Vehicle segmented path acceleration planning method based on neural network model

Abstract

The invention relates to a vehicle segmented path acceleration planning method based on a neural network model, and belongs to the technical field of automatic driving. The method comprises the following steps: firstly, establishing a segmented path planning model, namely acquiring vehicle and road information, determining candidate state sets of a transfer area and an end point area, generating candidate path sets of an initial section and an end section, establishing a comprehensive cost function and solving an optimal combined path; and then, through data acquisition and neural network acceleration model construction, the vehicle optimal segment path generation model is converted into a neural network model, so that the operation speed of the algorithm is remarkably improved, and real-time application is realized. The method also optimizes the selection range of transit points and candidate paths, reduces the risk of planning failure of the vehicle in a complex environment, comprehensively considers the characteristics of different segment paths through a unified cost function, screens out a global optimal path, and avoids the problem of suboptimal combined paths caused by sequential path planning.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving and relates to a vehicle segmented path acceleration planning method based on a neural network model. Background Art

[0002] In recent years, the safe and efficient driving of smart cars has become the future development trend of the automotive industry. Autonomous driving technology, which uses road traffic information and vehicle status information to improve the safety of transportation, the travel efficiency and intelligence level of vehicles, is receiving widespread attention. The core tasks that autonomous driving vehicles need to complete mainly include environmental perception and positioning, decision making and path planning, and motion control.

[0003] As one of the core functional modules of autonomous vehicles, path planning plays a bridging role between upper-level perception and lower-level control, which is of great significance. The local path generated by combining the global path and the surrounding environment information has a direct impact on the driving quality of autonomous driving. The goal of path planning is to generate a safe, comfortable and driver-friendly path for the autonomous vehicle.

[0004] Among them, vehicle segmented path planning is an excellent path planning method that can meet the path requirements of drivers and passengers at different stages. For example, the invention "A vehicle segmented path planning method based on vertex risk assessment" with Chinese patent publication number CN116817948B proposes a vehicle segmented path planning method that coordinates safety and comfort goals.

[0005] However, based on the relevant research literature on existing segmented path planning methods, including papers, patents and books, the current segmented path planning methods for autonomous driving vehicles still have the following shortcomings: it is difficult to ensure real-time performance when there are many segments and candidate paths; the use of a fixed transfer point calculation method limits the selection range of candidate paths, resulting in an increased possibility of planning failure in complex situations; and the use of a sequential selection method to calculate the optimal solution for each path segment may result in a non-optimal combined path. Summary of the invention

[0006] In view of this, the purpose of the present invention is to provide a vehicle segmented path acceleration planning method based on a neural network model. The method includes three parts: segmented path planning model establishment, data collection and acceleration model construction, and acceleration model application. Compared with the existing segmented path planning method, the present invention significantly improves the running speed of the algorithm and realizes the real-time application of segmented path planning; optimizes the selection range of transfer points and candidate paths, thereby reducing the risk of vehicle planning failure in complex environments; and comprehensively considers the characteristics of different segmented paths through a unified cost function, screens out the global optimal path, and avoids the problem of suboptimal combined paths caused by sequential planning paths.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A vehicle segmented path acceleration planning method based on a neural network model, the method comprising:

[0009] The segmented path planning model is established, including:

[0010] Obtain vehicle and road information;

[0011] Determine a set of candidate states of the transit area, and determine a set of candidate states of the terminal area;

[0012] Determine an initial candidate path set, determine a final candidate path set, and determine a combined candidate path set;

[0013] Establish comprehensive cost function and solve the optimal combination path;

[0014] Construct a segmented path planning model, with vehicle and road information as input and the optimal path as output;

[0015] Build data collection and neural network acceleration models, including:

[0016] Analyze the factors affecting the optimal path and determine the input and output of the neural network model;

[0017] Data generation and collection, collecting input and output variables to form training data sets;

[0018] The training and generation of the neural network acceleration model obtains an acceleration model whose input is the variables affecting the optimal path and whose output is the transit point and end point of the optimal path;

[0019] Applications of neural network acceleration models include:

[0020] Vehicle and road information acquisition;

[0021] Solve the optimal path and obtain the optimal transfer point and end point;

[0022] According to the vehicle and road information, the optimal transfer point and the end point information, the initial quartic polynomial path and the final quintic polynomial path are calculated and generated to obtain the optimal combined path.

[0023] Furthermore, the obtaining of vehicle and road information includes obtaining relevant information of the host vehicle, the obstacle vehicle and the road.

