CN109727490A - An adaptive correction prediction method for surrounding vehicle behavior based on driving prediction field - Google Patents

Abstract

The invention discloses an adaptive correction prediction method for surrounding vehicle behaviors based on a driving prediction field. Step 1: discretization of surrounding vehicle behaviors and data set preprocessing: dividing surrounding vehicle behaviors into N typical behaviors according to the horizontal and vertical directions; step 2: Obtain the time series data of the participating vehicles in the traffic environment: Each vehicle participating in the traffic environment uses the positioning system to obtain the position, speed, and acceleration of the vehicle at each moment in real time; Step 3: Establish a driving prediction field: establish a vehicle based on safety, efficiency, driving The driving prediction field E _P of the three elements of comfort, E _P =E _S +E _E +E _C ; Step 4: Establish a surrounding vehicle behavior prediction model based on the maximum likelihood estimation method; Step 5: Real-time prediction of surrounding vehicle behavior and model automatic Adapt to correction. The invention comprehensively considers the safety, efficiency and driving comfort affecting the driver's behavior, establishes a driving prediction field in the driving area of the target vehicle and analyzes it qualitatively and quantitatively, and proposes a new idea for the behavior prediction of surrounding vehicles.

Description

An adaptive correction prediction method for surrounding vehicle behavior based on driving prediction field

technical field

The invention belongs to the technical field of intelligent driving, and in particular relates to an adaptive correction prediction method for surrounding vehicle behavior based on a driving prediction field.

Background technique

Nowadays, both advanced driver assistance systems and fully autonomous vehicles have aroused extensive research interests of scholars in various fields. There is no doubt that automotive intelligence has become one of the most important trends and trends in the development of the automotive industry. The reason for this is that smart vehicles not only provide more efficient, safer and cleaner performance in transportation systems, but also free up human manipulation of the vehicle during driving. The real traffic environment is often complex and highly uncertain. In this environment, human beings are actually very good drivers, because they can continuously learn and have the ability to infer the behavioral intentions of surrounding traffic participants and predict their movement. Therefore, the "driving brain" of intelligent vehicles can achieve true "smart" driving only if it achieves the sublimation of the perception of the increasingly traffic environment like human beings. That is, from a technical point of view, the vehicle should be able to predict the future motion behavior of multiple targets in the surrounding complex traffic environment to improve its ability to understand the traffic environment, which is of great significance for formulating reasonable and efficient trajectory planning and decision control. In recent years, with the development of visual perception technology and communication technology, lidar, millimeter-wave radar, and camera multi-sensor fusion systems enable intelligent vehicles to monitor the surrounding traffic environment in real time. The device can also help the vehicle to accurately obtain additional information around it, which facilitates the behavior prediction of surrounding target vehicles.

At present, scholars at home and abroad have carried out a lot of research work on the identification and prediction of the driver's driving behavior or driving intention, and have achieved excellent results. When they process the feature input data of the prediction model, they often obtain the vehicle operating parameters of the vehicle (steering wheel angle, longitudinal acceleration, and the distance between the vehicle and the center line of the lane, etc.) or driver parameters (average gaze in the left rearview mirror). However, if we start from the detailed data of the surrounding vehicles and their drivers, it will undoubtedly increase the transmission of in-vehicle communication equipment or remote communication platforms. The requirements of efficiency also endanger the problem of information security, so this method is not suitable for the behavior prediction of surrounding vehicles. Another feasible solution is to directly predict the behavior through the prototype trajectories of the surrounding target vehicles, which can achieve higher computational efficiency, but such methods use the predicted target vehicle as an independent individual when performing motion prediction. It is difficult to carry out stable and accurate long-term motion behavior prediction in the traffic environment by conducting research without ignoring the influence of the surrounding traffic environment participants such as other vehicles (including the host vehicle) on their behavior. Whether it is a human-driven vehicle or a vehicle with autonomous driving capabilities, it is an agent that responds to the stimuli of the surrounding environment.

Therefore, when predicting the behavior of surrounding vehicles, it is necessary to excavate the long-term influencing factors of vehicle behavior. The invention proposes an adaptive prediction method for surrounding vehicle behavior based on a driving prediction field, and proposes a driving prediction field considering three factors of safety, efficiency, and driving comfort in decision-making, and based on this, a surrounding vehicle behavior prediction model is established. And combined with the behavior recognition results of the predicted vehicle, adaptive correction is used to improve the prediction accuracy.

SUMMARY OF THE INVENTION

Aiming at the requirements of long-term stability and reliability of the surrounding vehicle behavior prediction, the present invention proposes a driving prediction field based on the influence of the driver's behavioral decision-making elements, and based on this, an adaptive prediction method for surrounding vehicle behavior is proposed, which can real-time It can accurately predict the behavior of surrounding target vehicles and provide a reference for the intelligent vehicle's own decision-making planning. The object of the present invention can be achieved through the following technical solutions. An adaptive prediction method for surrounding vehicle behavior based on a driving prediction field, which specifically includes:

Step1: Discretization of surrounding vehicle behavior and data set preprocessing;

The surrounding vehicle behaviors are divided according to the combination of lateral and vertical aspects, and the discretization is divided into N typical behaviors b _i . The NGSIM traffic data set is denoised and an effective data set is extracted. According to the discretization method of vehicle behavior, each data is marked to calibrate the corresponding behavior type.

Step2: Obtain the time series data of the participating vehicles in the traffic environment;

Each vehicle participating in the traffic environment uses the on-board GPS and IMU joint positioning system to obtain the position (x, y), velocity (V _x , V _y ) and acceleration (a _x , a _y ) of the vehicle at each moment in real time. The host vehicle uses the D2D (Device-To-Device) proximity communication service (ProSe) of the LTE module in the V2V communication technology to obtain the status and timing information of the surrounding vehicles in the traffic environment in real time. For the surrounding target vehicles that need to predict the vehicle behavior, the vehicles in front and rear of the current lane and the vehicles in the adjacent lanes are taken as the influencers of the behavior.

Step3: Establish a driving prediction field;

Aiming at the fact that the driver is an agent that responds to the incentives of the surrounding environment by "seeking benefits and avoiding disadvantages", a driving prediction field EP based on the three factors of safety, efficiency, and driving comfort decision-making is established, where _{EP =} E _S +E _E +E _C , wherein the safety prediction field E _S , the efficiency prediction field E _E , and the driving comfort prediction field E _C . Let the forecast cycle time be ΔT.

(1) Establish a safety prediction field

The unit safety potential value of any point in the driving area of the target vehicle affected by the surrounding jth vehicle

Among them, (X, Y) is the position of any point in the driving area of the target vehicle; G _S is the undetermined constant of the driving safety prediction field; δ _j is the vehicle type coefficient of the j-th vehicle around; M _j is the equivalent of the j-th vehicle around The mass ratio is the reciprocal of the product of the length, width and height of the jth vehicle; (x _[j] , y _[j] ) is the current position vector of the jth vehicle around the target vehicle; is the velocity vector of the jth vehicle around the target vehicle at the current moment; is the acceleration vector of the jth vehicle around the target vehicle at the current moment; ΔT is the behavior prediction cycle time of the surrounding target vehicle; || ||2 is the 2-norm sign of the vector.

Then the unit safety potential value of any point in the driving area of the target vehicle

(2) Establish an efficiency prediction field

The unit efficiency potential value of any point in the target vehicle's driving area

Y is the longitudinal position of any point in the driving area of the target vehicle; G _E is the undetermined constant of the driving efficiency prediction field; M ₀ is the equivalent mass ratio of the target vehicle, which is the reciprocal of the product of the length, width and height of the target vehicle; y _[0] is the target vehicle The longitudinal position of the vehicle at the current moment;

(3) Establish a driving comfort prediction field

The unit driving comfort potential value at any point in the target vehicle's driving area

(X, Y) is the position of any point in the driving area of the target vehicle; G _C is the undetermined constant of the driving comfort prediction field; (x _[0] , y _[0] ) is the position vector of the target vehicle at the current moment; is the velocity vector of the target vehicle at the current moment; ΔT is the behavior prediction cycle time of the surrounding target vehicle; || ||2 is the 2-norm sign of the vector.

Step4: Establish a behavior prediction model of surrounding vehicles;

Fit the similarity trajectory corresponding to each vehicle behavior b _i , and set the area where the target vehicle travels according to the similarity trajectory as Calculate the predicted driving field strength and sum of each behavior of the surrounding target vehicle

Among them, K _S is the weight coefficient of the field strength sum for safety prediction; _{KE is the weight coefficient of the efficiency prediction field strength sum; K S is} the weight coefficient of the driving comfort prediction field strength sum;

Process field strength and normalization into probabilities for each vehicle behavior

Write the likelihood function L(θ) = _{ΠP_predict} ( _bi ), where θ = {K _S , K _E , K _C }.

Based on the preprocessed NGSIM data set in Step1, the maximum likelihood estimator is calculated using the mature conjugate gradient method That is, the initial weight coefficients K _{S_0} , K _{E_0} , K _{C_0} .

Then get the surrounding vehicle behavior prediction model function

Step5: Real-time prediction of surrounding vehicle behavior and model adaptive correction

The main vehicle is located in the real traffic environment. Firstly, according to Step 2, the time series data of the participating vehicles in the traffic environment are obtained in real time. According to the prediction model established in Step 4, the vehicle behavior probability of the surrounding target vehicles within the prediction cycle time ΔT is predicted in real time, and the behavior corresponding to the maximum probability is used as Prediction of surrounding vehicle behavior. In order to further improve the prediction accuracy, the present invention also proposes a weight coefficient regression correction method. Taking the lateral and longitudinal displacement, velocity, and acceleration of surrounding target vehicles as observation variables, online recognition of surrounding target vehicle behaviors is carried out based on the HMM model, and the recognition probability P _recognize (b _i ) of each behavior within the prediction cycle time is obtained.

Write the behavior prediction model probability function as

P _{predict_k} (bi )=f _k ( _{K S_k} ,K _{E_k} ,K _{C_k} )

That is, P _{predict_k} (b _i ) is a function of K _{S_k} , K _{E_k} , K _{C_k} , where K _{S_k} is the weight coefficient of the sum of the time safety field strengths of the kth prediction cycle; K _{E_k} is the time efficiency field field of the kth prediction cycle The weight coefficient of the strong sum; K _{C_k} is the weight coefficient of the driving field field strength sum at the kth prediction cycle time.

Construct a cost function between predicted and identified values

Among them, N is the number of surrounding vehicle behavior types, and N=9.

Construct Correction Function where is the α correction rate coefficient.

The weight coefficients are corrected online by the correction function to predict the cycle time ΔT one by one to realize adaptive prediction and further improve the accuracy of the behavior recognition and prediction of surrounding vehicles.

Beneficial effects of the present invention:

(1) Starting from the influence of the surrounding traffic participating vehicles on the behavior of the target vehicle, the defect that only the predicted target vehicle is regarded as a single independent individual is alleviated, and the long-term stable prediction of the behavior recognition of surrounding vehicles can be realized.

(2) Comprehensively considering the safety, efficiency and driving comfort that affect the behavior of the driver, a driving prediction field is established in the driving area of the target vehicle, and qualitative and quantitative analysis are carried out, which proposes a new idea for the prediction of surrounding vehicle behavior.

(3) A prediction model of surrounding vehicle behavior based on driving prediction field is proposed, which has better prediction accuracy and less prediction time.

(4) A regression correction method for the prediction of surrounding vehicle behavior is proposed. The HMM model identifies the vehicle behavior results in the prediction cycle time and continuously adapts the weight coefficient vector of the prediction model to further improve the prediction accuracy of surrounding vehicle behavior.

Description of drawings

Fig. 1 is a block diagram of the adaptive correction prediction method of surrounding vehicle behavior based on the driving prediction field;

Figure 2. Discretization of surrounding vehicle behavior;

Fig. 3 Traffic environment reference vehicle of the target vehicle;

(a). The target vehicle is in the middle lane; (b). The target vehicle is in the right lane; (c). The target vehicle is in the left lane);

Figure 4. The area where a certain behavior _bi of the surrounding target vehicle passes;

Figure 5 A typical traffic environment H;

Fig. 6 Safely predicted field strength distribution in H traffic environment;

Fig. 7 Efficient prediction field strength distribution in H traffic environment;

Figure 8. Field strength distribution for driving comfort prediction in H traffic environment;

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

As shown in Figure 1, the implementation of the present invention includes the following steps:

Step1: Discretization of surrounding vehicle behavior and data set preprocessing

According to the characteristics of many uncertain factors and complex and difficult to distinguish the behavior of surrounding target vehicles, the possible behaviors are divided into two directions: horizontal behavior and vertical behavior for combined division. From Lane Change to Left, Lane Keep, Lane Change to Right in the lateral behavior, SpeedIncrease, Speed Keep, Deceleration ( Speed Decrease) discretizes the behavior of surrounding vehicles into N typical behaviors b _i , N=9, which are left lane change deceleration (LCL-SD), left lane change uniform speed (LCL-SK), left lane change acceleration (LCL -SI), maintain lane deceleration (LK-SD), maintain lane constant speed (LK-SK), maintain lane acceleration (LK-SI), right lane change deceleration (LCR-SD), right lane change deceleration (LCR-SK) , right lane change deceleration (LCR-SI). As shown in Fig. 2, each trajectory corresponds to each behavior bi, where _1≤i≤9 . Using Python's Pandas data analysis package to develop a computer program, denoise the NGSIM traffic data set and extract an effective data set, and label each data to calibrate the corresponding behavior type according to the discretization method of vehicle behavior.

Step2: Obtain the time series data of the participating vehicles in the traffic environment

Each vehicle participating in the traffic environment has an independent ID, and uses the on-board GPS and IMU joint positioning system to obtain the position (x, y), speed (V _x , V _y ), acceleration (a _x , a _y ). Considering the real-time and robustness of data transmission in the subsequent steps, the data collection frequency is 50Hz, that is, 0.02s is the duration of the two data intervals before and after the acquisition. The host vehicle connects to the V2V communication network through the PC5 interface, and uses the D2D (Device-To-Device, inter-device) proximity communication service (ProSe) in the LTE module to obtain real-time status and timing information of the surrounding vehicles in the traffic environment. The full name of the V2V communication technology in English is Vehicle-To-Vehicle Technology. Because it can overcome the Doppler effect caused by high-speed movement and the complex communication environment, it is widely used in the field of smart cars, and its D2D module can be established without going through a base station. Intercommunication between adjacent vehicle terminals. The host vehicle locks a surrounding vehicle as the target vehicle for behavior prediction, and constructs the traffic environment of the target moving vehicle according to the real-time state time series data obtained in the above steps. In the real traffic environment, human drivers can obtain surrounding traffic information through the forward viewing area, rear-view mirrors and rear-view cameras, and the intelligent driving system can also perceive visual perception through cameras, lidar and millimeter-wave radar. The module achieves the same purpose, so the influence of traffic information on the target car is propagated alternately before and after the neighbors. When constructing the surrounding traffic environment of the target vehicle, take the front and rear vehicles in the current lane of the target vehicle and the front and rear vehicles in the adjacent lane as the influencers of its behavior, and set the number of surrounding affected vehicles as h. Figure 3 shows the ways in which the target vehicle is in different lanes and the surrounding vehicles on the three-way one-way lane, a is the target vehicle is in the middle lane, and h is 6 at this time; b is the target vehicle is in the left lane, and h is 4 at this time; c means that the target vehicle is in the right lane, and h is 4 at this time.

Step3: Establish a driving prediction field

Whether it is a human-driven traffic vehicle or a traffic vehicle with automatic driving ability, it is an intelligent body that responds to the incentives of the surrounding vehicle-road collaborative traffic environment, and the driving behavior of surrounding vehicles has the It has great uncertainty and variability, and is affected by the checks and balances of various expected benefits of the self-vehicle. Therefore, a driving prediction field EP based on decision influencing factors is established below in the surrounding target vehicle driving area to _quantify the influencing factors generated by surrounding vehicle behaviors. According to the three factors (safety, efficiency, and driving comfort) that affect the driver's behavior, the driving prediction field is divided into three sub-forecasting fields, namely, the safety prediction field ES, the efficiency prediction field _{E E ,} and the driving comfort prediction field E _C . where E _P =E _S +E _E +E _C . For the convenience of description, a typical traffic environment H shown in FIG. 5 is taken.

(1) Establish a safety prediction field

The safety prediction field characterizes the effect of safety on the driver and operator of the surrounding target vehicle. Taking h traffic vehicles in the front and rear of the target vehicle as the "charges" to generate the safety potential, and taking the position, speed and acceleration of the h traffic vehicles in the front and rear as the main variables that affect the safety potential value.

Write out the unit safety potential value of any point in the driving area of the target vehicle affected by the surrounding jth vehicle

Among them, (X, Y) is the position of any point in the driving area of the target vehicle; G _S is the undetermined constant of the driving safety prediction field; δ _j is the vehicle type coefficient of the j-th vehicle around; M _j is the equivalent of the j-th vehicle around The mass ratio is the reciprocal of the product of the length, width and height of the jth vehicle; (x _[j] , y _[j] ) is the current position vector of the jth vehicle around the target vehicle; is the velocity vector of the jth vehicle around the target vehicle at the current moment; is the acceleration vector of the jth vehicle around the target vehicle at the current moment; ΔT is the behavior prediction cycle time of the surrounding target vehicle; || || ₂ is the 2-norm sign of the vector.

Then the unit safety potential value of any point in the driving area of the target vehicle

As shown in Fig. 6, the simulation obtains the field strength distribution of the safety prediction field in the H traffic environment.

(2) Establish an efficiency prediction field

The Efficiency Prediction Field characterizes the impact of safety on surrounding target vehicle drivers. Taking the target vehicle as the "charge" that generates the efficiency field potential, the longitudinal position of the target vehicle is the main variable that affects the efficiency potential value.

Write the unit efficiency potential value at any point in the target vehicle's driving area

Y is the longitudinal position of any point in the driving area of the target vehicle; G _E is the undetermined constant of the driving efficiency prediction field; M ₀ is the equivalent mass ratio of the target vehicle, which is the reciprocal of the product of the length, width and height of the target vehicle; y _[0] is the target vehicle The longitudinal position of the vehicle at the current moment;

As shown in Figure 7, the simulation obtains the field strength distribution of the efficiency prediction in the H traffic environment.

(3) Establish a driving comfort prediction field

The driving comfort prediction field characterizes the influence of driving comfort on the surrounding target vehicle driver. Taking the target vehicle as the "charge" that generates the driving comfort field potential, the lateral and longitudinal acceleration of the target vehicle going to a certain position in the driving area is taken as the main variable that affects the driving comfort potential value.

Write the unit driving comfort potential value at any point in the driving area of the target vehicle

(X, Y) is the position of any point in the driving area of the target vehicle; G _C is the undetermined constant of the driving comfort prediction field; (x _[0] , y _[0] ) is the position vector of the target vehicle at the current moment; is the velocity vector of the target vehicle at the current moment; ΔT is the behavior prediction cycle time of the surrounding target vehicle; || || ₂ is the 2-norm sign of the vector.

As shown in Figure 8, the field strength distribution for driving comfort prediction in H traffic environment is obtained by simulation.

Step4: Establish a prediction model of surrounding vehicle behavior

Taking the target vehicle in the middle lane as an example, the safe area in the drivable area is divided into 9 behavior hot areas according to the behavior type. The decision-making layer simulating the surrounding target vehicles to generate and execute a certain vehicle behavior b _i will fuzzy estimate the balance effect of the predicted field strength and field strength of each vehicle. The location of the target vehicle. Fit the similarity trajectory corresponding to each vehicle behavior b _i , and take the area swept by the vehicle under the trajectory as the integral area

Normalize the predicted driving field strength sum, that is, convert the predicted driving field strength sum of each vehicle behavior into the predicted probability corresponding to the vehicle behavior,

Write the likelihood function L(θ) = ΠP _predict ( _bi ), where θ = {K _S , K _E , K _C }.

Using the preprocessed NGSIM data set in Step1 as the sample set, use the mature conjugate gradient method to calculate the maximum likelihood estimator That is, the initial weight coefficients K _S,0 , K _E,0 , K _C,0 .

Then get the surrounding vehicle behavior prediction model function

Step5: Real-time prediction of surrounding vehicle behavior and model adaptive correction

Human drivers or intelligent driving systems operating surrounding vehicles are also affected by their own decision-making mechanisms in the surrounding traffic environment, and their expectations for safety, efficiency, and driving comfort have distinct personalities, and they are expected to pursue long-term driving behavior predictions. is stable in the time domain. The main vehicle is located in the real traffic environment. First, the time series data of the participating vehicles in the traffic environment are obtained in real time according to Step 2, and the typical behavior probability P _predict (b _i ) of the surrounding target vehicles within the prediction cycle time ΔT is predicted in real time according to the prediction model established in Step 4, The vehicle behavior type corresponding to the output maximum behavior probability is the prediction result.

In order to further improve the prediction accuracy, the present invention also proposes a weight coefficient regression correction method, so as to reduce the error rate caused by the weight coefficient deviation in the real scene to the behavior prediction result. When a prediction time period ΔT expires, a Hidden Markov Model (HMM) is used to identify the typical behavior types of surrounding vehicles within the prediction period time ΔT.

Let the hidden Markov model be a quintuple (Q, V, A, B, π). The observable state is expressed as V={V ₁ , V ₂ ,...,VM }, where _M is the number of observed states; the hidden state is expressed as Q={Q ₁ , Q ₂ ,...,Q _N }, where N is the hidden state Number of. I is a state sequence of length T, O is a corresponding observation sequence, I={I ₁ ,I ₂ ,...,I _T }, O={O ₁ ,O ₂ ,...,O _T }.

A=[a _ij ] _N×N is the hidden state transition probability matrix, and its elements represent the transition probability between each hidden state in the HMM model. in,

a _ij =P(I _t+1 =Q _j ∣I _t =Q _i ),i=1,2...,N; j=1,2...,N

is the probability that the hidden state is Q _j at time t+1 under the condition that the hidden state is Q _i at time t.

B=[b _j (k)] _N×M is a confusion matrix, and its elements represent the transition probability between each hidden state and observed state in the HMM model. in,

b _j (k)=P(O _t =V _k ∣I _t =Q _j ),k=1,2...,M; j=1,2...,N

Indicates that at time t, the hidden state Q _j is the probability that the observed state is O _t under the condition.

π=(π _i ) is the initial state probability matrix, wherein π _i =P(I ₁ =Q _i ), _i =1,2,...,N is the probability of each hidden state Qi at the initial time t=1.

The behavior recognition of surrounding vehicles with HMM model can be divided into two stages:

a. Model training and learning: Initialize each vehicle behavior recognition model to obtain initial parameters N, M, A, B, π; extract the feature data O _t ( Lateral displacement, velocity, acceleration and longitudinal displacement, velocity, acceleration), the observation sequence O=O ₁ O ₂ O ₃ ... O _ΔT composed of characteristic data O _t at each time point. The observation sequence is used as the input of HMM parameter learning. According to the initial parameters after model initialization, the Baum-Welch iterative algorithm is used to adjust the parameters of the model λ=(A, B, π) to maximize the probability function, and gradually update the model parameters. Obtain the optimal HMM model for each behavior type;

b. Online test identification: Then use the trained vehicle behavior HMM identification model to extract and encode the surrounding target vehicles to be identified to form the observation sequence as the model input, and use the forward algorithm to calculate the observation sequence under each HMM model. The probability P _{recognize_k} (b _i ).

Write the behavior prediction model probability function as

P _{predict_k} (bi )=f _k ( _{K S_k} ,K _{E_k} ,K _{C_k} )

That is, P _{predict_k} (b _i ) is a function of K _{S_k} , K _{E_k} , K _{C_k} , where k represents the kth prediction cycle time ΔT.

Write the residual sum of squares function of the predicted probability value and the recognition probability value, that is, the cost function

Construct Correction Function where is the α correction rate coefficient.

Then, K _{S_k+1} , K _{E_k+1} , K _{C_k+1} are substituted into the parameters of the updated prediction model to predict the behavior of the target vehicle within the next passing prediction cycle time ΔT. The weight coefficients are corrected online by the correction function to predict the cycle time one by one, so as to realize the adaptive prediction and further improve the accuracy of the behavior recognition and prediction of surrounding vehicles.

The series of detailed descriptions listed above are only specific descriptions for the feasible embodiments of the present invention, and they are not used to limit the protection scope of the present invention. Changes should all be included within the protection scope of the present invention.

Claims (8)

1. a surrounding vehicle behavior adaptive correction prediction method based on driving prediction field, is characterized in that, comprises the steps: Step 1: Discretization of surrounding vehicle behavior and data set preprocessing; Step 2: Obtain the time series data of the participating vehicles in the traffic environment; Step 3: Establish a driving prediction field, including a safety prediction field, an efficiency prediction field, and a driving comfort prediction field; Step 4: Establish a model for predicting the behavior of surrounding vehicles Step 5: Real-time prediction of surrounding vehicle behavior and model adaptive correction. 2. a kind of surrounding vehicle behavior adaptive correction prediction method based on driving prediction field according to claim 1, is characterized in that, the concrete process of described step 1 comprises: The surrounding vehicle behavior is divided according to the combination of horizontal and vertical aspects, and the discretization is divided into N typical behaviors b _i . The NGSIM traffic data set is denoised and an effective data set is extracted. The data calibration corresponds to the behavior type. 3. a kind of surrounding vehicle behavior adaptive correction prediction method based on driving prediction field according to claim 2, is characterized in that, described N typical behaviors are specifically: Left lane change deceleration, left lane change uniform speed, left lane change acceleration, maintain lane deceleration, maintain lane uniform speed, maintain lane acceleration, right lane deceleration, right lane deceleration, right lane deceleration. 4. a kind of surrounding vehicle behavior adaptive correction prediction method based on driving prediction field according to claim 2, is characterized in that, the concrete realization of described step 2 comprises: Each vehicle participating in the traffic environment uses the on-board GPS and IMU joint positioning system to obtain the position (x, y), speed (V _x , V _y ) and acceleration (a _x , a _y ) of the vehicle at each moment in real time; the main vehicle uses The D2D proximity communication service of the LTE module in the V2V communication technology obtains the status and timing information of the surrounding vehicles in the traffic environment in real time. 5. a kind of surrounding vehicle behavior adaptive correction prediction method based on driving prediction field according to claim 1, is characterized in that, the concrete realization of described step 3 comprises: The establishment method of the safety prediction field: Taking the front and rear of h surrounding traffic vehicles of the target vehicle as the "charges" to generate the safety field potential, and taking the position, speed and acceleration of the front and rear h surrounding traffic vehicles as the main variables affecting the safety potential value; Write out the unit safety potential value of any point in the driving area of the target vehicle affected by the surrounding jth vehicle Among them, (X, Y) is the position of any point in the driving area of the target vehicle; G _s is the undetermined constant of the driving safety prediction field; δ _j is the vehicle type coefficient of the jth vehicle around; M _j is the equivalent of the jth vehicle around The mass ratio is the reciprocal of the product of the length, width and height of the jth vehicle; (x _[j] , y _[j] ) is the current position vector of the jth vehicle around the target vehicle; is the velocity vector of the jth vehicle around the target vehicle at the current moment; is the acceleration vector of the jth vehicle around the target vehicle at the current moment; ΔT is the behavior prediction cycle time of the surrounding target vehicle; |||| ₂ is the 2-norm symbol of the vector; Then the unit safety potential value of any point in the driving area of the target vehicle The establishment method of the efficiency prediction field: Taking the target vehicle as the "charge" that generates the efficiency field potential, and taking the longitudinal position of the target vehicle as the main variable affecting the efficiency potential value; Write the unit efficiency potential value at any point in the target vehicle's driving area Y is the longitudinal position of any point in the driving area of the target vehicle; G _E is the undetermined constant of the driving efficiency prediction field; M ₀ is the equivalent mass ratio of the target vehicle, which is the reciprocal of the product of the length, width and height of the target vehicle; y _[0] is the target vehicle The longitudinal position of the vehicle at the current moment; The establishment method of the driving comfort prediction field: Taking the target vehicle as the "charge" that generates the driving comfort field potential, the lateral and longitudinal acceleration of the target vehicle going to a certain position in the driving area is taken as the main variable affecting the driving comfort potential value; Write the unit driving comfort potential value at any point in the driving area of the target vehicle (X, Y) is the position of any point in the driving area of the target vehicle; G _C is the undetermined constant of the driving comfort prediction field; (x _[0] , y _[0] ) is the position vector of the target vehicle at the current moment; is the velocity vector of the target vehicle at the current moment; ΔT is the behavior prediction cycle time of the surrounding target vehicle; |||| ₂ is the 2-norm sign of the vector. 6. a kind of surrounding vehicle behavior adaptive correction prediction method based on driving prediction field according to claim 1, is characterized in that, the concrete realization of described step 4 comprises: Taking the target vehicle in the middle lane as an example, the safe area in the drivable area is divided into 9 behavioral hotspots according to the type of behavior; the decision-making layer simulating the surrounding target vehicle to generate and execute a certain vehicle _behavior will fuzzy estimate the prediction field of each driving sub-prediction field. According to the principle of similarity, the center point of each behavior hot zone is taken as the position of the target vehicle at the end of each behavior type; the similarity trajectory corresponding to each vehicle behavior _bi is fitted, and the The area swept under the trajectory is the integration area Among them, K _S is the weight coefficient of the field strength sum for safety prediction; _{KE is the weight coefficient of the efficiency prediction field strength sum; K S is} the weight coefficient of the driving comfort prediction field strength sum; Normalize the sum of the predicted driving field strength, that is, convert the predicted field strength of each vehicle behavior into the predicted probability corresponding to the connected vehicle behavior, Write out the likelihood function L(θ)=ΠP _predict (b _i ), where θ={K _S , K _E , K _C }; Using the preprocessed NGSIM data set in step 1 as the sample set, use the mature conjugate gradient method to calculate the maximum likelihood estimator That is, the initial weight coefficients K _S,0 , K _E,0 , K _C,0 . Obtain the surrounding vehicle behavior prediction model function 7. a kind of surrounding vehicle behavior adaptive correction prediction method based on driving prediction field according to claim 1, is characterized in that, the concrete realization of described step 5 comprises: First, obtain the time series data of the participating vehicles in the traffic environment in real time according to step 2, predict the typical behavior probability P _predict (b _i ) of the surrounding target vehicles in the prediction cycle time ΔT in real time according to the prediction model established in step 4, and output the vehicle corresponding to the maximum behavior probability. The behavior type is predicted outcome. 8 . The method for adaptively correcting and predicting the behavior of surrounding vehicles based on the driving prediction field according to claim 7 , further comprising: adopting a weight coefficient regression correction method for the prediction result, and the correction method is specifically: 9 . : When a prediction time period ΔT ends, use the Hidden Markov Model to identify the typical behavior types of surrounding vehicles within the prediction period time ΔT; Let the hidden Markov model be a quintuple (Q, V, A, B, π), and the observed states are represented as V={V ₁ , V ₂ , ..., V _M }, where M is the number of observed states ; Hidden state is expressed as Q={Q ₁ , Q ₂ ,...,Q _N }, N is the number of hidden states, I is the state sequence of length T, O is the corresponding observation sequence, I={I ₁ , I ₂ , ..., I _T }, O = {O ₁ , O ₂ , ..., O _T }; A=[a _ij ] _N×N is the hidden state transition probability matrix, and its elements represent the transition probability between each hidden state in the HMM model; wherein, a _ij =P(I _t+1 =Q _j |I _t =Q _i ), i=1, 2...,N; j=1, 2...,N is the probability that the hidden state is Q _j at time t+1 under the condition that the hidden state is Qi _i at time t; B=[b _j (k)] _N×M is a confusion matrix, and its elements represent the transition probability between each hidden state and observed state in the HMM model; among them, b _j (k)=P(O _t =V _k |I _t =Q _j ), k=1,2...,M; j=1,2...,N Indicates that at time t, the hidden state Q _j is the probability that the observed state is O _t ; π=(π _i ) is the initial state probability matrix, wherein π _i =P(I ₁ =Q _i ), _i =1,2,...,N is the probability of each hidden state Qi at the initial time t=1; The behavior recognition of surrounding vehicles with HMM model is divided into two stages: a. Model training and learning: Initialize each vehicle behavior prediction model to obtain initial parameters N, M, A, B, π; extract the feature data O _t at each collection time point in the NGSIM traffic data set processed in step 1 , namely the observation sequence O=O ₁ O ₂ O ₃ ......O _ΔT composed of the lateral displacement, velocity, acceleration and longitudinal displacement, velocity, acceleration, characteristic data O _t at each time point; take the observation sequence as HMM The input of parameter learning is based on the initial parameters after the model is initialized, and the Baum-Welch iterative algorithm is used to adjust the parameters of the model λ=(A, B, π) to maximize the probability function, gradually update the model parameters, and finally obtain the parameters of each behavior type. Optimal HMM model; b. Online test identification: Using the trained vehicle behavior HMM model, the surrounding target vehicles to be identified are extracted and encoded to form the observation sequence as the model input, and the forward algorithm is used to calculate the probability P of the observation sequence under each HMM model. _{recognize_k} (bi); Write the behavior prediction model probability function as P _{predict_k} (bi )=f _k (K _{S_k} , _{K E_k} , K _{C_k} ) That is, P _{predict_k} (b _i ) is a function of K _{S_k} , K _{E_k} , K _{C_k} , and k represents the kth prediction cycle time ΔT; Write the residual sum of squares function of the predicted probability value and the recognition probability value, that is, the cost function Construct Correction Function where is the α correction rate coefficient; Then substitute K _{S_k+1} , K _{E_k+1} , and K _{C_k+1} into the parameters of the updated prediction model to predict the next target vehicle behavior within the prediction cycle time ΔT, and use the correction function to predict the cycle time one by one to correct the weight coefficients online to achieve Adaptive prediction improves the accuracy of surrounding vehicle behavior recognition and prediction.