CN117032203A - Svo-based intelligent control method for automatic driving - Google Patents

Abstract

An automatic driving intelligent control system based on SVO theory 1) completes structured data acquisition and semi-structured data analysis of a highway driving data set, and extracts structured data to simulate highway conditions; constructing S-shaped roads with different curvature radiuses by utilizing a Matlab automatic driving tool box; 2) Training an automatic driving simulation LSTM neural network model by utilizing a Matlab automatic driving simulation data set, and carrying out common reasoning among a plurality of automobiles by introducing a 'Social' pooling layer; 3) Generating a lane change decision of the unmanned vehicle based on the fuzzy logic according to the prediction result and the current driving scene; 4) Optimizing the transverse path planning of the model by adopting a B spline curve algorithm; the global path is then planned based on the optimization principle. In path planning, utilizing a dynamic algorithm to longitudinally plan the speed, and transversely planning a local path by using a B-spline; and finally, controlling the vehicle to complete automatic driving by using the MPC.

Description

Svo-based intelligent control method for automatic driving

Technical Field

The invention belongs to the technical field of intelligent traffic, and relates to a high-speed vehicle track prediction method based on a neural network and a vehicle path planning method based on traffic prediction.

Background

With the development of unmanned technologies, an automatic driving car traveling on a public road is urgently required to improve efficiency and safety. Currently, autopilot cars lack an understanding of human behavior and therefore can only take conservative driving strategies. A conservative driving strategy can cause traffic congestion, especially at high-speed intersections. At the same time, the conservative behavior can make the automatic driving automobile more likely to be collided with by the rear-end collision of a human driver or have other traffic accidents. In the analysis of a single traffic accident in california autopilot, 57% of the traffic accidents are rear-end collisions of autopilot by human drivers, many of which occur because the behavior of autopilot is unexpected to human drivers. In order for an autonomous car to safely travel on a public road with a human driver, an autonomous system must know the intention and driving style of the human driver and make driving behavior in a driving manner conforming to the human driver.

The autopilot field is an important field of artificial intelligence research and mainly comprises a sensor, a high-speed chip, a GPU and the like. How to improve the visual ability of an automobile and how to copy the human visual response ability in a computer system is the most troublesome technical problem for the research and development personnel at present. The currently developed automotive Vision system is still very primitive. When a current person suddenly appears on the lane, even if the unmanned automobile scanner senses in time, the automatic driving system cannot immediately make an operation for controlling the automatic driving automobile to avoid, and the defect of the control system limits the danger avoiding function of the automatic driving. The difficulty with unmanned vehicles is that in addition to the need for the vehicle to maintain focus on other surrounding vehicles, the vehicle must also be able to monitor a range of other objects, such as passers-by, lanes, stop lines, traffic signs, and surrounding traffic lights, and the like, and current automated driving prediction systems are not able to address so many problems at the same time, particularly human-related problems, such as: is a vehicle currently traveling on a road surface ahead not to terminate traveling at a place other than several hundred meters? In addition, weather effects cannot be ignored, for example, when an unmanned automobile is snowy on a road surface, the problem that a guideboard and other detailed information are difficult to see is often encountered, and therefore the automobile is difficult to accurately position by means of limited information, and an unmanned system is difficult to make a correct decision.

Through a search of the prior art, wilko Schwarting, alyssa Pierson et al published an article entitled "social behavior for autonomous vehicles (social behavior of an autopilot) on Proceedings of the National Academy of Sciences of the United States of America in 2019. These articles are based on SVO theory and propose methods to apply socioeconomic tools to unmanned car predictions. Although the social psychological tool is applied to unmanned car prediction, it is difficult for an unmanned control system that needs to travel on a highway together with a plurality of human-driven traffic vehicles to fully satisfy the control requirements.

It has also been found by search that in order to further optimize the vehicle path planning algorithm, chang Guo et al, in 2021, published article entitled "The inverse variance-flatness relation in stochastic gradient descent is critical for finding flat minimum" (the inverse variance-flatness relationship in random gradient descent is crucial for finding a flat minimum) ", put forward the merits of several path planning curves. The automatic driving path planning is different from the path planning, the path planning contains time information, has a relation with speed, and the comfort evaluation index influencing the riding comfort is dominant and comprises indexes such as acceleration, jerk and the like; the path planning does not contain time information, so that analysis is only performed from the path curve itself, and the evaluation index of the reference path length and the path average curvature is important.

