG. 5.
The structure of the hourglasset of this embodiment is shown in fig. 6, and is formed by combining one coding layer and four Hourglass modules.
3) And training the key point detection model by using the vehicle axle load key point data set.
Because the HourglassNet is deeper, in order to prevent the phenomenon that the gradient disappears, an intermediate supervision mode is adopted for training, namely the loss function of the model is the sigmoid cross entropy between the feature diagram output by each Hourglass module and the labeling data. In addition, in this embodiment, each Hourglass module in the HourglassNet outputs a feature map of 64 × 64, so the loss function is:
where i, j is the number of rows and columns of the feature map, g i,j Is the data of the annotation,
is the result of the output of the kth Hourglass module.
Training the keypoint detection model using the keypoint data set generated in step 21), and testing the test set, as shown in fig. 7. As can be seen from fig. 7, the HourglassNet keypoint detection model can identify the positions of the vehicle keypoints more accurately.
4) Based on the detection result of the key point detection model, the axle load coordinate in the image coordinate system is transformed into the world coordinate system through the homography matrix to obtain the axle load coordinate in the world coordinate system, the motion trail of the vehicle axle load is optimally estimated by using Kalman filtering and a kinematic bicycle model of the vehicle, and the real-time spatial distribution monitoring of the vehicle axle load is realized.
41 Coordinate transformation)
Since the directly detected keypoint coordinates are in the image coordinate system, a transformation of the coordinate system is required, which needs to be determined by on-site calibration. According to the principle of homography transformation, as long as a plurality of non-collinear points (the number is more than or equal to 4) are marked on the spot, a homography matrix can be obtained through calculation, and therefore the coordinates in the image coordinate system are converted into the coordinates in the real coordinate system.
Assuming a homography matrix
The coordinate of the point i in the image coordinate system is
The coordinates in the real coordinate system are
n is the number of reference points, then:
and obtaining a least square solution of H by solving the equation so as to determine the conversion relation between the image coordinate system and the world coordinate system as follows:
42 To build a kinematic bicycle model
The motion of a vehicle on a road surface belongs to the motion of a rigid body on a two-dimensional plane. The present embodiment describes the motion of the vehicle using a kinematic bicycle model, and tracks and corrects the trajectory of the on-axis point of action based on the kinematic bicycle model.
The kinematic bicycle model makes the following three assumptions about the motion of the car:
1. only the movement of the car in the two-dimensional plane is considered, not the movement in the vertical direction;
2. the speed and direction of two front wheels of the automobile are equal, and the speed and direction of the rear wheels are also equal, so that the front wheels and the rear wheels can be described by one wheel respectively;
3. similar to a bicycle, the turning angle of the body of the car is also controlled by the front wheel.
The basic assumption of the kinematic bicycle model is shown in fig. 8 (a).
As shown in FIG. 8 (b), the kinematic bicycle model is used to track and correct the trajectory of the axle load application point, and it uses four variables to describe the current state of the vehicle: coordinates (x, y) and speed (v) of the center point, and attitude (θ) of the vehicle body; and variables describing the real-time location of the axle load point of action of a vehicle include: the length (h) and width (w) of the rectangle, the coordinates (x, y) of the center of the rectangle, and the attitude (θ) of the vehicle body are shown in fig. 8 (c).
According to the principle of rigid motion, the kinematic equation of the bicycle model is as follows:
wherein, the first and the second end of the pipe are connected with each other,
is the included angle between the speed direction of the central point and the x axis;
since the velocity is a vector and is not beneficial to subsequent calculation, the projection of the velocity in the coordinate axis direction is defined as v x And v y Then there is
The kinematic equation for the bicycle model is equivalent to:
43 Optimal estimation of the motion trajectory of the vehicle axle load by combining Kalman filtering and the kinematic bicycle model of the vehicle
Although the key point detection model can detect the key points of the vehicle in the image so as to deduce the real-time position of the axle load action point, the method still has measurement errors and cannot track and detect the vehicle, so that the detection result is not suitable for directly monitoring the axle load space-time distribution. In order to eliminate the measurement error and track the vehicle, the invention further processes the detection result by adopting Kalman filtering.
According to an equivalent kinematics equation of a bicycle model, five parameters (h, w, x, y, theta) are needed for determining the real-time state of each axle load, and since the size parameters (h, w) of a rectangle do not change along with time and the calculation of the posture parameter theta of the vehicle body depends on the value of h, the flow for optimally estimating the motion track of the axle load of the vehicle is shown in fig. 9, and the method specifically comprises the following steps:
431 Determining optimal estimation value (x) of center point coordinates and corresponding movement speed based on Kalman filtering opt ,y opt ,v x,opt ,v y,opt ) And realize the tracking of the vehicle;
in the embodiment, the motion state of the central point (i.e., point M in fig. 8 (b)) of the airborne rectangle is tracked and optimized, and the key point detection model can only detect the real-time state of the vehicle in a single frame image and cannot detect the current motion speed, so that the system state S at the time t t As the position and velocity of the centre point M, i.e. S t ＝(x t ,y t ,v x,t ,v y,t ) T The measurement state being the position of the centre point M, i.e. M t ＝(x t ,y t ) T ；
According to a system state equation in Kalman filtering, the state transition and measurement equation of the vehicle axle load rectangular central point is as follows:
wherein u is t Is process noise.