[0024] Further, the determining of the transfer area candidate state set includes:

[0025] When the obstacle vehicle is stationary, the closest position and the farthest position are determined by the defined boundary conditions;

[0026] Determine the closest position: When the front edge of the main vehicle and the rear edge of the obstacle vehicle are in the same straight line, and the heading of the main vehicle is parallel to the lane, its position is on the oy axis and maintains a safe distance from the obstacle vehicle;

[0027] Determine the farthest position: the rear edge of the main vehicle and the rear edge of the obstacle vehicle are in the same straight line, the heading of the main vehicle is still parallel to the lane, and a safe distance is maintained on the oy axis;

[0028] With the right rear corner of the vehicle as the rotation center, rotate counterclockwise at a preset interval angle, with a maximum rotation angle of 75 degrees, and record the total rotation angle;

[0029] Starting from the right rear corner of the main vehicle, move the main vehicle forward along the road direction, set the moving interval to a fixed value, and the maximum moving position is the farthest position. Record the total number of moving positions.

[0030] For any rotation and movement, the actual position of the center of mass of the main vehicle is uniquely determined;

[0031] The candidate state set of the transit area is expressed as:

[0032]

[0033] Among them, N x is the total number of mobile positions in the transit area, N θ is the total number of rotation angles, the candidate state of the transit area for the i-th rotation and j-th movement S(i,j)=(P(i,j),θ i ), S(N x ,N θ ) indicates the Nth x The Nth rotation and θ The candidate status of the transit area for the next move.

[0034] Further, determining the candidate state set of the end area includes:

[0035] Determine the nearest and farthest positions;

[0036] Starting from the nearest position, the main vehicle is moved forward along the road direction at fixed intervals until the farthest position, and the total number of moved positions is recorded as;

[0037] For any movement, the state of the main vehicle is uniquely determined;

[0038] The candidate state set of the destination area is expressed as:

[0039] S′={S′(1),S′(2),…,S′(N′ x )}

[0040] Among them, N′ xis the total number of moving positions of the destination area, and the candidate state of the destination area of ​​the kth move S′(k)=x′ k , S′(N′ x ) represents the Nth x The candidate states of the destination area of ​​the next move.

[0041] Further, determining the initial candidate path set includes:

[0042] For a candidate state in the transfer area candidate state set, a quartic polynomial path curve is established from the initial state of the main vehicle to the candidate state;

[0043] Assume the coefficient function of the quartic polynomial is, solve the boundary conditions to get the quartic polynomial function;

[0044] The initial candidate path set is expressed as:

[0045]

[0046] Among them, the initial candidate path function of the i-th rotation and j-th movement is 1f i,j , Nth x The Nth rotation and θ The initial candidate path function of the move is

[0047] Further, determining the terminal segment candidate path set includes:

[0048] For a candidate state in the transfer area candidate state set, a quintic polynomial path curve is established from the candidate state to a candidate state in the terminal area;

[0049] Assume the coefficient function of the quintic polynomial is, solve the boundary conditions to get the quintic polynomial function;

[0050] The terminal candidate path set is expressed as:

[0051]

[0052] Among them, the terminal candidate path function of the kth move is 2f k , N′ x The terminal candidate path function of the move is

[0053] Furthermore, the determining of the combined candidate path set includes combining the initial candidate path set and the final candidate path set to obtain a complete candidate path set.

[0054] Furthermore, the establishment of a comprehensive cost function and the solution of an optimal combination path include:

[0055] Establish road risk potential field function and obstacle vehicle risk potential field function;

[0056] Discretize the initial path of the combined candidate path, obtain the path points, and calculate the coordinates of the four corner points of the main vehicle;

[0057] Calculate the average potential field value of the initial path;

[0058] Calculate the path length of the final path;

[0059] Calculate the maximum curvature of the combined path;

[0060] Comprehensively consider safety, efficiency and comfort indicators to establish a comprehensive cost function;

[0061] Filter out the optimal combination path.

[0062] Further, the construction of the data acquisition and neural network acceleration model includes:

[0063] Analyze the influencing factors of the optimal combination path and determine the input and output of the neural network model;

[0064] Randomly generate the main vehicle's position, speed, lane number, and the obstacle vehicle's position, speed, and width;

[0065] The randomly generated parameters are input into the optimal combined path solving model to obtain the optimal transfer point and the optimal end point;

[0066] Organize and store valid data samples to obtain parameter input With target output The mapping data set is Indicates the heading angle of the main vehicle, Indicates the speed of the main vehicle, lane indicates the lane number where the main vehicle is located, is the coordinates of the left rear corner of the obstacle vehicle in the initial state, that is, the obstacle vehicle is in the right lane, or the right rear corner, that is, the obstacle vehicle is in the left lane, on the ox and oy axes, W obs Indicates the width of the obstacle vehicle, They are the coordinates of the optimal transfer point on the ox axis, the coordinates of the optimal transfer point on the oy axis, the heading angle at the optimal transfer point, and the coordinates of the optimal end point on the ox axis obtained by the optimal combined path solution model;

[0067] By input Characterized by To achieve this goal, we designed a neural network acceleration model.