In summary, the problems of the prior art are: the existing automatic driving prediction system cannot be used for simultaneously solving the complex situations of the problems, particularly, most of path planning algorithms for processing the problems related to people cannot adapt to the requirement of driving at high speed (3) the unmanned automobile is difficult to accurately position by means of limited information, and the unmanned system is difficult to make a correct decision. The significance of solving the technical problems is that: based on the development of the existing artificial intelligence technology and the development of the intelligent traffic technology, the more efficient and reliable high-speed vehicle track prediction method can improve the accuracy of traffic prediction and enrich the information quantity of traffic prediction, provide a new thought for future high-speed traffic management measures and promote the development of the intelligent traffic field technology and the application of artificial intelligence in the traffic field.

Disclosure of Invention

Aiming at the problems existing in the prior art, an automatic driving intelligent control method based on SVO theory is provided, namely, a social psychological tool (SVO) is integrated into the control of an automatic driving automobile.

The technical scheme of the invention is realized in such a way that the SVO-based automatic driving intelligent control method comprises the following steps: step 1: completing structured data acquisition and semi-structured data analysis of a highway driving data set (NGSIM) in the United states, and extracting unstructured data to simulate high-speed road conditions; constructing S-shaped roads with different curvature radiuses by utilizing a Matlab automatic driving tool box, acquiring left and right boundaries of the road, and simulating 2 historical traffic vehicles and driving paths thereof;

step 2: and training an LSTM neural network by utilizing the Matlab automatic driving simulation data set, and capturing the motion characteristics and the hidden states of the surrounding vehicles in real time by utilizing the LSTM network by using the model. The model enables a system to perform common reasoning among a plurality of automobiles by introducing a 'society' pooling layer, in each period of time, an LSTM community receives pooling of hidden state information from an adjacent LSTM community, and track prediction of an unmanned automobile is completed by utilizing a socially pooled LSTM network;

step 3: according to the prediction result and the current driving scene, the complex nonlinear relation of the vehicle driving environment and the lane change decision is researched by utilizing the fuzzy theory, whether the vehicle changes lanes or not is determined by some fuzzy driving environment information and parameters, and then, the global path is planned based on the optimization principle.

Step 4: and (3) optimizing transverse path planning of the model by adopting a B-SPLINE (B SPLINE) curve algorithm, analyzing from a path curve, establishing an evaluation index of a reference path length and a path average curvature, screening the model from two aspects of model selection and performance evaluation, and optimizing longitudinal speed planning of the model by adopting a dynamic planning algorithm.

Step 5: the vehicle dynamics model is adopted, the model is converted into a linear state of a space equation, and parameters are introduced to complete relevant control: longitudinal control, namely realizing speed control of the automobile by controlling the brake/accelerator of the automobile, and transverse control, namely realizing course control by adjusting the angle of a steering wheel, and realizing Model Predictive Control (MPC).

Information pooling is a flexible tool that can abstract the intellectual relationships of time-domain space and space-domain space into low-dimensional quantifiable embeddings (e.g., normalized continuous vectors) based on the advantages of deep neural networks and extensive open-source programming. The time information of the nodes and the edges is captured into an evolution graph (EvolvingGraph).

Each track of the same scene in the model has an independent LSTM network. The LSTM then connects to each other through the social pool (S-pool) layer. Unlike conventional LSTM, this pooling layer can enable spatially close LSTM networks to share information with each other. Thus, they can characterize the strength of interaction between human drivers through aggregation operations such as average aggregation, weighted aggregation, and graph evolution messaging (or graph messaging). Furthermore, the pooling operation may embed information in low-dimensional latency in both time and space dimensions, either independently or simultaneously, with different neural network structures.

An autopilot algorithm conforming to the Social Value Orientation (SVO) is shown in table 1.

