CN111489585A - Vehicle and pedestrian collision avoidance method based on edge calculation - Google Patents

Abstract

The invention discloses a vehicle-pedestrian collision avoidance method based on edge computing, which belongs to the field of Internet of Things. In the edge computing environment, communication between vehicles and pedestrian equipment is realized through an edge server, communication delay is reduced, and Kalman filtering is used to predict vehicles The position of the pedestrian device is used to predict the pedestrian trajectory, which overcomes the limitation of the line of sight and improves the collision detection accuracy. , a collision risk model based on time to collision (TTC) is established, and the simulation results show that the collision avoidance algorithm proposed in the present invention can effectively reduce the collision risk of vehicles and pedestrians.

Description

A vehicle-pedestrian collision avoidance method based on edge computing

technical field

The invention relates to the technical field of the Internet of Things, in particular to a vehicle-pedestrian collision avoidance method based on edge computing.

Background technique

In 2017, Hassan Artail proposed a scheme to use cellular technology to realize pedestrian and vehicle communication in Avoiding car-pedestrian collisions using a VANET to cellular communication framework to detect possible vehicle-pedestrian collisions. In this technical solution, a remote cloud server is used as an intermediate server to indirectly realize the communication between vehicles and pedestrians, which overcomes the limitation of sight distance, but the system delay is very high and difficult to determine, which is extremely sensitive to the delay of pedestrian safety algorithm programs. It is undoubtedly a fatal flaw; in addition, the pedestrian and vehicle position prediction algorithms in this technical solution are extremely rough. In most cases, the positions where pedestrians and vehicles may appear at the same time are all classified as unsafe areas, and the probability of false alarms is relatively high. high.

SUMMARY OF THE INVENTION

The technical purpose of the present invention is to provide a vehicle-pedestrian collision avoidance method based on edge computing, which not only reduces the communication delay between vehicles and pedestrians, but also improves the early warning accuracy of vehicle-pedestrian collision accidents.

The above-mentioned technical purpose of the present invention is achieved through the following technical solutions:

A vehicle-pedestrian collision avoidance method based on edge computing, comprising the following steps:

Step1: Establish a network model for pedestrian device-to-vehicle communication: Arrange roadside units (RSUs)/base stations at intersections or road edges, and mobile edge computing (MEC) servers connected to the roadside units (RSUs)/base stations, so The pedestrian device is a smart phone with a GPS module, and the pedestrian device and the vehicle can communicate within the coverage of the network model;

In this mode, the MEC server can be connected to the roadside unit (RSU)/base station, the vehicle communicates with the RSU/base station through LTE-V2X, and the pedestrian communicates with the RSU/base station through a smartphone; computing tasks can be performed in the vehicle Or on the MEC server; such a system not only utilizes the resources of the vehicle in the traditional Internet of Vehicles, but also relieves the pressure of the computing resources on the vehicle by running the computing task on the MEC server when the computing task is heavy, and also utilizes the smart phone. resources such as location services on the Internet, allowing smartphones to participate in the application of pedestrian safety protection;

Step2: According to the current information and historical information of vehicles and pedestrian equipment, use Kalman filter to predict the future positions of vehicles and pedestrians respectively;

Step3: According to the predicted vehicle and pedestrian paths in Step2, according to the mutual motion relationship, run the vehicle-pedestrian collision avoidance algorithm to give early warning of possible collisions.

Further, the specific steps of predicting the future position of the vehicle in the Step2 are:

Assuming that the car has a constant rotation rate and speed, the state vector of the selected vehicle is:

X=[xyv θ ω] ^T

θ is the yaw angle, which is the angle between the tracked target vehicle and the x-axis in the current vehicle coordinate system, the counterclockwise direction is positive, and the value range is [0, 2π), and ω is the yaw angular velocity; The motion states of roads and traffic regulations are generally straight or turning. Both motion states can be represented by this state vector. When ω≠0, it is non-straight travel. At this time, the state transfer function of the model is:

When ω=0, it is a straight line, and the state transition function of the model becomes:

For the multivariate functions of (1) and (2), the multivariate Taylor series is used for expansion, the higher order number is ignored, and only the Jacobian matrix is used for linearization. In this model, the partial derivative of each element can be obtained. Jacobian Matrix J _A , when ω≠0:

When ω=0, the Jacobian matrix is:

Calculate and deal with the uncertainty Q caused by noise; in the network model, the introduction of noise mainly comes from two sources: linear acceleration and yaw angular acceleration noise, the linear acceleration and yaw angular acceleration noise will affect the state quantity (x, y, v, θ, ω), the expression of the influence of the linear acceleration and yaw angular acceleration on the state is as follows:

为直线上和转角上的加速度，Q就是处理噪声的协方差矩阵，其表达式为：where μ _a ,

is the acceleration on the straight line and on the corner, Q is the covariance matrix for dealing with noise, and its expression is:

Q=E[n·n ^T ]=E[GμμT G ^T ]=G·E[μμ ^T ]·G ^{T (} 4)

in

Therefore, the calculation formula of the covariance matrix Q of processing noise in the CTRV model is:

When establishing the measurement equation, the data collected by the lidar and radar sensors are selected to perform sensor fusion to predict the vehicle trajectory;

The first is lidar, whose data is in the Cartesian coordinate system, which can detect the vehicle position, but has no speed information, and its measurement value is the coordinates (x, y) of the target vehicle; therefore, the measurement model of lidar is still linear. , and its measurement matrix is:

The corresponding measurement equation is:

The second is the radar, whose measured target vehicle data is in the polar coordinate system, and the radar's prediction is mapped to the measurement space is nonlinear ^[16] , and its expression is:

At this time, h(x) is used to represent such a nonlinear mapping, then when solving the Kalman gain, the nonlinear process should also be linearized using Taylor's formula, and the Jacobian matrix of h(x) should be solved with reference to the prediction process. :

The unknown parameters J _H , _JA , Q, R (processing noise, generally provided by the manufacturer) in the prediction and correction model have been solved, and then the data can be filtered according to the network model, so as to realize the future position of the vehicle Prediction.

Further, the specific steps of determining the future position of the pedestrian in the Step2 are:

The location of the pedestrian is determined according to the information collected by the GPS module of the smartphone, but there are still many errors in the dynamic positioning of the GPS module, such as satellite clock errors, etc. In order to eliminate the errors and improve the positioning accuracy, the Kalman filter method is applied to "Current" statistical model of GPS dynamic filtering;

Starting directly from the positioning results output by the GPS receiver, the total error caused by various error sources of GPS positioning in all directions is equivalent to the sum of a current mean and a colored noise that conforms to the first-order Markov process; according to " The current "statistical model", the pedestrian's state vector is selected as:

The state equation of the system can be defined as:

Among them, A is the system state transition matrix:

A=diag[A _x , A _y , A _z ]

A _y , A _z and so on;

W(t) is:

=[0 0 W _ax W _x 0 0 W _ay W _y 0 0 W _az W _z ] ^T

is the system noise vector;

的高斯白噪声；

是三个坐标轴上的“当前”加速度均值；τ _ax , τ _ay , τ _az are the acceleration-related time constants respectively; τ _x , τ _y , τ _z are the relevant time constants corresponding to the Markov process; W _ax , W _ay , and W _az are respectively

Gaussian white noise;

is the "current" mean acceleration on the three axes;

When establishing the measurement equation, take the positioning results L _x , _Ly , and L _z of the GPS receiver in the directions of the three coordinate axes as the measurement values, then:

L _x = x + ∈ _x + ω _lx

L _y =y+∈ _y +ω _ly

L _z = z+∈ _z +ω _lz

Written in matrix form:

L=[L _x , _Ly , L _z ] ^T

V is the observation noise V=[ω _lx , ω _ly , ω _lz ] ^T

H is the observation equation

Then the corresponding measurement equation is

L=HX+V (7)

So far, we have established the measurement equation of the system with the observation equation of the GPS system, and established the state transition equation of the system with the "current" statistical model and the error model of the GPS system. According to the above system equation and observation equation, let the sampling period be T. , through the discretization process ^[18] , the discrete Kalman filter equation is established as follows:

K(k+1)=P(k+1,k)H ^T (k+1)[H(k+1)P(k+1)H ^T (k+1)+R(k+1)] ^-1

(9)

P(k+1, k)=Φ(k+1, k)P(k)Φ ^T (k+1, k)+Q(k) (10)

P(k+1)=[I-K(k+1)H(k+1)]P(k+1,k) (11)

where Φ _l (k+1, k)=

diag[Φ _lx (k+1, k), Φ _ly (k+1, k), Φ _lz (k+1, k)]

Φ _ly , Φ _lz and so on;

In formula (10), Φ(k+1, k) is the discretization matrix of the system transition matrix A:

Φ(k+1, k)=diag[Φ _x (k+1, k), Φ _y (k+1, k), Φ _z (k+1, k)]

Φ _y (k+1, k), Φ _z (k+1, k) and so on, just replace τ _ax , τ _x in (12) with τ _ay , τ _y and τ _az , τ _z is Can;

Q(k) is the discretization matrix of the system noise covariance matrix Q

R(k) is the discretization matrix of the observation noise covariance matrix R of the system

Q(k) and R(k) are both positive definite matrices;

So far, the unknown parameters in the prediction and correction models have been solved, and then the data can be processed according to the conventional Kalman filter procedure.

Further, the specific steps of the vehicle-pedestrian collision avoidance algorithm in the Step 3 are:

Step3.1: Vehicles and pedestrian equipment periodically collect their own state information through their own sensors, and upload it to the MEC server through the RSU/base station, and run the corresponding Kalman filter on the MEC server according to the state information of pedestrians and vehicles Predict the trajectories of pedestrians and vehicles;

In the above Step3.1, the movement of pedestrians needs to be regarded as nonlinear, because the movement of pedestrians has great uncertainty, and EKF is a filter for predicting the position of nonlinear moving objects; compared with pedestrians, driving on the road Due to the restrictions of the road, the motion of the vehicle can be roughly divided into two categories, one is straight, and the yaw angular velocity is zero, which can be regarded as a linear motion; the other is when the yaw angular velocity is not zero when turning is nonlinear sports;

Step3.2: Run the collision avoidance algorithm to calculate the time to collision (TTC) and judge the possibility of vehicle-pedestrian collision according to the value of TTC to warn the vehicle driver and pedestrians;

The TTC (time to collision) is:

L ₁ , L ₂ are the trajectory lengths of vehicles and pedestrians to the collision point respectively, V ₁ , V ₂ are the current speeds of vehicles and pedestrians, respectively, t _v , t _p are the communication delays from vehicles and pedestrians to the edge server, respectively, according to this The value of TTC to determine whether a collision occurs;

When TTC=0, that is, if the current vehicle and pedestrian keep the current state without taking measures, a collision may occur. At this time, the warning information is sent to the vehicle, and the vehicle is notified to take measures before the accident occurs, so as to avoid the collision.

Further, in the Step 3.2, due to the unavoidable error in the algorithm, set TTCR (time to collision range) and threshold ξ, when

TTCR≤ξ

Collision warning is carried out to eliminate the error caused by trajectory prediction and the influence of vehicle length.

The beneficial effects of the present invention are:

First, the present invention establishes a feasible communication method between vehicles and pedestrians in the current network environment through the MEC server, effectively reducing the communication delay. The accuracy of pedestrian collision early warning effectively reduces the probability of collision between vehicles and pedestrians in non-line-of-sight situations such as blind spots on rainy and foggy roads.

Second, the present invention establishes a feasible communication mode between vehicles and pedestrians in the current network environment through the MEC server, and the introduction of the MEC greatly relieves the burden of the cloud server, improves the efficiency of the system, and increases the throughput of the system.

Third, the present invention introduces a pedestrian device-smartphone, which overcomes the limitation of non-line-of-sight distance, effectively expands the application scenarios, and adds the device that pedestrians carry with them that can provide location services anytime and anywhere into the Internet of Vehicles, which significantly improves the performance of the Internet of Vehicles. The safety of non-motorized road users, and the introduction of smartphones is bound to have a positive impact on the current vehicle-focused status of pedestrian safety research.

Fourth, the present invention also uses Kalman filtering to predict the future positions of vehicles and pedestrians, which further improves the accuracy of vehicle-pedestrian collision accident warning.

Description of drawings

1 is a system model diagram of the present invention;

Figure 2 is a graph showing the effect of vehicle speed on the success rate of collision warning;

Figure 3 is a graph of the impact data of pedestrian speed on the success rate of collision warning;

Figure 4 is a graph of the results of the comparative test data.

Detailed ways

In order to describe the present invention in more detail and facilitate the understanding of those skilled in the art, the present invention will be further described below with reference to the accompanying drawings and embodiments. Limit the invention.

Example:

A vehicle-pedestrian collision avoidance method based on edge computing, as shown in Figure 1, includes the following steps:

Step1: Establish a network model for pedestrian device-to-vehicle communication: Arrange roadside units (RSUs)/base stations at intersections or road edges, and mobile edge computing (MEC) servers connected to the roadside units (RSUs)/base stations, so The pedestrian device is a smart phone with a GPS module, and the pedestrian device and the vehicle can communicate within the coverage of the network model;

In this mode, the MEC server can be connected to the roadside unit (RSU)/base station, the vehicle communicates with the RSU/base station through LTE-V2X, and the pedestrian communicates with the RSU/base station through a smartphone; computing tasks can be performed in the vehicle Or on the MEC server; such a system not only utilizes the resources of the vehicle in the traditional Internet of Vehicles, but also relieves the pressure of the computing resources on the vehicle by running the computing task on the MEC server when the computing task is heavy, and also utilizes the smart phone. resources such as location services on the Internet, allowing smartphones to participate in the application of pedestrian safety protection;

Step2: According to the current information and historical information of vehicles and pedestrian equipment, use Kalman filter to predict the future positions of vehicles and pedestrians respectively;

Step3: According to the predicted vehicle and pedestrian paths in Step2, according to the mutual motion relationship, run the vehicle-pedestrian collision avoidance algorithm to give early warning of possible collisions.

Further, the specific steps of predicting the future position of the vehicle in the Step2 are:

Assuming that the car has a constant rotation rate and speed, the state vector of the selected vehicle is:

X=[xyv θ ω] ^T

θ is the yaw angle, which is the angle between the tracked target vehicle and the x-axis in the current vehicle coordinate system, the counterclockwise direction is positive, and the value range is [0, 2π), and ω is the yaw angular velocity; The motion states of roads and traffic regulations are generally straight or turning. Both motion states can be represented by this state vector. When ω≠0, it is non-straight travel. At this time, the state transfer function of the model is:

When ω=0, it is a straight line, and the state transition function of the model becomes:

For the multivariate functions of (1) and (2), the multivariate Taylor series is used for expansion, the higher order number is ignored, and only the Jacobian matrix is used for linearization. In this model, the partial derivative of each element can be obtained. Jacobian Matrix J _A , when ω≠0:

When ω=0, the Jacobian matrix is:

Calculate and deal with the uncertainty Q caused by noise; in the network model, the introduction of noise mainly comes from two sources: linear acceleration and yaw angular acceleration noise, the linear acceleration and yaw angular acceleration noise will affect the state quantity (x, y, v, θ, ω), the expression of the influence of the linear acceleration and yaw angular acceleration on the state is as follows:

为直线上和转角上的加速度，Q就是处理噪声的协方差矩阵，其表达式为：where μ _a ,

is the acceleration on the straight line and on the corner, Q is the covariance matrix for dealing with noise, and its expression is:

Q=E[n·n ^T ]=E[GμμT G ^T ]=G·E[μμ ^T ]·G ^{T (} 4)

in

Therefore, the calculation formula of the covariance matrix Q of processing noise in the CTRV model is:

When establishing the measurement equation, the data collected by the lidar and radar sensors are selected to perform sensor fusion to predict the vehicle trajectory;

The first is lidar, whose data is in the Cartesian coordinate system, which can detect the vehicle position, but has no speed information, and its measurement value is the coordinates (x, y) of the target vehicle; therefore, the measurement model of lidar is still linear. , and its measurement matrix is:

The corresponding measurement equation is:

The second is the radar, whose measured target vehicle data is in the polar coordinate system, and the radar's prediction is mapped to the measurement space is nonlinear ^[16] , and its expression is:

At this time, h(x) is used to represent such a nonlinear mapping, then when solving the Kalman gain, the nonlinear process should also be linearized using Taylor's formula, and the Jacobian matrix of h(x) should be solved with reference to the prediction process. :

The unknown parameters J _H , _JA , Q, R (processing noise, generally provided by the manufacturer) in the prediction and correction model have been solved, and then the data can be filtered according to the network model, so as to realize the future position of the vehicle Prediction.

Further, the specific steps of determining the future position of the pedestrian in the Step2 are:

The location of the pedestrian is determined according to the information collected by the GPS module of the smartphone, but there are still many errors in the dynamic positioning of the GPS module, such as satellite clock errors, etc. In order to eliminate the errors and improve the positioning accuracy, the Kalman filter method is applied to "Current" statistical model of GPS dynamic filtering;

Starting directly from the positioning results output by the GPS receiver, the total error caused by various error sources of GPS positioning in all directions is equivalent to the sum of a current mean and a colored noise that conforms to the first-order Markov process; according to " The current "statistical model", the pedestrian's state vector is selected as:

The state equation of the system can be defined as:

Among them, A is the system state transition matrix:

A=diag[A _x , A _y , A _z ]

A _y , A _z and so on;

W(t) is:

=[0 0 W _ax W _x 0 0 W _ay W _y 0 0 W _az W _z ] ^T

is the system noise vector;

的高斯白噪声；

是三个坐标轴上的“当前”加速度均值；τ _ax , τ _ay , τ _az are the acceleration-related time constants respectively; τ _x , τ _y , τ _z are the relevant time constants corresponding to the Markov process; W _ax , W _ay , and W _az are respectively

Gaussian white noise;

is the "current" mean acceleration on the three axes;

When establishing the measurement equation, take the positioning results L _x , _Ly , and L _z of the GPS receiver in the directions of the three coordinate axes as the measurement values, then:

L _x = x + ∈ _x + ω _lx

L _y =y+∈ _y +ω _ly

L _z = z+∈ _z +ω _lz

Written in matrix form:

L=[L _x , _Ly , L _z ] ^T

V is the observation noise V=[ω _lx , ω _ly , ω _lz ] ^T

H is the observation equation

Then the corresponding measurement equation is L=HX+V (7)

So far, we have established the measurement equation of the system with the observation equation of the GPS system, and established the state transition equation of the system with the "current" statistical model and the error model of the GPS system. According to the above system equation and observation equation, let the sampling period be T. , through the discretization process ^[18] , the discrete Kalman filter equation is established as follows:

K(k+1)=P(k+1,k)H ^T (k+1)[H(k+1)P(k+1)H ^T (k+1)+R(k+1)] ^-1

(9)

P(k+1, k)=Φ(k+1, k)P(k)Φ ^T (k+1, k)+Q(k) (10)

P(k+1)=[I-K(k+1)H(k+1)]P(k+1,k) (11)

where Φ _l (k+1, k)=

diag[Φ _lx (k+1, k), Φ _ly (k+1, k), Φ _lz (k+1, k)]

Φ _ly , Φ _lz and so on;

In formula (10), Φ(k+1, k) is the discretization matrix of the system transition matrix A:

Φ(k+1, k)=diag[Φ _x (k+1, k), Φ _y (k+1, k), Φ _z (k+1, k)]

Φ _y (k+1, k), Φ _z (k+1, k) and so on, just replace τ _ax , τ _x in (12) with τ _ay , τ _y and τ _az , τ _z is Can;

Q(k) is the discretization matrix of the system noise covariance matrix Q

R(k) is the discretization matrix of the observation noise covariance matrix R of the system

Q(k) and R(k) are both positive definite matrices;

So far, the unknown parameters in the prediction and correction models have been solved, and then the data can be processed according to the conventional Kalman filter procedure.

Further, the specific steps of the vehicle-pedestrian collision avoidance algorithm in the Step 3 are:

Step3.1: Vehicles and pedestrian equipment periodically collect their own state information through their own sensors, and upload it to the MEC server through the RSU/base station, and run the corresponding Kalman filter on the MEC server according to the state information of pedestrians and vehicles Predict the trajectories of pedestrians and vehicles;

In the above Step3.1, the movement of pedestrians needs to be regarded as nonlinear, because the movement of pedestrians has great uncertainty, and EKF is a filter for predicting the position of nonlinear moving objects; compared with pedestrians, driving on the road Due to the restrictions of the road, the motion of the vehicle can be roughly divided into two categories, one is straight, and the yaw angular velocity is zero, which can be regarded as a linear motion; the other is when the yaw angular velocity is not zero when turning is nonlinear sports;

Step3.2: Run the collision avoidance algorithm to calculate the time to collision (TTC) and judge the possibility of vehicle-pedestrian collision according to the value of TTC to warn the vehicle driver and pedestrians;

The TTC (time to collision) is:

L ₁ , L ₂ are the trajectory lengths of vehicles and pedestrians to the collision point respectively, V ₁ , V ₂ are the current speeds of vehicles and pedestrians, respectively, t _v , t _p are the communication delays from vehicles and pedestrians to the edge server, respectively, according to this The value of TTC to determine whether a collision occurs;

When TTC=0, that is, if the current vehicle and pedestrian keep the current state without taking measures, a collision may occur. At this time, the warning information is sent to the vehicle, and the vehicle is notified to take measures before the accident occurs, so as to avoid the collision.

Further, in the Step 3.2, due to the unavoidable error in the algorithm, set TTCR (time to collision range) and threshold ξ, when

TTCR≤ξ

Collision warning is carried out to eliminate the error caused by trajectory prediction and the influence of vehicle length.

As can be seen from Figure 2, when TTC=0, the higher the vehicle speed, the lower the warning accuracy, because the higher the vehicle speed, the greater the impact of errors and delays on it. When the threshold ξ increases, the early warning accuracy rate will first increase and reach the highest value, because the increase of the threshold will gradually reduce the influence of the system error. After reaching the peak, the early warning accuracy begins to decline, because the time difference, that is, the larger the threshold, the less accurate the trajectory prediction is. And after the threshold is increased to a certain extent, there is a situation that one or both of the vehicle and the pedestrian have passed the collision point, and the probability of such a situation will increase with the increase of the threshold.

Moreover, it can be seen that the curve of the warning accuracy rate will be “steeper” when the vehicle speed is higher, and the curve will be smoother when the vehicle speed is lower. Because the higher the vehicle speed, the more sensitive it is to the delay, the lower the vehicle speed, the more accurate the trajectory prediction is, and the corresponding accuracy rate is higher.

It can be seen from Figure 3 that the pedestrian speed and vehicle speed have roughly the same impact on the system, and the trend of the curve also conforms to the trend that the faster the speed is, the more "steep" it is, and the slower the speed is, the more "smooth". When TTC=0, speed and early warning success rate are inversely correlated. The difference is that when the threshold ξ is the same, the early warning accuracy rate in this part is generally higher than the accuracy rate in 5.3, because the pedestrian speed is slower than the vehicle, and the impact on the system will not be as obvious as the vehicle.

Figure 4 is a comparison test result diagram. In the present invention, the collision point is calculated according to the trajectory of the vehicle and pedestrian, and each pedestrian and vehicle have their own trajectory, so the total number of vehicle pedestrians will affect the system performance. This part studies the impact of the total number of pedestrians and vehicles on the system. And contrast with the approach in Avoiding car-pedestrian collisions using a VANETto cellular communication framework. It can be seen from Figure 4 that as the total number of pedestrians and vehicles increases, the warning accuracy rate will decrease. Because at this time, there will be a situation where pedestrians and vehicles will not collide, but an early warning will be given only because the distance is relatively close. This is due to the influence of the threshold ξ. The system may determine this situation as a collision.

And the trajectory increases. A single vehicle or pedestrian may receive multiple warnings. At this time, the total number of warnings will increase. When a vehicle receives accurate warning information to decelerate and successfully avoid a collision, other warning information it receives will be invalid. The number of successful early warnings of vehicles or pedestrians is less than or equal to 1, that is, the upper limit of the number of successful early warnings is small, but will only increase with the increase of the total number and the upper limit is very high, so the change trend of the early warning success rate is as follows: The total number of pedestrians and vehicles increases and gradually decreases. Compared with the algorithm in Avoiding car-pedestrian collisions using a VANET tocellular communication framework, the algorithm in this paper has a higher early warning success rate.

The specific embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned embodiments, and can also be made within the scope of knowledge possessed by those of ordinary skill in the art without departing from the spirit of the present invention. Various changes.

Claims (5)

1. a vehicle-pedestrian collision avoidance method based on edge computing, is characterized in that: comprise the following steps: Step1: Establish a network model for pedestrian device-to-vehicle communication: Arrange roadside units (RSUs)/base stations at intersections or road edges, and mobile edge computing (MEC) servers connected to the roadside units (RSUs)/base stations, so The pedestrian device is a smart phone with a GPS module, and the pedestrian device and the vehicle can communicate within the coverage of the network model; Step2: According to the current information and historical information of vehicles and pedestrian equipment, use Kalman filter to predict the future positions of vehicles and pedestrians respectively; Step3: According to the predicted vehicle and pedestrian paths in Step2, according to the mutual motion relationship, run the vehicle-pedestrian collision avoidance algorithm to give early warning of possible collisions. 2. a kind of vehicle-pedestrian collision avoidance method based on edge computing according to claim 1, is characterized in that: the concrete step of predicting the future position of vehicle in described Step2 is: Assuming that the car has a constant rotation rate and speed, the state vector of the selected vehicle is: X=[xyv θ ω] ^T θ is the yaw angle, which is the angle between the tracked target vehicle and the x-axis in the current vehicle coordinate system, the counterclockwise direction is positive, and the value range is [0, 2π), and ω is the yaw angular velocity; The motion states of roads and traffic regulations are generally straight or turning. Both motion states can be represented by this state vector. When ω≠0, it is non-straight travel. At this time, the state transfer function of the model is:

When ω=0, it is a straight line, and the state transition function of the model becomes:

For the multivariate functions of (1) and (2), the multivariate Taylor series is used for expansion, the higher order number is ignored, and only the Jacobian matrix is used for linearization. In this model, the partial derivative of each element can be obtained. Jacobian Matrix J _A , when ω≠0:

When ω=0, the Jacobian matrix is:

Calculate and deal with the uncertainty Q caused by noise; in the network model, the introduction of noise mainly comes from two sources: linear acceleration and yaw angular acceleration noise, the linear acceleration and yaw angular acceleration noise will affect the state quantity (x, y, v, θ, ω), the expression of the influence of the linear acceleration and yaw angular acceleration on the state is as follows:

where μ _a ,

is the acceleration on the straight line and on the corner, Q is the covariance matrix for dealing with noise, and its expression is: Q=E[n·n ^T ]=E[GμμT G ^T ]=G·E[μμ ^T ]·G ^{T (} 4) in

Therefore, the calculation formula of the covariance matrix Q of processing noise in the CTRV model is:

When establishing the measurement equation, the data collected by the lidar and radar sensors are selected to perform sensor fusion to predict the vehicle trajectory; The first is lidar, whose data is in the Cartesian coordinate system, which can detect the vehicle position, but has no speed information, and its measurement value is the coordinates (x, y) of the target vehicle; therefore, the measurement model of lidar is still linear. , and its measurement matrix is:

The corresponding measurement equation is:

The second is the radar, whose measured target vehicle data is in the polar coordinate system, and the radar's prediction is mapped to the measurement space is nonlinear ^[16] , and its expression is:

At this time, h(x) is used to represent such a nonlinear mapping, then when solving the Kalman gain, the nonlinear process should also be linearized using Taylor's formula, and the Jacobian matrix of h(x) should be solved with reference to the prediction process. :

The unknown parameters J _H , _JA , Q, R (processing noise, generally provided by the manufacturer) in the prediction and correction model have been solved, and then the data can be filtered according to the network model, so as to realize the future position of the vehicle Prediction. 3. a kind of vehicle-pedestrian collision avoidance method based on edge computing according to claim 1, is characterized in that: the concrete step of determining the future position of pedestrian in described Step2 is: The location of the pedestrian is determined according to the information collected by the GPS module of the smartphone, but there are still many errors in the dynamic positioning of the GPS module, such as satellite clock errors, etc. In order to eliminate the errors and improve the positioning accuracy, the Kalman filter method is applied to "Current" statistical model of GPS dynamic filtering; Starting directly from the positioning results output by the GPS receiver, the total error caused by various error sources of GPS positioning in all directions is equivalent to the sum of a current mean and a colored noise that conforms to the first-order Markov process; according to " The current "statistical model", the pedestrian's state vector is selected as:

The state equation of the system can be defined as:

Among them, A is the system state transition matrix: A=diag[A _x , A _y , A _z ]

A _y , A _z and so on;

W(t) is: =[0 0 W _ax W _x 0 0 W _ay W _y 0 0 W _az W _z ] ^T is the system noise vector; τ _ax , τ _ay , τ _az are the acceleration-related time constants respectively; τ _x , τ _y , τ _z are the relevant time constants corresponding to the Markov process, respectively; W _ax , W _ay , and W _az are respectively

Gaussian white noise;

is the "current" mean acceleration on the three axes; When establishing the measurement equation, take the positioning results L _x , _Ly , and L _z of the GPS receiver in the directions of the three coordinate axes as the measurement values, then: L _x = x + ∈ _x + ω _lx L _y =y+∈ _y +ω _ly L _z = z+∈ _z +ω _lz Written in matrix form: L=[L _x , _Ly , L _z ] ^T V is the observation noise V=[ω _lx , ω _ly , ω _lz ] ^T H is the observation equation

Then the corresponding measurement equation is L=HX+V (7) So far, we have established the measurement equation of the system with the observation equation of the GPS system, and established the state transition equation of the system with the "current" statistical model and the error model of the GPS system. According to the above system equation and observation equation, let the sampling period be T. , through the discretization process ^[18] , the discrete Kalman filter equation is established as follows:

K(k+1)=P(k+1,k)H ^T (k+1)[H(k+1)P(k+1)H ^T (k+1)+R(k+1)] ^-1 (9) P(k+1, k)=Φ(k+1, k)P(k)Φ ^T (k+1, k)+Q(k) (10) P(k+1)=[I-K(k+1)H(k+1)]P(k+1,k) (11) where Φ _l (k+1, k)= diag[Φ _lx (k+1, k), Φ _ly (k+1, k), Φ _lz (k+1, k)]

Φ _ly , Φ _lz and so on; In formula (10), Φ(k+1, k) is the discretization matrix of the system transition matrix A: Φ(k+1, k)=diag[Φ _x (k+1, k), Φ _y (k+1, k), Φ _z (k+1, k)]

Φ _y (k+1, k), Φ _z (k+1, k) and so on, just replace τ _ax , τ _x in (12) with τ _ay , τ _y and τ _az , τ _z is Can; Q(k) is the discretization matrix of the system noise covariance matrix Q

R(k) is the discretization matrix of the observation noise covariance matrix R of the system

Q(k) and R(k) are both positive definite matrices; So far, the unknown parameters in the prediction and correction models have been solved, and then the data can be processed according to the conventional Kalman filter procedure. 4. a kind of vehicle-pedestrian collision avoidance method based on edge computing according to claim 1, is characterized in that: the concrete steps of vehicle-pedestrian collision avoidance algorithm in described Step3 are: Step3.1: Vehicles and pedestrian equipment periodically collect their own state information through their own sensors, and upload it to the MEC server through the RSU/base station, and run the corresponding Kalman filter on the MEC server according to the state information of pedestrians and vehicles Predict the trajectories of pedestrians and vehicles; In the above Step3.1, the movement of pedestrians needs to be regarded as nonlinear, because the movement of pedestrians has great uncertainty, and EKF is a filter for predicting the position of nonlinear moving objects; compared with pedestrians, driving on the road Due to the restrictions of the road, the motion of the vehicle can be roughly divided into two categories, one is straight, and the yaw angular velocity is zero, which can be regarded as a linear motion; the other is when the yaw angular velocity is not zero when turning is nonlinear sports; Step3.2: Run the collision avoidance algorithm to calculate the time to collision (TTC) and judge the possibility of vehicle-pedestrian collision according to the value of TTC to warn the vehicle driver and pedestrians; The TTC (time to collision) is:

L ₁ , L ₂ are the trajectory lengths of vehicles and pedestrians to the collision point, respectively, V ₁ , V ₂ are the current speeds of vehicles and pedestrians, respectively, t _v , t _p are the communication delays from vehicles and pedestrians to the edge server, respectively, according to this The value of TTC to determine whether a collision occurs; When TTC=0, that is, if the current vehicle and pedestrian keep the current state without taking measures, a collision may occur. At this time, the warning information is sent to the vehicle, and the vehicle is notified to take measures before the accident occurs, so as to avoid the collision. 5. a kind of vehicle-pedestrian collision avoidance method based on edge computing according to claim 4, is characterized in that: in described Step3.2, because there is an unavoidable error in the algorithm, therefore, set TTCR (time to collisionrange) and threshold ξ, when TTCR≤ξ Collision warning is carried out to eliminate the error caused by trajectory prediction and the influence of vehicle length.

Images