According to the principle of Kalman filtering, the coordinates of the central point M and the optimal estimation value of the corresponding movement speed are obtained through three steps of initialization, prediction and updating, and the tracking of the vehicle is realized:
a) Initialization:
when the center point M is first detected, initializing the system state S, and determining S 1 Wherein, the position is initialized to a measured value, and the speed is initialized according to the design speed of the urban road;
b) A prediction step:
wherein Q is u t Covariance matrix of (2), P t A covariance matrix that is a predicted value;
according to the predicting step, calculating a predicted value of the coordinate of the M point of the current frame by using the state information of the central point M of each vehicle obtained from the previous frame, calculating the distance between the predicted value and the measured value of the coordinate of each M point, and matching the predicted value and the measured value into the same vehicle if the distance is smaller than a pre-configured threshold value;
c) And (3) updating:
wherein R is the variance of the observation noise;
and according to the updating step, after the prediction of each prediction step is completed, updating the vehicle state so as to realize the tracking of the vehicle.
432 Calculating an optimal estimated value (h) of the rectangular size based on the result of the tracking detection opt ,w opt )；
The dimensions (h, w) of the rectangle formed by the axle-loaded contact points for each vehicle remain constant during vehicle movement. 431 Using Kalman filtering to realize vehicle tracking and M-point track optimization, counting the detection results of the same vehicle in different frames by using the tracking result, and determining the maximum likelihood estimation values (h) of h and w opt ,w opt )。
Assuming that the vehicle k passes through the region to be detected in the m-n frames of the video, n-m +1 detection results of the vehicle can be obtained according to the detection result of each frame. And counting the data to respectively obtain a frequency distribution table of h and a frequency distribution table of w, and taking the middle point of the interval with the highest frequency as the maximum likelihood estimation for calculating the axle load real-time position.
433 Computing an optimal estimate of body attitude θ from an equivalent kinematics equation for the bicycle model opt 。
According to the equivalent kinematic equation of the bicycle model:
(h, w, x, y, v) of the vehicle at each instant x ,v y ) Has been determined, and theta is t+1 Can be prepared from (h, v) x ,v y ) And theta t Determine, therefore, only for θ 1 The value of each time θ can be calculated by initialization. In conclusion, the axle load spatial distribution of the vehicle at each moment has been completely determined.
This example provides a specific test case to further illustrate the feasibility and effectiveness of the method of the present invention.
The field test selects to obtain the traffic monitoring video in a period of time on a road at a certain place through shooting. Through marking and calculation, the homography matrix from the image coordinate system to the road surface coordinate system is obtained as follows:
therefore, in order to ensure the accuracy, the present embodiment selects a region of 22.50m × 55.00m as the detection region of the axial load space-time distribution, as shown in fig. 10.
A 1000 frame (16 second) video segment was chosen as an example case. In this video, a total of 8 vehicles pass through the detection area, which is respectively marked as vehicle 1, vehicle 2, \ 8230;, vehicle 8. Based on the method for monitoring the vehicle axle load position, the optimal estimation of the axle load rectangular center track is obtained as shown in fig. 11. As can be seen from the figure, the track of the center of the axle load rectangle optimized by Kalman filtering is smoother and more suitable for the actual situation.
After the track of the central point of the axle load rectangle is obtained, the size h multiplied by w of the axle load rectangle and the motion attitude angle theta of the axle load rectangle at each moment can be obtained by utilizing a maximum likelihood estimation method and a kinematic bicycle model of a vehicle. According to the optimized values of the parameters (x, y, h, w, theta), the optimal estimation (x, y, h, w, theta) of the state of each frame can be obtained. Fig. 12 shows the detection result of the vehicle axle load spatial distribution at a part of the time. As can be seen from the figure, the position of each vehicle within the monitored area on the axle load can be accurately calculated.
Fig. 13 shows the axle load rectangles of the vehicles 1 to 8 and the trajectories of their center points in the monitoring area. As can be seen from the figure, the track of the axle load action points is smooth, and the axle load action points all move along the lane, which is consistent with the practical situation, so that the method provided by the invention can accurately sense the space-time distribution of the vehicle.