[0068] Use loss function and learning rate to optimize hyperparameters to ensure the accuracy of optimal transit point and end point prediction;

[0069] Through neural network training, the input is the variables that affect the optimal path, and the output is the acceleration model of the transit points and end points of the optimal path.

[0070] Furthermore, the application of the neural network acceleration model includes:

[0071] Real-time information collection;

[0072] Steering to collision avoidance decisions;

[0073] Information is fed into the neural network model;

[0074] Path planning, calculate and generate the initial quartic polynomial path and the final quintic polynomial path, so as to obtain the optimal combined path;

[0075] Path tracing module.

[0076] The beneficial effects of the present invention are:

[0077] (1) The existing vehicle segment path planning method needs to generate multiple candidate paths for each segment and screen out the optimal path based on information such as obstacle vehicles, vehicle motion status, lanes and road geometry. When there are many segments and candidate paths, it is difficult to ensure real-time performance, which limits the application of this method in actual vehicles. The present invention pre-trains the vehicle optimal segment path generation model, eliminates the need for online comparison of multiple paths, thereby significantly accelerating the algorithm operation and realizing the real-time application of vehicle segment path planning.

[0078] (2) Traditional methods usually use a fixed transfer point calculation method, which is simple but limits the selection range of transfer points and candidate paths, resulting in an increased probability of planning failure in complex situations. When constructing the optimal segmented path generation model, the present invention comprehensively considers multiple transfer point calculation methods, optimizes the selection range of transfer points and candidate paths, and thus reduces the probability of vehicle planning failure in complex situations.

[0079] (3) The traditional method selects the optimal solution for each segment of the path in sequence and plans the next segment of the path based on the endpoint information of the previous segment. This segment-by-segment planning method cannot fully consider the mutual influence between the path segments, which may result in a non-optimal combined path. When establishing a vehicle optimal segmented path generation model, the present invention comprehensively considers the characteristics of different segmented paths through a unified cost function, selects the global optimal path, and avoids the problem of insufficient combined path optimization caused by the sequential planning method.

[0080] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0082] Figure 1 This is a framework diagram of a vehicle segmented path acceleration planning method based on a neural network model;

[0083] Figure 2 Schematic diagram for data acquisition;

[0084] Figure 3 This is a schematic diagram of the nearest location of the transfer area;

[0085] Figure 4 This is a schematic diagram of the farthest location in the transfer area;

[0086] Figure 5 This is a schematic diagram of the nearest and farthest positions of the end area;

[0087] Figure 6 Create a schematic diagram for the cost function. DETAILED DESCRIPTION

[0088] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0089] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0090] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0091] The present invention proposes a vehicle segmented path acceleration planning method based on a neural network model, and the specific steps include:

[0092] S1: Segmented path planning model establishment

[0093] S11: Vehicle and road information acquisition: Acquire relevant information of the main vehicle, obstacle vehicle and road.

[0094] S12: Determine the candidate state set of the transfer area and the terminal area: Analyze the implementation approach of the segmented path planning, construct the feasible range of the transfer point and the terminal, and establish the corresponding candidate state set.

[0095] S13: Determine the initial segment, final segment and combined candidate path sets: According to the candidate state sets of the transit area and the terminal area, based on the fourth and fifth order polynomial path boundary conditions, generate the initial segment and final segment candidate path sets respectively, and combine them to obtain a complete candidate path set.

[0096] S14: Comprehensive cost function and optimal combination path solution: By comprehensively considering evaluation indicators such as safety, efficiency and comfort, all candidate paths in the combination path set are evaluated to screen out the optimal combination path.

[0097] S15: Segmented path planning model: The above whole process from information acquisition to optimal path solution is constructed into a vehicle segmented path planning model, the input is vehicle and road information, and the output is the optimal path.

[0098] S2: Data collection and accelerated model building

[0099] S21: Analysis of factors affecting the optimal path: Analyze the variables that affect the optimal combination path and determine the input and output of the neural network model.

[0100] S22: Data generation and collection: Randomly generate the above variables, and obtain the transfer points and end points of the optimal path through the vehicle segmented path planning model. Collect input and output variables to form a training data set.

[0101] S23: Training and generation of neural network acceleration model: input the training data into the neural network to complete the acceleration model training. The input of the model is the variables that affect the optimal path, and the output is the transit point and end point of the optimal path.

[0102] S3: Application of Neural Network Acceleration Model

[0103] S31: Vehicle and road information acquisition: real-time acquisition of host vehicle, obstacle vehicle and road information.

[0104] S32: Optimal path solution: Input the vehicle and road information into the neural network acceleration model to obtain the optimal transfer point and end point. According to the vehicle and road information, the optimal transfer point and end point information, the initial quartic polynomial path and the final quintic polynomial path are calculated and generated to obtain the optimal combined path.

[0105] S33: The present invention significantly improves the real-time and accuracy of vehicle segmented path planning by converting the vehicle path planning problem into a fast solution process of a neural network, and has broad application prospects.

[0106] like Figure 1 As shown, the present invention proposes a vehicle segmented path acceleration planning method based on a neural network model, and its specific implementation steps are as follows:

[0107] It should be noted that the analysis of the present invention is based on the situation that the main vehicle is located in the right lane in the initial state, but if the main vehicle is initially located in the left lane, the processing method is exactly the same. Therefore, the situation that the main vehicle is initially located in the left lane is also included in the scope of application of the present invention.

[0108] (1) Vehicle and road information acquisition:

[0109] like Figure 2 As shown, relevant information about the main vehicle, obstacle vehicles and roads is obtained in real time, including the following:

[0110] Main vehicle information: includes the main vehicle's physical dimensions (length, width, and the geometric position of the center of mass in the vehicle), motion state (speed, heading angle, and the position of the center of mass in the global coordinate system), and lane information.

[0111] Obstacle vehicle information: including the size (width) and movement status (speed) of the obstacle vehicle.

[0112] Road information: mainly includes key parameters such as lane width.

[0113] (2) Determine the candidate state set of the transfer area:

[0114] The state of the main vehicle in the transfer area is determined by its position and heading angle (i.e., the vehicle's posture). Figure 3and Figure 4 The boundary conditions defined are determined.

[0115] Latest location: Figure 3 As shown, when the front edge of the main vehicle and the rear edge of the obstacle vehicle are located on the same straight line and the heading of the main vehicle is parallel to the lane, its position is located on the oy axis and maintains a safe distance from the obstacle vehicle.

[0116] Farthest position: Figure 4 As shown in the figure, the rear edge of the main vehicle and the rear edge of the obstacle vehicle are located in the same straight line, the heading of the main vehicle is still parallel to the lane, and a safe distance is maintained on the oy axis.

[0117] exist Figure 3 Under the defined "closest position" condition, the right rear corner of the main vehicle is selected as the rotation center, and the preset interval angle θ step Rotate counterclockwise, the maximum rotation angle is 75 degrees, and the total rotation angle is N θ .

[0118] At the same time, starting from the right rear corner of the main vehicle, the main vehicle is moved forward along the road direction, and the moving interval is set to a fixed value x step , the maximum moving position is Figure 4 At the defined "farthest position", the total number of recorded moving positions is N x .

[0119] For any i-th rotation and j-th movement, denoted as (θ i ,x j ). Where: θ i Indicates the rotation angle (i.e. the heading angle of the main vehicle), x j represents the coordinate of the right rear corner of the vehicle on the ox axis, i∈{1,2,…,N θ}, j∈{1,2,…,N x}.

[0120] θ i The calculation method is:

[0121] θ i =(i-1)θ step

[0122] x j The calculation method is:

[0123] x j =x nearest +(j-1)x step

[0124] Among them, x nearest The calculation method is:

[0125]

[0126] in, Indicates the coordinate of the left rear corner of the obstacle vehicle on the ox axis, L e Indicates the length of the main vehicle.

[0127] For any i-th rotation and j-th movement, the actual position of the center of mass of the main vehicle can be uniquely determined and calculated using the following formula:

[0128]

[0129] Among them, P(i,j) represents the center of mass position of the main vehicle after the i-th rotation and j-th movement, which is specifically expressed as l r It represents the distance from the center of mass of the main vehicle to the rear edge of the main vehicle. We Indicates the width of the main vehicle. It represents the coordinate of the left rear corner of the obstacle vehicle on the oy axis, and γ represents the safety distance.

[0130] Therefore, the candidate state set of the transit area can be expressed as:

[0131]

[0132] Among them, the candidate state of the transfer area for the i-th rotation and j-th movement is S(i,j)=(P(i,j),θ i ).

[0133] The above shows the calculation method of the candidate state set of the transfer area when the obstacle vehicle is stationary. After determination, the transfer area candidate status set can be uniquely determined. It remains unchanged during the movement and no additional calculation is required. Changes during exercise, calculated as:

[0134]

[0135] in, is the coordinate of the left rear corner of the obstacle vehicle on the ox axis in the initial state, is the coordinate of the center of mass of the main vehicle on the ox axis in the initial state, is the speed of the obstacle vehicle in the initial state, is the speed of the main vehicle in the initial state, To ensure the collision safety factor, l f It is the distance from the center of mass of the main vehicle to the front edge of the main vehicle.

[0136] The above method can be used to determine the and That is, the candidate state set of the transfer area can be uniquely determined, which is exactly the same as the case where the obstacle vehicle is stationary. Therefore, the following will no longer distinguish whether the obstacle vehicle is moving or not.

[0137] (3) Determine the candidate state set of the end area:

[0138] In the end area, the state of the main vehicle is mainly determined by its coordinate position on the ox axis, because the end point is located at the center line of the adjacent lane and the heading angle is 0 (that is, completely parallel to the lane). The range of the end area is as follows Figure 5 shown.

[0139] Closest position: When the vehicle is traveling in a straight line along the current heading angle, the coordinate of the intersection of the center of mass of the vehicle and the center line of the adjacent lane on the ox axis is marked as the closest position x′ nearest .

[0140] Farthest position: The nearest position is moved forward a distance L along the road direction. set The coordinates after.

[0141] like Figure 5 As shown, starting from the nearest position, with a fixed interval x′ step Move the vehicle forward along the road until it reaches the farthest position. The total number of positions moved is recorded as N′. x .

[0142] For any k-th move, the state of the main car is x′ k , x′ k represents the coordinate of the main vehicle on the ox axis when it moves to the center line of the adjacent lane, k∈{1,2,…,N′ x}.

[0143] x′ k The calculation method is:

[0144] x′ k = x′ nearest +(k-1)x′ step

[0145] Among them, x′ nearest The calculation method is:

[0146]

[0147] Among them, y adj is the coordinate of the center line of the adjacent lane on the oy axis, Δx′ constant Represents a constant greater than 0.

[0148] Therefore, the candidate state set of the destination area can be expressed as:

[0149] S′={S′(1),S′(2),…,S′(N′ x)}

[0150] Among them, the candidate state of the end area of ​​the kth move S′(k)=x′ k .

[0151] (4) Determine the initial candidate path set:

[0152] For a candidate state S(i,j)=(P(i,j),θ i ), establish a quartic polynomial path curve from the initial state of the main vehicle to the candidate state S(i,j). Let the coefficient function of the quartic polynomial be y=1f i,j (x), then the boundary condition of the quartic polynomial can be expressed as:

[0153]

[0154] in, is the coordinate of the center of mass of the main vehicle on the ox axis in the initial state, is the coordinate of the center of mass of the main vehicle on the oy axis in the initial state, is the heading angle of the main vehicle in the initial state, is the curvature of the vehicle in the initial state, θ i Determined by the candidate state S(i,j). Solving the above boundary conditions can obtain the quartic polynomial function 1f i,j .

[0155] Therefore, the initial candidate path set can be expressed as:

[0156]

[0157] Among them, the initial candidate path function of the i-th rotation and j-th movement is 1f i,j .

[0158] (5) Determine the final candidate path set:

[0159] For a candidate state S(i,j)=(P(i,j),θ i ), establish a path from the candidate state to a candidate state S′(k)=x′ in the end region k The fifth-order polynomial path curve. Suppose the coefficient function of the fifth-order polynomial is y = 2f k (x), then the boundary condition of the quintic polynomial can be expressed as:

[0160]

[0161] y adj =2f k (x′k )

[0162] 0=2f k ′(x′ k )

[0163] 0=2f k ″(x′ k )

[0164] Solving the above boundary conditions gives us the quintic polynomial function 2f k .

[0165] Therefore, the final candidate path set can be expressed as:

[0166]

[0167] Among them, the terminal candidate path function of the kth move is 2f k .

[0168] (6) Determine the combined candidate path set:

[0169] The total number of functions in the initial candidate path set is The total number of functions in the terminal candidate path set is N′ x Therefore, the total number of functions of the combined candidate path set F is Any combined candidate path can be expressed as f(i,j,k)={1f i,j ,2f k}.

[0170] (7) Establishment of comprehensive cost function and solution of optimal combination path:

[0171] The cost function is calculated for all combined paths in the combined candidate path set F. The cost function is established as follows: Figure 6 As shown in the figure, it mainly considers the safety, efficiency and comfort indicators of the path.

[0172] The goal of the initial path is to avoid collision between the main vehicle and the obstacle vehicle, and to prevent the main vehicle from crossing the road boundary. To this end, the road risk potential field function P is established. road and the obstacle vehicle risk potential field function P obs .

[0173] P road The calculation method is:

[0174]

[0175] Where: y is the coordinate of a point on the oy axis, y L ,y R Δy is the coordinates of the left and right edges of the vehicle on the oy axis when the main vehicle is driving along the center lines of the left and right lanes.L , Δy R Represents y L ,y R The distances from the left and right road edges are respectively. When the main vehicle vertex exceeds the road boundary, P road Set to a large value.

[0176] P obs The calculation method is:

[0177]

[0178] in, is the vehicle speed, v obs is the speed of the main vehicle, ε is a very small number used to prevent the denominator from being 0, x is the coordinate of a point on the ox axis, δ x ,δ y are the standard deviation coefficients in the horizontal and vertical directions, y el_L ,y el_R The length and width of the obstacle vehicle are the left and right vertical axis coordinates of the rectangular envelope. When the main vehicle vertex overlaps with the obstacle vehicle envelope, P obs Set to a large value. The center of the obstacle vehicle risk potential field is located at the intersection of the rear edge of the obstacle vehicle and the center axis of the vehicle. Compared with the center of the obstacle vehicle set, it is more realistic in the actual collision avoidance environment.

[0179] Since the two corner points on the left side of the main vehicle are likely to exceed the road boundary when the main vehicle changes lanes on the left, P is calculated for the two corner points on the left side of the main vehicle. road , and the two corners on the right side of the main vehicle are prone to collision with the obstacle vehicle, so P is calculated for the two corners on the right side of the main vehicle. obs For the initial path 1f of the combined candidate path f(i,j,k) i,j , discretize it to get n path points. For each path point, the coordinates of the four corner points of the main vehicle can be calculated:

[0180]

[0181] Among them, x g ,y g , are the horizontal coordinate, vertical coordinate and heading angle of each path point respectively.

[0182] First section route 1f i,j The average potential field value of is:

[0183]

[0184] Where R is the average potential field value, They represent the road potential field values ​​of the left front corner and the left rear corner, and the obstacle vehicle potential field values ​​of the right front corner and the right rear corner, respectively. n is the number of discrete path points.

[0185] Final Path 2f k The path length is:

[0186]

[0187] Wherein, L represents the length of the final path.

[0188] The maximum curvature of the combined path is:

[0189]

[0190] Among them, κ is the maximum curvature of the combined path, are the maximum curvatures of the initial path and the final path, respectively, and are calculated as follows:

[0191]

[0192]

[0193] R is the safety index, L is the efficiency index, and κ is the comfort index. Comprehensively considering R, L, and κ, a comprehensive cost function is established:

[0194]

[0195] Among them, {i opt ,j opt ,k opt} defines the optimal combined path, λ1,λ2,λ3 are the coefficients of R, L,κ.

[0196] (8) Data acquisition and neural network acceleration model training:

[0197] Analyze steps (1) to (7), when the size information of the main vehicle and the road width information are known, and assuming that the obstacle vehicle is moving at a constant speed parallel to the lane, the factors affecting the optimal combined path are:

[0198] Main car: position speed Lane number (lane - the left lane is 1, the right lane is 2).

[0199] Obstacle Car: Location speed Width (W obs ).

[0200] Since the positions of the main vehicle and the obstacle vehicle are relative, the position of the main vehicle can be fixed at (0,0) when constructing the data set; and since the moving obstacle vehicle can be treated as a stationary obstacle vehicle after conversion, the speed of the obstacle vehicle is set to 0 when constructing the data set.

[0201] The data set collection process is as follows:

[0202] ① Based on steps (1) to (7), establish the optimal combination path solution model.

[0203] ② Randomly generate the main vehicle and lane, obstacle course W obs .

[0204] ③ Input the randomly generated parameters into the optimal combined path solving model to obtain the optimal transfer point and optimal end point recommended by the model. If the calculation fails or the path is not feasible, the sample will be eliminated.

[0205] ④Organize and store valid data samples to obtain parameter input With target output The mapping dataset.

[0206] ⑤ Input Characterized by With this goal in mind, a neural network acceleration model is designed. During network training, the loss function and learning rate are used to optimize hyperparameters to ensure the accuracy of the prediction of the optimal transit point and end point.

[0207] Through neural network training, the process of generating the optimal path in similar scenarios can be accelerated, and the real-time performance of the path planning model can be improved.

[0208] (9) Application process of neural network acceleration model:

[0209] During vehicle driving, the established two-stage collision avoidance path neural network model can be used to quickly solve the optimal path. The specific application steps are as follows:

[0210] ① Real-time information collection. When the main vehicle is driving, the following information is collected in real time:

[0211] ② Steering collision avoidance decision. At the decision-making level, the current driving state and environment are evaluated based on real-time information, and a steering collision avoidance command is issued. Our algorithm requires steering collision avoidance decision information as an enabling instruction.

[0212] ③Information input into the neural network model. The real-time collected information is input into the trained neural network model for rapid reasoning to obtain the optimal transfer point and the optimal end point.

[0213] ④ Path planning. According to the information of the transfer point and the end point obtained by solving, the path is calculated using the following steps: The first path: establish a quartic polynomial from the initial state of the main vehicle to the optimal transfer point, and calculate the coefficients of the polynomial. The boundary conditions include the initial position of the main vehicle, the coordinates of the target transfer point, and other constraints.

[0214] The second path: establish a quintic polynomial from the optimal transfer point to the end point, and also calculate the coefficients of the polynomial to ensure that the path is smooth and meets the dynamic constraints.

[0215] ① Combined path generation. Combine the polynomials of the first and second segments to obtain a complete optimal combined path. This path should have good smoothness, small curvature and collision avoidance capabilities.

[0216] ② Path tracking module. The generated optimal combined path is input into the vehicle’s path tracking module. Through the control algorithm, the vehicle will track the generated path in real time to ensure safe and smooth driving.

[0217] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A vehicle segmented path acceleration planning method based on a neural network model, characterized in that: The method includes: The segmented path planning model is established, including: Obtain vehicle and road information; Determine a set of candidate states of the transit area, and determine a set of candidate states of the terminal area; Determine an initial candidate path set, determine a final candidate path set, and determine a combined candidate path set; Establish comprehensive cost function and solve the optimal combination path; Construct a segmented path planning model, with vehicle and road information as input and the optimal path as output; Build data collection and neural network acceleration models, including: Analyze the factors affecting the optimal path and determine the input and output of the neural network model; Data generation and collection, collecting input and output variables to form training data sets; The training and generation of the neural network acceleration model obtains an acceleration model whose input is the variables affecting the optimal path and whose output is the transit point and end point of the optimal path; Applications of neural network acceleration models include: Vehicle and road information acquisition; Solve the optimal path and obtain the optimal transfer point and end point; According to the vehicle and road information, the optimal transfer point and the end point information, the initial quartic polynomial path and the final quintic polynomial path are calculated and generated to obtain the optimal combined path.

2. The vehicle segmented path acceleration planning method based on a neural network model according to claim 1 is characterized in that: The obtaining of vehicle and road information includes obtaining relevant information of the host vehicle, the obstacle vehicle and the road.

3. The vehicle segmented path acceleration planning method based on a neural network model according to claim 2 is characterized in that: Determining the candidate state set of the transit area includes: When the obstacle vehicle is stationary, the closest position and the farthest position are determined by the defined boundary conditions; Determine the closest position: When the front edge of the main vehicle and the rear edge of the obstacle vehicle are in the same straight line, and the heading of the main vehicle is parallel to the lane, its position is on the oy axis and maintains a safe distance from the obstacle vehicle; Determine the farthest position: the rear edge of the main vehicle and the rear edge of the obstacle vehicle are in the same straight line, the heading of the main vehicle is still parallel to the lane, and a safe distance is maintained on the oy axis; With the right rear corner of the vehicle as the rotation center, rotate counterclockwise at a preset interval angle, with a maximum rotation angle of 75 degrees, and record the total rotation angle; Starting from the right rear corner of the main vehicle, move the main vehicle forward along the road direction, set the moving interval to a fixed value, and the maximum moving position is the farthest position. Record the total number of moving positions. For any rotation and movement, the actual position of the center of mass of the main vehicle is uniquely determined; The candidate state set of the transit area is expressed as: Among them, N x is the total number of mobile positions in the transit area, N θ is the total number of rotation angles, the candidate state of the transit area for the i-th rotation and j-th movement S(i,j)=(P(i,j),θ i ), S(N x ,N θ ) indicates the Nth x The Nth rotation and θ The candidate status of the transit area for the next move.

4. The vehicle segmented path acceleration planning method based on a neural network model according to claim 3 is characterized in that: Determining the candidate state set of the end area includes: Determine the nearest and farthest positions; Starting from the nearest position, the main vehicle is moved forward along the road direction at fixed intervals until the farthest position, and the total number of moved positions is recorded as; For any movement, the state of the main vehicle is uniquely determined; The candidate state set of the destination area is expressed as: S′={S′(1),S′(2),…,S′(N′ x )} Among them, N′ x is the total number of moving positions of the destination area, and the candidate state of the destination area of ​​the kth move S′(k)=x k ′,S′(N′ x ) represents the Nth x The candidate states of the destination area of ​​the next move.

5. The vehicle segmented path acceleration planning method based on a neural network model according to claim 4 is characterized in that: Determining the initial candidate path set includes: For a candidate state in the transfer area candidate state set, a quartic polynomial path curve is established from the initial state of the main vehicle to the candidate state; Assume the coefficient function of the quartic polynomial is, solve the boundary conditions to get the quartic polynomial function; The initial candidate path set is expressed as: Among them, the initial candidate path function of the i-th rotation and j-th movement is 1f i,j , Nth x The Nth rotation and θ The initial candidate path function of the move is 6. The vehicle segmented path acceleration planning method based on a neural network model according to claim 5 is characterized in that: Determining the terminal candidate path set includes: For a candidate state in the transfer area candidate state set, a quintic polynomial path curve is established from the candidate state to a candidate state in the terminal area; Assume the coefficient function of the quintic polynomial is, solve the boundary conditions to get the quintic polynomial function; The terminal candidate path set is expressed as: Among them, the terminal candidate path function of the kth move is 2f k , N′ x The terminal candidate path function of the move is 7. The vehicle segmented path acceleration planning method based on a neural network model according to claim 6 is characterized in that: The determining of the combined candidate path set includes combining the initial candidate path set and the final candidate path set to obtain a complete candidate path set.

8. The vehicle segmented path acceleration planning method based on a neural network model according to claim 7 is characterized in that: The step of establishing a comprehensive cost function and solving an optimal combination path includes: Establish road risk potential field function and obstacle vehicle risk potential field function; Discretize the initial path of the combined candidate path to obtain the path points and calculate the coordinates of the four corner points of the main vehicle; Calculate the average potential field value of the initial path; Calculate the path length of the final path; Calculate the maximum curvature of the combined path; Comprehensively consider safety, efficiency and comfort indicators to establish a comprehensive cost function; Filter out the optimal combination path.

9. The vehicle segmented path acceleration planning method based on a neural network model according to claim 8, characterized in that: The construction of the data acquisition and neural network acceleration model includes: Analyze the influencing factors of the optimal combination path and determine the input and output of the neural network model; Randomly generate the main vehicle's position, speed, lane number, and the obstacle vehicle's position, speed, and width; The randomly generated parameters are input into the optimal combined path solving model to obtain the optimal transfer point and the optimal end point; Organize and store valid data samples to obtain parameter input With target output The mapping data set is Indicates the heading angle of the main vehicle, Indicates the speed of the main vehicle, lane indicates the lane number where the main vehicle is located, is the coordinates of the left rear corner of the obstacle vehicle in the initial state, that is, the obstacle vehicle is in the right lane, or the right rear corner, that is, the obstacle vehicle is in the left lane, on the ox and oy axes, W obs Indicates the width of the obstacle vehicle, They are the coordinates of the optimal transfer point on the ox axis, the coordinates of the optimal transfer point on the oy axis, the heading angle at the optimal transfer point, and the coordinates of the optimal end point on the ox axis obtained by the optimal combined path solution model; By input Characterized by To achieve this goal, we designed a neural network acceleration model. Use loss function and learning rate to optimize hyperparameters to ensure the accuracy of optimal transit point and end point prediction; Through neural network training, the input is the variables that affect the optimal path, and the output is the acceleration model of the transit points and end points of the optimal path.

10. The vehicle segmented path acceleration planning method based on a neural network model according to claim 9, characterized in that: Applications of the neural network acceleration model include: Real-time information collection; Steering to collision avoidance decisions; Information is fed into the neural network model; Path planning, calculate and generate the initial quartic polynomial path and the final quintic polynomial path, so as to obtain the optimal combined path; Path tracing module.

Images