CN-115950445-B - Highway commercial vehicle track planning method integrating complex network theory and deep neural network and vehicle electronic equipment

CN115950445BCN 115950445 BCN115950445 BCN 115950445BCN-115950445-B

Abstract

The invention discloses a method for planning the track of a commercial vehicle on an expressway by combining a complex network theory and a deep neural network and vehicle electronic equipment, which are used for evaluating important vehicle nodes and extracting a risk tree through a model, the method is used for expanding the action space of the self-vehicle track planning, modeling various vehicle nodes into a field form and evaluating the collision risk of the planned track. Meanwhile, track planning is directly carried out under an XYT coordinate system in a highway scene, the predicted track of the other vehicle and frames (Bounding Box) of other obstacles are mapped into a grid map generated according to the coordinate system, and a Deep Neural Network (DNN) is utilized for convolution operation to obtain a safe and feasible collision-free track, so that the influence of different sizes of commercial vehicles and passenger vehicles on track planning is reduced, and meanwhile, the problems of difficult road modeling and overlarge calculated amount are solved. The method can be used for planning a high-quality track which can simultaneously give consideration to safety, efficiency and flexibility without converting a coordinate system in the whole intelligent driving system frame.

Inventors

FU XINKE
CAI YINGFENG
WANG HAI
CHEN LONG
JIANG TAO
DONG ZHAOZHI
LIU QINGCHAO

Assignees

江苏大学

Dates

Publication Date: 20260512
Application Date: 20230109

Claims (6)

1. The highway commercial vehicle track planning method integrating the complex network theory and the deep neural network is characterized by comprising the following steps: step 1, dynamically modeling a driving environment based on a cognitive theory of a complex network; step 2, generating a complex network according to the model built in the step 1, evaluating important nodes and extracting a risk tree; Mapping Bounding Box of the predicted track of the other vehicle and other obstacles into a grid map generated according to an XYT coordinate system, wherein the XYT coordinate system is a coordinate system formed by introducing a Cartesian coordinate system into a time dimension; Step 4, sampling track points on the grid map, taking Bounding Box of the sampling points as convolution kernels, carrying out convolution operation by using a deep neural network DNN to obtain a safe and feasible collision-free track, evaluating the track, and selecting an optimal track; The specific process of the step 1 comprises complex network modeling, in particular: firstly, based on a complex network theory, taking a vehicle and other vehicles as network nodes, and constructing a dynamic complex network model: Wherein, the Is a dynamic complex network model; For a set of nodes in a network, As a node in the network, Is the number of nodes in the network; For a set of edges of a node in a network, As an edge of a node in the network, The number of node edges in the network; is the weight of the edge; Is a movable area of the node; Modeling as a smooth bounded surface: Wherein, the Is the boundary of a smooth bounded curved surface; the specific process of step 1 further comprises dynamic modeling, in particular: Step 1.2, carrying out dynamic modeling on vehicle nodes based on a two-degree-of-freedom vehicle model and a motion point model; the two-degree-of-freedom model of the vehicle is as follows: Wherein the state variable of the model is yaw rate And lateral velocity ; Steering angle for front wheel; And Respectively, a front side force and a rear side force generated during driving, and a longitudinal speed Considered as a time-varying parameter, when the tire cornering characteristics are within a linear range, the model is expressed as: And The front wheel cornering stiffness and the rear wheel cornering stiffness are respectively; And Respectively the center of gravity of the front axle and a rear axle center of gravity; Is of mass; considering the vehicle nodes as moving particles with a center of gravity, and building a moving point model relative to a desired path: Wherein, the And Heading angle deviation and transverse path deviation respectively; is the distance along the desired path; curvature for the desired path; The specific process of step 1 further comprises variable gaussian security field modeling, in particular: step 1.3, a variable Gaussian safety field model based on risk center transfer is provided according to a field theory, and a static safety field is described by a two-dimensional Gaussian function: Wherein, the Is the field intensity coefficient; And As the coordinates of the center of the risk, And The radii of the major axis and the minor axis of the ellipse are respectively the radius of the major axis and the minor axis of the ellipse, the ellipse is the enlargement of the inscribed ellipse of the vehicle frame, and the ellipse is equivalently expressed by the transverse-longitudinal ratio of the vehicle; when the vehicle moves, the risk center is shifted along with the movement, and the new risk center is that : Wherein, the Is a velocity vector; Is a regulatory factor, and has Or (b) (; Is that And (3) with The axis angle, the dynamic security field can be expressed as: And Radii representing major and minor axes of the ellipse after risk center transfer; The specific process of step 1 further includes dividing the cognitive domain, specifically: Step 1.4, dividing the periphery of the node into a first cognitive domain, a second cognitive domain and an outside domain space according to the sensitivity and the response time of a human driver to the distance; The range of the first cognitive domain is: Wherein, the Is a first threshold; a first cognitive response time for a human driver; maximum approaching speed for other nodes in the environment; The range of the second cognitive domain is: Wherein, the Is a first threshold; A second cognitive response time for a human driver, the second cognitive outside-domain space being defined as an outside-domain space; within the framework of a variable gaussian security field, a risk awareness function between nodes is established: Wherein, the Is a node At the node The field strength at the location of the field, Is a node Scalar speed, direction angle of (2) Is a node Velocity vector of (2) And node Is of the field strength vector of (2) Is arranged at the lower end of the cylinder, Is a risk cognitive regulatory factor.
2. The method for planning the track of the commercial highway vehicle by fusing the complex network theory and the deep neural network according to claim 1, wherein the specific process of the step 2 comprises the following steps: Step 2.1, evolution process of complex network: 1) The method comprises the steps that a host node is set as a master node, the master node and other nodes in a first cognitive domain are connected, the weight of corresponding connecting edges is calculated, the corresponding connecting edges are ordered according to the weight, if a plurality of nodes exist in the connection direction with the master node, the nearest node is taken as a node in the first cognitive domain, other nodes with larger distances are taken as a node in a second cognitive domain, and the nodes in the second cognitive domain are treated in the same way; 2) In the first cognitive domain, sequentially selecting environment nodes according to the weight sequence determined in the last step, calculating weights of the environment nodes and other nodes, sequencing, connecting node pairs with the largest weights, if the weights are lower than a set threshold, not connecting, and if the edges exist, not repeatedly connecting; 3) Selecting an environment node in the second cognitive domain and a node in the first cognitive domain, and connecting a node pair with the largest weight; 4) In the second cognitive domain, sequentially selecting environment nodes, calculating the weights of the environment nodes relative to other nodes, sequencing, connecting node pairs with the largest weights, if the weights are lower than a set threshold, not connecting, and if the edges exist, not repeatedly connecting; step 2.2, evaluating important nodes and generating a risk tree, wherein the step comprises the following steps: 1) Set and node Aggregate with all neighboring nodes as Node strength Is a node The sum of the weights of the adjacent nodes is: 2) Defining the average value of all node intensities in the network as the network intensity The following steps are: 3) Intensity of node The ratio to the sum of the intensities of all nodes is defined as For evaluating the importance of a node, there are: Will be The largest node is defined as important node, the important node in the first cognitive domain is defined as a first type important node, and the important node in the second cognitive domain is defined as a second type important node; 4) The tree generated by the main node, the first class important node and the second class important node is defined as a risk tree.
3. The method for planning the track of the commercial highway vehicle by fusing the complex network theory and the deep neural network according to claim 1, wherein the specific process of the step 3 comprises the following steps: For a dynamic vehicle, expanding the transverse and longitudinal proportion of the variable Gaussian safety field constructed by each node in the step 1 on the original rectangular boundary of the vehicle to form a new circumscribed rectangle, defining the new circumscribed rectangle as Bounding Box, converting the new circumscribed rectangle into a probability field, representing the probability of collision of the vehicle with the related vehicle node track by field intensity, wherein the probability in the original boundary of the vehicle is 1, the probability in the area from the original boundary to Bounding Box is reduced along with the reduction of the field intensity, and finally converting the variable Gaussian safety field into the probability field; Step 3.2, setting the current time in the coordinate system as , Representing the length of time of a planning cycle, will Section prediction of other vehicle track according to Taking the form of interpolation conversion Bounding Box and projecting to In the grid map at the moment, for the vehicle nodes in the risk tree extracted in the step 2, all possible predicted tracks are input into the grid map at the same time, and for other vehicle nodes, only the predicted track with the highest probability is input into the grid map.
4. The method for planning the track of the commercial highway vehicle by fusing the complex network theory and the deep neural network according to claim 1, wherein the specific process of the step 4 comprises the following steps: 4.1, sampling on a grid map by using a Nonlinear model predictive control (Nonliean) MPC method to obtain a series of discrete tracks at equal time intervals; Sampling the obtained track Time of day according to At the time of Interpolation is internally taken to obtain Bounding Box, the result is regarded as kernel with the value of 1, DNN is utilized to carry out convolution operation on the kernel and a grid map comprising other vehicle prediction tracks and obstacle projections, if the convolution result is greater than 0, the track is regarded as collision risk, and otherwise, the track is regarded as collision risk-free, so that a series of safe and feasible collision-free discrete tracks are directly obtained under an XYT coordinate system; Step 4.2 the cost of planning the trajectory is determined by a comprehensive cost function And (3) determining: Wherein, the Is a correlation coefficient; A distance function of the track points and other vehicle predicted tracks is used for punishing tracks too close to the other vehicle predicted tracks; The distance function between the track point and the road boundary is used for punishing the track which is too close to the road boundary; the track point is a distance function between the track point and the front vehicle and is used for punishing the track which is too close to the front vehicle; the track is a distance function between the track point and the reference line, and is used for punishing the track which is far away from the reference line; Determining the weight of each part and the design of related functions through inverse reinforcement learning and imitation learning; and adding the costs of the track points at each moment of each track to obtain the total cost of the track, and selecting the track with the minimum cost as the optimal track.
5. The method for planning the commercial vehicle track of the highway by fusing the complex network theory and the deep neural network according to claim 4, wherein the specific process of the step 4 further comprises the steps of optimizing and smoothing the discrete track through a non-linear MPC, and outputting an optimal track without coordinate transformation.
6. An electronic device for a vehicle, wherein the electronic device for a vehicle is a control device or a storage device, the method of any one of claims 1 to 5 is provided in the control device, and the program of the method of any one of claims 1 to 5 is stored in the storage device.

Description

Highway commercial vehicle track planning method integrating complex network theory and deep neural network and vehicle electronic equipment Technical Field The invention belongs to the field of intelligent driving in artificial intelligence, and particularly relates to a highway commercial vehicle track planning method and vehicle electronic equipment integrating a complex network theory and a deep neural network. Background The commercial vehicle has the advantages of stronger carrying capacity, faster running speed and the like, promotes the rapid flow of logistics, people flow and information flow, and plays an irreplaceable important role in the aspects of social comprehensive transportation systems, engineering construction and the like. The expressway has the characteristics of relatively closed traffic environment, structured road, relatively simple road condition and the like, and is expected to realize large-scale application of high-level intelligent driving commercial vehicles. However, the commercial vehicle has the characteristics of large volume, heavy weight, large blind area of the driver's visual field and the like, so that the track planning of the commercial vehicle is more difficult than that of the commercial vehicle. How to find a high-quality track which can simultaneously achieve safety, efficiency and flexibility is an important subject of the track planning of the highway commercial vehicle. The track planning module generates the track driven by the vehicle by receiving and processing various perception information and predicting the tracks of other vehicles, and is one of key technologies of intelligent driving. In order to solve the problems of difficult road modeling and excessive calculation, the current mainstream track planning method generally adopts a path speed decomposition method, converts a Cartesian coordinate system (Cartesian) into a lane coordinate system (Frenet), and projects tracks predicted by obstacles and other vehicles to the Frenet coordinate system to carry out track planning of the vehicle. However, currently, the mainstream intelligent driving system generally adopts a multi-mode track prediction method to generate probabilities of multiple predicted tracks of other vehicles at the same time, and single-mode track planning needs to enable the own vehicle to avoid multiple possible tracks of other vehicles at the same time under the Frenet coordinate system, which often promotes the own vehicle to select a very conservative planned track. Meanwhile, the problems that time is more than short for alignment and the like can occur in the process of interconversion of the Cartesian coordinate system and the Frenet coordinate system, and when the road curvature is large, the quality of a track generated under the Frenet coordinate system is greatly reduced. Disclosure of Invention In order to solve the problems, the invention introduces a cognitive theory based on a complex network to model a driving environment and is used for guiding the track planning of the bicycle. The method comprises the steps of evaluating important vehicle nodes through a model, extracting a risk tree, expanding an action space of vehicle track planning, modeling various vehicle (commercial vehicle and passenger vehicle) nodes into a field form, and evaluating collision risk of a planned track. Meanwhile, track planning is directly carried out under an XYT coordinate system (a Cartesian coordinate system is introduced into a time dimension) in an expressway scene, a predicted other vehicle track and frames (Bounding Box) of other obstacles are mapped into a grid map generated according to the coordinate system, a Deep Neural Network (DNN) is utilized for carrying out convolution operation, and a safe and feasible collision-free track is obtained, so that the influence of different sizes of commercial vehicles and passenger vehicles on track planning is reduced, and meanwhile, the problems of difficulty in road modeling and overlarge calculated amount are solved. The method can be used for planning a high-quality track which can simultaneously give consideration to safety, efficiency and flexibility without converting a coordinate system in the whole intelligent driving system frame. The invention aims to provide an intelligent driving track planning method suitable for a highway commercial vehicle, which is used for modeling a driving environment by introducing a cognitive theory based on a complex network and guiding track planning of a vehicle, and simultaneously, carrying out convolution operation by utilizing a Deep Neural Network (DNN) under an XYT coordinate system to generate a safe and feasible collision-free track, evaluating the collision-free track, and finally, selecting a high-quality track which can simultaneously give consideration to safety, efficiency and flexibility. In order to achieve the above purpose, the invention adopts the following technical scheme that the m