The foregoing detailed description of the preferred embodiments of the invention has been presented. It should be understood that numerous modifications and variations could be devised by those skilled in the art in light of the present teachings without departing from the inventive concepts. Therefore, the technical solutions that can be obtained by a person skilled in the art through logical analysis, inference or limited experiments based on the prior art according to the concepts of the present invention should be within the scope of protection defined by the claims.
Claims (10)
1. A method for monitoring the axle load position of a vehicle is characterized by comprising the following steps:
detecting vehicles in the real-time monitoring picture based on the target detection model;
extracting geometric characteristics of the vehicle, establishing an indirect inference method of axle load positions, and establishing a vehicle axle load key point data set and a key point detection model by taking the indirect inference method as a basis;
training a key point detection model by using a vehicle axle load key point data set;
based on the detection result of the key point detection model, the axle load coordinate in the image coordinate system is transformed into the world coordinate system through the homography matrix to obtain the axle load coordinate in the world coordinate system, the motion trail of the vehicle axle load is optimally estimated by using Kalman filtering and a kinematic bicycle model of the vehicle, and the real-time spatial distribution monitoring of the vehicle axle load is realized.
2. The method for monitoring the axle load position of the vehicle according to claim 1, wherein the indirect estimation method of the axle load position is based on the assumption that the distance between the tires at two sides of the vehicle is equal to the vehicle width and the assumption that the contact points between the tires of the vehicle and the ground enclose a rectangle according to the detected axle load position and the window position at one side of the vehicle, and the axle load position shielded by the vehicle body is calculated according to the rectangle characteristic, so that the positions of all the axle load acting points are obtained through monocular images.
3. The method for monitoring the vehicle axle load position according to claim 1, wherein the keypoint detection model is established based on HourglassNet, hourglassNet performs primary feature extraction on an image through a plurality of convolution layers, and performs secondary feature extraction on features in the image through a plurality of coding-decoding structures, wherein one coding-decoding structure is a Hourglass module, a thermodynamic diagram is generated by stacking a plurality of Hourglass modules, so as to detect keypoints in the image, and the output of the previous Hourglass module is directly used as the input of the next Hourglass module.
4. The vehicle axle load position monitoring method according to claim 3, wherein the basic structure of the Hourglass module comprises two branches, one branch extracts deep level features of an input array through multiple pooling operations and gradually restores the resolution to the original size through upsampling, and the other branch only extracts the features without changing the size of the array; the Hourglass module adds the array of two branches when outputting, so that the neural network can combine information of multiple scales in the image.
5. The method as claimed in claim 3, wherein the key point detection model is trained in an intermediate supervision manner, and the loss function of the model is set as the cross entropy of sigmoid between the feature map output by each Hourglass module and the labeled data.
6. A method for monitoring the position of the axle load of a vehicle according to claim 1, characterized in that a homography matrix is assumed
The coordinates of point i in the image coordinate system are
The coordinates in the real coordinate system are
i =1,2,3, \8230, n, n being the number of reference points, then:
and obtaining a least square solution of H by solving the equation so as to determine the conversion relation between the image coordinate system and the world coordinate system as follows:
7. the method of claim 1, wherein the kinematic bicycle model is used for tracking and correcting the locus of the axle load point, and the kinematic bicycle model uses four variables to describe the current state of the vehicle: coordinates (x, y) and speed (v) of the center point, and attitude (θ) of the vehicle body; variables describing the real-time location of the axle load point of action of a vehicle include: the length (h) and width (w) of the rectangle, the coordinates (x, y) of the center of the rectangle, and the attitude angle (theta) of the vehicle body;
according to the principle of rigid body motion, the kinematic equation of the bicycle model is as follows:
wherein the content of the first and second substances,
is the included angle between the speed direction of the central point and the x axis;
the projection of the velocity in the direction of the coordinate axis is defined as v x And v y Then there is
The kinematic equation for the bicycle model is equivalent to:
8. the method for monitoring the vehicle axle load position according to claim 7, wherein five parameters (h, w, x, y, θ) are required for determining the real-time state of each vehicle axle load according to the equivalent kinematic equation of the bicycle model, and since the dimension parameters (h, w) of the rectangle do not change with time and the calculation of the vehicle body attitude parameter θ depends on the value of h, the optimal estimation of the motion trajectory of the vehicle axle load specifically comprises the following steps:
kalman filtering based determination of central point coordinates and optimal estimation value (x) of corresponding movement speed opt ,y opt ,v x,opt ,v y,opt ) And realize the tracking of the vehicle;
calculating an optimal estimated value (h) of the rectangular size based on the result of the tracking detection opt ,w opt )；
Calculating an optimal estimate of body attitude θ from an equivalent kinematics equation for a bicycle model opt 。
9. The method according to claim 8, wherein the system state S at time t is t As the position and velocity of the centre point M, i.e. S t ＝(x t ,y t ,v x,t ,v y,t ) T The measurement state being the position of the centre point M, i.e. M t ＝(x t ,y t ) T ；
According to a system state equation in Kalman filtering, a state transition and measurement equation of a vehicle axle load rectangular central point is as follows:
wherein u is t Is process noise.