Table 1 automatic driving algorithm meeting social criteria

Human drivers, when interacting with others, create a wide variety of social preferences. Different SVO psychological tendencies, namely Lihexose and Litawnian, can be obtained by carrying out quantitative calculation on the psychological tendencies. The SVO psychological trend measures whether a rider sets his or her priority to travel on a road higher than other riders. The invention uses a non-interactive baseline algorithm to calculate other driver vehicle models except the self-driving vehicle to obtain the optimal strategy. By utilizing the data set and the track history, other drivers are modeled as dynamic driving traffic vehicles, and meanwhile, the performance standard obtained by the base line prediction is compared with the standard of an SVO theoretical model, and different static SVOs are generated aiming at different driving interaction behaviors: 1) Static lyxotropy SVO; 2) Static litaxenic SVO; 3) Dynamic social SVO. The optimal static SVO corresponds to the optimal SVO estimate that remains constant throughout the driving interaction.

The path planning is related to the speed, and the comfort evaluation indexes influencing the riding comfort are dominant, such as the indexes of acceleration, jerk and the like; the path planning does not contain time information, so that analysis is only performed from the path curve itself, and an evaluation index of the reference path length and the path average curvature is established.

The invention adopts two path evaluation indexes: path length and path average curvature. The commonly used multi-objective optimization methods are: pareto optimization front, linear weighting (multi-objective to single objective), NSGA non-dominant ordering genetic algorithm, and the like. According to the complexity of the road changing path screening, the invention selects a linear weighting method.

f＝ω ₁ f _length +ω ₂ f _curvature

Wherein f _length And f _curvature Respectively representing the path length and the path average curvature function constructed from the weight coefficients.

The path planning algorithm adopted by the invention is a B-spline curve algorithm. The B-spline curve is defined as:

wherein p is _i Is the vertex of the feature polygon; b (B) _i K is called a k-order (k-1 th order) basis function, and the order of the B-spline algorithm is the number of times plus 1. In the non-zero definition domain, the basis functions of K-1 are minimized in each node of the repetition degree K; in the internal nodes of each complex k, the p-k+1 of the non-zero basis function is the most significantAnd a plurality of times, wherein p is the number of times of the basis function.

Model predictive control takes model, prediction and control as core ideas.

The model generally selects a vehicle kinematic model or a dynamics model and seeks to transform the vehicle model into a linear state space equation. In a scene of low-speed running of the vehicle, the kinematic model of the vehicle can meet the requirements, and the kinematic model is relatively simple. It is apparent that the model employed in the present invention needs to run at high speed, so that a vehicle dynamics model is selected and converted into a linear state space equation.

Selecting a state vector, and constructing a system state equation:

wherein: m is the vehicle mass, delta is the front wheel steering angle, v is the vehicle speed, and phi is the steering speed.

The prediction is a process of obtaining a series of vehicle state quantities by continuously recursively obtaining the state quantities according to a state space equation model of the vehicle, and the state and output of the system at the future moment are expressed as a matrix:

as can be seen from the formula, both the state and the output in the prediction domain can be determined by x (k): current state quantity sum deltau of system

The control essence is a quadratic programming problem in that the control quantity is constructed at each moment so that the cumulative objective function is optimal:

wherein h=Φ ^T Q _e Φ+R _e ,g ^T ＝2E(t)Q _e Φ,Q _e ,R _e Is an extended matrix of the weight matrix Q, R. Q and R are weight matrixes, and Q is more than or equal to 0 and R is satisfied>Positive definite matrix of 0.

The invention provides an automatic driving control system framework, which creatively integrates a social psychological tool (SVO) into the design of an automatic driving automobile. Automatic driving control based on SVO theory is divided into three main parts, namely prediction decision, path planning and MPC control. Under the framework of the built prediction-decision, planning and control system, an evaluation model is built on the basis of the NGSIM data set, and the system is verified to be capable of effectively improving the efficiency and safety of the unmanned automobile running on the expressway.

B-spline (B-spline) is a special representation of spline curves in mathematical sub-discipline numerical analysis. B-splines are a generalization of the betz curve (also known as bezier curve) and can further be generalized to non-uniform rational B-splines (NURBS) that can build accurate models for more general geometries. B-spline curves are widely used in industrial fields such as vehicles and aerospace, for example: in order to make the automobile run stably during automatic driving automobile path planning, the second derivative of the running path needs to be continuous (currently, an AGV trolley mainly performs path planning through a straight line and an arc, and the problem of jitter exists in the process of performing conversion between the straight line and the arc due to inconsistent acceleration in two stages), so that a B-spline curve needs to be used. By using the B-Spline curve, the local path of the unmanned vehicle can be planned shorter and more accurately.

Model predictive control (MPC for short) is an excellent process control model in unmanned technology. The model uses an optimization algorithm to calculate a series of control input sequences over a finite time frame and then optimize the sequences. In the application field of tracking the vehicle track, the MPC modeling selects modeling based on a kinematic motion state equation, and the modeling based on a dynamic motion state equation can also be selected according to a control mode of the robot. The control itself is a quadratic programming problem with the aim of constructing control quantities at each moment so that the cumulative objective function is optimal. MPC is mainly divided into 3 key steps: model prediction, rolling optimization, feedback correction. And the MPC is utilized to build a control module, so that the unmanned vehicle is controlled more accurately.

A recurrent neural network (RNN, recurrentNeuralNetworks) is a common neural network architecture. The circulating network can be extended to longer sequences, most of the circulating network can process sequences with variable lengths, and the limitation of the traditional neural network in processing sequence information is solved.

The social value orientation (social value orientation, SVO) is based on strong interactivity of the social system between people. The driving behavior of a person is governed by two broad categories of regulations, legal regulations and social regulations. Traffic regulations constitute legal regulations, while the orientation of the social value of a person constitutes social regulations. In real traffic, when a human driver pushes efficient safety behavior of a road, the human driver drives according to implicit social specifications and legal rules, which makes the behavior of the human driver difficult to express by mathematical logic.

The beneficial effects are that: compared with the prior art, the invention has the advantages that (1) the defect that the existing automatic driving prediction system cannot simultaneously solve the complex situation of the problem is perfected, the unmanned system is helped to better process the driving problem related to the person, the unmanned system is helped to know the intention and driving style of the human driver, and the driving behavior is made in a driving mode which is fit with the human driver. (2) And comparing most path planning algorithms, and screening out a path planning algorithm (3) which meets the requirement of driving at a high speed to help the unmanned automobile to accurately position by means of limited information, so that the unmanned system makes a correct decision in the high-speed driving.

Drawings

FIG. 1 is a diagram of a high-speed simulated unmanned vehicle driving scenario employed by an embodiment of the present invention.

Fig. 2 is a diagram of predicted training results of LSTM network trajectory prediction according to an embodiment of the present invention.

FIG. 3 is a graph of trace prediction X error for an embodiment of the present invention.

FIG. 4 is a graph of trajectory prediction Y error in accordance with an embodiment of the present invention.

FIG. 5 is a graph comparing actual trajectories with predicted trajectories according to an embodiment of the present invention;

FIG. 6 is a view of obtaining a fuzzy logic based surface of view based on Matlab;

fig. 7 is a graph showing the road diameter contrast of the road change curve.

Fig. 8 is a graph of lane change curve path curvature versus.

FIG. 9 is a graph of screening based on multiple objective evaluation functions.

Fig. 10 is a high-speed unmanned control simulation diagram (initial), with the upper diagram being a control simulation implementation diagram and its lower diagram S-T curves.

Fig. 11 is a high-speed unmanned control simulation diagram (judgment), and the upper diagram is a control simulation implementation diagram and the lower diagram S-T curve thereof.

Fig. 12 is a simulation diagram (plan) of the high-speed unmanned control, the upper diagram is a control simulation implementation diagram and the lower diagram S-T curves thereof.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the following detailed description of the embodiments of the present invention is given with reference to the accompanying drawings: the embodiment is implemented on the premise of the technical scheme of the invention, and detailed implementation modes and specific operation processes are given. It should be understood that the specific examples described herein are for illustrative purposes only and that the scope of the present invention is not limited to the following examples.

Examples: the present embodiment employs the high-speed simulation driving scenario of fig. 1, and proposes a system framework that integrates a psychology tool (SVO) into the control of an autopilot. The system is divided into three main parts, namely prediction decision, path planning and MPC control. Firstly, sharing human driving parameters into an LSTM network, namely building the LSTM network with social pooling, so as to predict the track of the existing traffic vehicle on the road, and making a lane change decision of the unmanned vehicle based on fuzzy logic according to a prediction result and a current driving scene; the global path is then planned based on the optimization principle. In path planning, utilizing a dynamic algorithm to longitudinally plan the speed, and transversely planning a local path by using a B-spline; and finally, controlling the vehicle to complete automatic driving by using the MPC.

The prediction module of the embodiment mainly realizes that the NGSIM data set is obtained by Matlab automatic driving simulation, and along with different predicted roads, the obtained XY track coordinates have larger difference, and the difference can be directly reflected on the accuracy of LSTM network training, so that the standardization processing is also needed: and obtaining the average value and standard deviation of the data set by using the mean function and the std function, and converting the average value and standard deviation into a standardized data set with zero average value and unit variance. Meanwhile, in the later stage, when the LSTM network is predicted, the cell type data is used for storing the data sets for facilitating gradual training.

And an online prediction module: the basic step of constructing an online prediction module is to link an LSTM network model trained in advance by using a Stateful prediction module, and then roll and Predict 80 times at each moment point by using a for loop structure to obtain a future 8s traffic vehicle prediction track. Firstly, inputting the current actual position of the traffic vehicle, converting the actual position into a standard value, inputting the standard value into a Stateful prediction module, finally obtaining the output value of the Stateful prediction module, and inversely normalizing the output value into a normal position. And an offline prediction module: in order to combine the history track and the predicted track in a cycle and then use the combined history track and the predicted track for predicting the next cycle, an offline prediction module must be designed and built. Firstly, carrying out standardization processing on the historical actual running track coordinates, defining a prediction time domain length and an initial prediction frame, importing a previously trained LSTM network, starting to predict tracks of 8 seconds in the future for each moment point of two traffic vehicles, storing output results into corresponding variables, finally, inversely standardizing the stored output values into actual tracks, and converting the actual tracks into time sequence variables.

The network hierarchy is set to an input layer 2 level, an output layer 2 level and an hidden layer 250 level.

As shown in fig. 2, the maximum number of iterations, 'maxepchs', 500; the blue output line (upper line) represents RMSE (Root Mean Square Error) rms error, which is the square root of the ratio of the square of the deviation of the predicted value from the true value to the number of observations n. The measure is the deviation between the predicted value and the true value. The lower line in fig. 5 shows the L2 loss in trajectory prediction training, with the gradient decreasing as the error decreases. The specific parameter settings are shown in table 2.

Table 2 trajectory prediction training important parameter settings

Actual trajectory and predicted trajectory: in track prediction, rolling prediction is required to be continuously performed, and the aim is to continuously update the LSTM model, and fig. 3, fig. 4 and fig. 5 are respectively the X-axis error value and the Y-axis error value of the estimated point track and the actual track, and the actual track and the predicted track error, as can be seen from the figure, the error value is controlled within 4, the actual track almost coincides with the predicted track, and the accuracy of the predicted model is excellent according to the model reference standard provided by the argoferase: 3D Tracking and Forecasting with Rich Maps dataset.

The decision module of the embodiment adopts a fuzzy logic algorithm, the fuzzy logic is an uncertainty imitating human brain judgment, reasoning and other aspects, and the fuzzy logic is often used for judging and reasoning uncertain factors. In the process of establishing an intelligent vehicle lane change decision model, the artificial intelligence utilizes a fuzzy theory to research the complex nonlinear relation between the vehicle running environment and lane change decision, determines whether the vehicle changes lanes or not according to accurate running environment information and parameters, and has a great nonlinear relation between lane change intention generation and the information, wherein the nonlinear relation is difficult to determine according to the accurate running environment information and parameters.

Setting the lane change satisfaction degree of the vehicle, inputting a speed difference coefficient and a vehicle distance expected coefficient according to the lane change satisfaction degree, and blurring into five grades, wherein the speed difference coefficient V is selected from { small, medium, large and large } as a blurring subset, the domain is X, and the values are 0 and 1. The vehicle distance is expected to select { small, medium, large } as a fuzzy subset, and the domain is Y. Fig. 6 shows a fuzzy logic observation surface obtained in MATLAB, namely, outputting lane change willingness fai _h according to fai _d (speed difference coefficient) and fai _v (distance desired coefficient).

The path planning algorithm of this embodiment is a B-spline curve algorithm. The path planning is different from the path planning, the path planning contains time information, has a relation with speed, and the comfort evaluation index influencing the riding comfort is dominant and comprises indexes such as acceleration, jerk and the like; the path planning does not contain time information, so that analysis is only carried out from the path curve itself, evaluation indexes of reference path length and path average curvature are established, and the superiority of the model is stated from the aspects of model selection and performance evaluation.

Three curve models, a B-spline curve, a Dubins curve and a Bessel curve, were compared: as shown in FIG. 7, the comparison result of the lane change path shows that the curvature of the Dubin curve jumps twice, which indicates that if the vehicle follows the path, the unmanned vehicle stops immediately after running to the point C, the steering wheel needs to be corrected immediately, and when the unmanned vehicle stops immediately after running to the point D, the steering wheel needs to be rotated to an angle meeting the turning radius of the circular arc section, which is not as convenient as the B spline curve. As shown in fig. 8, the comparison result of the path curvature is that the path curvature of the Bessel curve meets the requirement of continuous change, but the curvature at the lane change point and the lane merging point is not 0, which can lead to the fact that the steering angle direction is not in situ after the vehicle runs to the lane merging point, namely, the steering wheel direction is incorrect, and the vehicle is deviated from the lane center line. The present invention selects B-spline curves, unlike B-spline curves, which are not as accurate. The B-spline curve is excellent in comprehensive evaluation of path length and average curvature, so the invention selects the B-spline curve. FIG. 9 is a multi-objective evaluation function-based screening curve formed by using the path length and the average curvature, wherein each channel-changing path is firstly constructed to be capable of being simulated, and then the optimal channel-changing path is screened out by solving the evaluation function value of each channel-changing path.

The control module of this embodiment employs MPC, i.e., predictive model control. Fig. 10, 11 and 12 are control simulation implementation diagrams of the unmanned control system on the system simulation expressway and S-T curves thereof, wherein two cars simulate the historical track of a traffic vehicle operated by a human driver, the car is a vehicle operated by the unmanned control system, and the curve in front of the car is a vehicle travel route predicted by a prediction module. Fig. 10 shows that at the first moment, the predicted travel route (the cross line) of the host vehicle and the saw-tooth obstacle region (the predicted travel region of the traffic vehicle) do not intersect, so that the speed (the dotted line) of the host vehicle is a smooth ascending curve, and the host vehicle is in an accelerating state, as can be seen from the following S-T curve. At this time, the speed of the rear traffic vehicle is obviously faster than that of the front traffic vehicle, the predicted track can show that the rear traffic vehicle is about to exceed the front traffic vehicle, the running route of the rear traffic vehicle is likely to influence the running route of the vehicle, and at this time, the control module controls the vehicle to normally run. Fig. 11 shows a second key time after three vehicles travel for a period of time, when the vehicle reaches the vicinity of the rear-side traffic vehicle, the lower S-T curve shows that the predicted travel route (cross line) of the vehicle and the saw-tooth obstacle area (predicted travel area) have obvious intersection states, and the travel track of the left-side traffic vehicle obviously covers the travel route of the vehicle. The prediction module then classifies the left-hand traffic psychology tendency as a litaxel tendency. When the right side is observed to pass through, the speed of the left side traffic vehicle is obviously in a descending state and does not occupy the front route of the vehicle for a long time, if the vehicle is decelerated at the moment, the high speed is in a parallel state of double vehicles, and accidents such as rear-end collision and the like can be possibly caused, so the control module of the invention controls the vehicle to continue to be in a driving stage before lane changing and does not perform deceleration treatment. Fig. 12 shows a third key time after three vehicles travel for a period of time, when the vehicle is traveling near the right lane traffic vehicle, the vehicle has already planned a local lane change path through the planning module, the front curve of the vehicle is the planned lane change path, and the lower S-T curve can show that the predicted traveling route (cross line) of the vehicle and the saw-tooth obstacle area (traffic vehicle predicted traveling area) recover to have no intersection state at this time, so that the vehicle speed (dotted line) is a smooth ascending curve, and the vehicle is in an accelerating traveling state. The control module controls the vehicle to finish the lane change behavior, finish the overtaking and continuously keep the speed to run at the high speed until the end point of the simulated high-speed road section.

The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, and alternatives falling within the spirit and principles of the invention.

Claims (4)

1. Automatic driving intelligent control system based on SVO theory, its characterized in that: integrating a socio-psychological tool (SVO) into the control of an autonomous car comprises the steps of:

step 1: completing structured data acquisition and semi-structured data analysis of a highway driving data set, and extracting structured data to simulate highway conditions; constructing S-shaped roads with different curvature radiuses by utilizing a Matlab automatic driving tool box, acquiring left and right boundaries of the S-shaped roads, and simulating 2 historical traffic vehicles and driving paths thereof;

step 2: training an automatic driving simulation LSTM neural network model by utilizing a Matlab automatic driving simulation data set, and capturing information of motion characteristics and hidden states of surrounding vehicles in real time during automatic driving simulation by utilizing an LSTM network by the model; the model carries out common reasoning among a plurality of automobiles by introducing a 'society' pooling layer, communities of the LSTM neural network receive society and information pooling of hidden state information from communities of adjacent LSTM neural networks in each period of time, and track prediction of the unmanned automobile is completed by utilizing the society and information pooled LSTM networks;

step 3: generating a lane change decision of the unmanned vehicle based on the fuzzy logic according to the prediction result and the current driving scene; then, planning a global path based on an optimization principle;

step 4: optimizing the transverse path planning of the model by adopting a B-SPLINE (B SPLINE) curve algorithm, and optimizing the longitudinal speed planning of the model by adopting a dynamic planning algorithm; the introduction parameters complete the relevant control: longitudinal control, namely realizing speed control of the automobile by controlling the automobile brake/accelerator, and transverse control, namely realizing course control by adjusting the angle of a steering wheel, and realizing Model Predictive Control (MPC);

the social and information pooling LSTM network abstracts the intelligent relationship between the time domain space and the space domain space into a low-dimensional quantifiable embedded LSTM network; capturing time information of nodes and edges into an evolution graph;

each track of the same scene in the automatic driving simulation LSTM neural network model is provided with an independent LSTM network; the LSTM then interconnects through a social pool (social and information pool, S-pool) layer; the pooling layer allows spatially close LSTM networks to share information with each other; characterizing the strength of interaction between human drivers by averaging, weighting and graph evolution message or graph messaging aggregation operations; pooling operation is to embed information independently or simultaneously in different neural network structures in a low-dimensional latent state of time domain and space domain dimensions;

an autopilot algorithm conforming to the Social Value Orientation (SVO) is shown in table 1;

table 1 automatic driving algorithm meeting social criteria

Human drivers, when interacting with others, create a wide variety of social preferences; different SVO psychological trends, namely Lihexose and Lisitant, are obtained through quantitative calculation of the psychological trends; SVO psychological trends measure whether a rider sets his or her priority of traveling on a road higher than other riders; calculating other driver vehicle models except the self-driving vehicle by using a non-interactive baseline algorithm to obtain an optimal strategy; modeling other drivers as dynamic driving traffic vehicles by utilizing the data set and the track history, and simultaneously comparing the performance standard obtained by the base line prediction with the standard of the SVO theoretical model to generate different static SVOs for different driving interaction behaviors: 1) Static lyxotropy SVO; 2) Static litaxenic SVO; 3) Dynamic social SVO; the optimal static SVO corresponds to an optimal SVO estimated value which is kept constant in the whole driving interaction process;

the path planning is related to the speed, and the comfort evaluation indexes influencing the riding comfort are dominant, such as the indexes of acceleration, jerk and the like; the path planning does not contain time information, so that analysis is only carried out from the path curve, and evaluation indexes of reference path length and path average curvature are established;

two path evaluation indexes are adopted: path length and path average curvature; the commonly used multi-objective optimization methods are: pareto optimization front, linear weighting method, namely multi-target single-target-to-NSGA non-dominant ordering genetic algorithm; the complexity of the problem is screened according to the road changing path;

selecting a linear weighting method;

f＝ω ₁ f _length +ω ₂ f _curvature

wherein f _length And f _curvature Respectively representing the path length and the path average curvature function constructed according to the weight coefficients;

adopting a path planning algorithm as a B-spline curve algorithm; the B-spline curve is defined as:

wherein p is _i Is the vertex of the feature polygon; b (B) _i K is called a k-order (k-1 times) basis function, and the order of the B-spline algorithm is the number of times added with 1; in the non-zero definition domain, the basis functions of K-1 are minimized in each node of the repetition degree K; in the internal nodes of each complex k, the non-zero basis function has the greatest p-k+1, where p is the number of times the basis function;

model predictive control takes model, prediction and control as core ideas;

the model selects a vehicle kinematic model or a dynamics model, and converts the vehicle kinematic model into a linear state space equation; in a scene of low-speed running of the vehicle, the kinematic model of the vehicle meets the requirements, and is relatively simple;

the vehicle dynamics model is adopted to run at a high speed, so that the vehicle dynamics model is selected and converted into a linear state space equation;

selecting a state vector, and constructing a system state equation:

wherein: m is the vehicle mass, delta is the front wheel rotation angle, v is the vehicle speed, phi is the steering speed;

the prediction is a process of obtaining a series of vehicle state quantities by continuously recursively obtaining the state quantities according to a state space equation model of the vehicle, and the state and output of the system at the future moment are expressed as a matrix:

as can be seen from the above equation, both the state and the output in the prediction domain pass through x (k): current state quantity sum deltau of system

Control is a quadratic programming problem in that the control quantity is constructed at each moment so that the cumulative objective function is optimal:

wherein h=Φ ^T Q _e Φ+R _e ,g ^T ＝2E(t)Q _e Φ,Q _e ,R _e An expansion matrix for the weight matrix Q, R; q and R are weight matrixes, and Q is more than or equal to 0 and R is satisfied>Positive definite matrix of 0.

2. The SVO theory-based intelligent control method for automatic driving according to claim 1, wherein: and an online prediction module: the basic steps of constructing an online prediction module are that a state future prediction module is used for linking an LSTM network model trained in advance, and then rolling prediction is performed 80 times at each moment point by using a for circulation structure, so that a future 8s traffic vehicle prediction track is obtained; firstly, inputting the current actual position of the traffic vehicle, converting the actual position into a standard value, inputting the standard value into a Stateful prediction module, finally obtaining the output value of the Stateful prediction module, and inversely normalizing the output value into a normal position; and an offline prediction module: in order to combine the history track and the predicted track in the cycle and then use the combined history track and the predicted track for predicting the next cycle, an offline prediction module is designed and built; firstly, carrying out standardization processing on the historical actual running track coordinates, defining a prediction time domain length and an initial prediction frame, importing a previously trained LSTM network, starting to predict tracks of 8 seconds in the future for each moment point of two traffic vehicles, storing output results into corresponding variables, finally, inversely standardizing the stored output values into actual tracks, and converting the actual tracks into time sequence variables.

3. The SVO theory-based intelligent control method for automatic driving according to claim 1, wherein: data set processing: dividing the model by adopting different training sets and verification sets, and carrying out multiple groups of different training and verification; training the LSTM network by adopting a cross-validation method to realize the core content of track prediction: firstly dividing a data set into 10 groups of data with equal length, numbering 1-10, randomly taking 1 data as a verification set and the rest 9 data as a training set, verifying the validity of a trained network by utilizing 4 groups of tracks of the data, circulating 3 groups, and finally comparing 3 groups of minimum values to confirm a training function; wherein, the network layer is set as an input layer 2 level, an output layer 2 level and an hidden layer 250 level;

deviation between predicted and actual values: l2 loss in track prediction training, along with the reduction of errors, gradient is also reduced; specific parameter setting:

table 2 trajectory prediction training important parameter settings

Actual trajectory and predicted trajectory: when track prediction is performed, rolling prediction is required to be performed continuously, and the LSTM model is updated continuously.

4. The SVO theory-based intelligent control method for automatic driving according to claim 1, wherein: the decision module adopts a fuzzy logic algorithm, the fuzzy logic is an uncertainty imitating human brain judgment and reasoning, and the fuzzy logic is often used for judging and reasoning uncertain factors; in the process of establishing an intelligent vehicle lane change decision model, the artificial intelligence utilizes a fuzzy theory to study the nonlinear relation between the vehicle running environment and lane change decision, and decides whether the vehicle changes lanes or not through accurate running environment information and parameters, and a great nonlinear relation exists between lane change intention generation and the information, wherein the nonlinear relation is difficult to determine through some accurate running environment information and parameters;

setting the lane change satisfaction degree of the vehicle, inputting a speed difference coefficient and a vehicle distance expected coefficient according to the lane change satisfaction degree, and blurring into five grades, wherein the speed difference coefficient V is selected from { small, medium, large and large } as a blurring subset, the domain is X, and the values are 0 and 1; the vehicle distance is expected to select { small, medium, large } as a fuzzy subset, and the domain is Y.