10. The method for monitoring the axle load position of the vehicle according to claim 9, wherein the tracking of the vehicle is realized by obtaining the coordinates of the central point M and the optimal estimated value of the corresponding movement speed thereof through three steps of initialization, prediction and update according to the kalman filtering principle:
a) Initialization:
when the central point M is first detected, the system state S is initialized, and S is determined 1 Wherein, the position is initialized to a measured value, and the speed is initialized according to the design speed of the urban road;
b) A prediction step:
wherein Q is u t Of the covariance matrix, P t A covariance matrix that is a predicted value;
according to the predicting step, the predicted value of the coordinate of the M point of the current frame is calculated by utilizing the state information of the central point M of each vehicle obtained from the previous frame, the distance between the predicted value and the measured value of the coordinate of each M point is calculated, and if the distance is smaller than a pre-configured threshold value, the vehicles are matched into the same vehicle;
c) And (3) updating:
wherein R is the variance of the observation noise;
and according to the updating step, after the prediction of each prediction step is completed, updating the vehicle state so as to realize the tracking of the vehicle.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211342233.0A CN115661749A (en) | 2022-10-30 | 2022-10-30 | Vehicle axle load position monitoring method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211342233.0A CN115661749A (en) | 2022-10-30 | 2022-10-30 | Vehicle axle load position monitoring method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115661749A true CN115661749A (en) | 2023-01-31 |
Family
ID=84994106
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211342233.0A Pending CN115661749A (en) | 2022-10-30 | 2022-10-30 | Vehicle axle load position monitoring method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115661749A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117874630A (en) * | 2024-03-13 | 2024-04-12 | 同济大学 | Optical fiber traffic on-axis signal processing method and device, electronic equipment and medium |
-
2022
- 2022-10-30 CN CN202211342233.0A patent/CN115661749A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117874630A (en) * | 2024-03-13 | 2024-04-12 | 同济大学 | Optical fiber traffic on-axis signal processing method and device, electronic equipment and medium |
| CN117874630B (en) * | 2024-03-13 | 2024-05-03 | 同济大学 | Optical fiber traffic on-axis signal processing method and device, electronic equipment and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111932580A (en) | Road 3D vehicle tracking method and system based on Kalman filtering and Hungary algorithm | |
| Tanzmeister et al. | Grid-based mapping and tracking in dynamic environments using a uniform evidential environment representation | |
| Song et al. | A model-based vehicle segmentation method for tracking | |
| CN109727273B (en) | Moving target detection method based on vehicle-mounted fisheye camera | |
| CN113126115B (en) | Semantic SLAM method and device based on point cloud, electronic equipment and storage medium | |
| Huang et al. | Lane detection based on inverse perspective transformation and Kalman filter | |
| Kim | Realtime lane tracking of curved local road | |
| JP2020052695A (en) | Object detection apparatus, object detection method, and computer program for object detection | |
| Tang et al. | Multiple-kernel based vehicle tracking using 3D deformable model and camera self-calibration | |
| Hbaieb et al. | Pedestrian detection for autonomous driving within cooperative communication system | |
| Pinggera et al. | High-performance long range obstacle detection using stereo vision | |
| CN111027461A (en) | Vehicle track prediction method based on multi-dimensional single-step LSTM network | |
| CN114419098A (en) | Moving target trajectory prediction method and device based on visual transformation | |
| CN115063454B (en) | Multi-target tracking matching method, device, terminal and storage medium | |
| Arsenali et al. | RotInvMTL: Rotation invariant multinet on fisheye images for autonomous driving applications | |
| CN115661749A (en) | Vehicle axle load position monitoring method | |
| Fan et al. | Dynamicfilter: an online dynamic objects removal framework for highly dynamic environments | |
| Zeng | An object-tracking algorithm for 3-D range data using motion and surface estimation | |
| Kang et al. | Robust visual tracking framework in the presence of blurring by arbitrating appearance-and feature-based detection | |
| Gökçe et al. | Recognition of dynamic objects from UGVs using Interconnected Neuralnetwork-based Computer Vision system | |
| Muffert et al. | Stix-fusion: A probabilistic stixel integration technique | |
| CN115471526A (en) | Automatic driving target detection and tracking method based on multi-source heterogeneous information fusion | |
| Wang et al. | Online drone-based moving target detection system in dense-obstructer environment | |
| Kang et al. | Real-time road surface marking detection from a bird’s-eye view image using convolutional neural networks | |
| Yildiz et al. | CNN based sensor fusion method for real-time autonomous robotics systems |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |