Information input method and device of intelligent vehicle decision-making system in complex scene

Abstract

The invention provides an information input method of an intelligent vehicle decision-making system in a complex scene, and the method comprises the steps: packaging the state information of a vehicle and surrounding traffic participants into a multi-dimensional feature vector, and achieving the numerical representation of the vehicle and environment states; integrating the feature vectors into a state matrix with a fixed dimension, and carrying out sorting and filling according to the relative positions and the number; dynamic and static environment elements around a vehicle are described by constructing a local aerial view, and the dynamic and static environment elements comprise road boundaries, lane lines, expected lanes and surrounding traffic participants; performing feature coding and dimension raising processing on the numerical state characterization and the risk field state characterization to serve as information input of intelligent vehicle behavior decision making; according to the method, the numerical state and the risk field state are effectively integrated, comprehensive and accurate information input is provided for behavior decision-making of the intelligent vehicle in a complex scene, and the decision-making capability and safety of the intelligent vehicle are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent driving technology, and in particular to an information input method and device for an intelligent vehicle decision-making system in complex scenarios. Background Art

[0002] In the field of intelligent vehicle technology, the decision-making system is the core component of autonomous driving, and its performance is directly related to the safety and efficiency of vehicle driving. In the face of complex traffic environments, how to fully and accurately perceive and understand the surrounding environment and make reasonable behavioral decisions is an important topic in current intelligent vehicle research.

[0003] Traditional intelligent vehicle decision-making systems often rely on simple environmental models, which can usually only capture limited traffic elements and have difficulty accurately describing the dynamic relationship between the vehicle and the surrounding environment. This limitation causes the decision-making system to be unable to make accurate and timely decisions when dealing with complex scenarios, thus affecting the driving safety of the vehicle.

[0004] In addition, when constructing the input of the intelligent vehicle decision-making system, how to reasonably represent and process information such as road boundaries, lane lines, desired lanes, and surrounding traffic participants is also a key factor affecting the performance of the decision-making system. Traditional representation methods are often too simple and cannot fully reflect the complexity and diversity of these traffic elements. Therefore, it is necessary to explore more refined modeling methods to improve the accuracy of environmental perception and the effectiveness of decision-making. Summary of the invention

[0005] The purpose of the present invention is to address the problems existing in the prior art and to provide an information input method and device for an intelligent vehicle decision-making system in complex scenarios, which effectively integrates the two representations of numerical state and risk field state, and provides a relatively complete information input for the behavioral decision-making of intelligent vehicles, thereby improving the decision-making ability and safety of intelligent vehicles in complex scenarios.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] An information input method for an intelligent vehicle decision-making system in a complex scenario comprises: encapsulating the state information of an ego vehicle and surrounding traffic participants into multi-dimensional feature vectors respectively, and representing the state of the vehicle and the environment in numerical form; integrating the constructed feature vectors into a state matrix of a fixed dimension, and sorting and filling them according to the relative positions and numbers of the ego vehicle and surrounding traffic participants; describing the dynamic and static environmental elements around the ego vehicle by constructing a local bird's-eye view, wherein the state representation in the form of the local bird's-eye view includes a road boundary field, a lane line field, a desired lane field, and a surrounding traffic participant field; and performing feature encoding and dimensionality upgrading on the numerical state representation and the risk field state representation as information input for intelligent vehicle behavior decision-making.

[0008] The status information of the vehicle and the surrounding traffic participants are respectively encapsulated into 9-dimensional feature vectors, which specifically include: existence, used to indicate whether the vehicle or traffic participant exists in the current scene; traffic participant type, used to indicate the type of vehicle or traffic participant; lane number, used to indicate the lane position of the vehicle or traffic participant; longitudinal distance from the vehicle, used to indicate the relative distance between the vehicle or traffic participant and the vehicle in the longitudinal direction; lateral distance from the vehicle, used to indicate the relative distance between the vehicle or traffic participant and the vehicle in the lateral direction; cosine value of the heading angle, used to indicate the cosine value of the heading angle of the vehicle or traffic participant; sine value of the heading angle, used to indicate the sine value of the heading angle of the vehicle or traffic participant; longitudinal speed of the vehicle, used to indicate the longitudinal speed of the vehicle or traffic participant; lateral speed of the vehicle, used to indicate the lateral speed of the vehicle or traffic participant.

[0009] The method integrates the constructed feature vectors into a state matrix of fixed dimension, and sorts and fills them according to the relative positions and numbers of the vehicle and surrounding traffic participants, including: filling the feature vectors into a numerical state matrix of fixed dimension, wherein the first row is used to fill in the state information of the vehicle, and the state vectors of the surrounding traffic participants existing in the current frame are filled in from the second row; when the number of traffic participants in the decision-making interest area is less than a predetermined value n, the remaining rows at the bottom of the state matrix are filled with 0; if the number of traffic participants exceeds the predetermined value n, they are sorted according to their distance from the vehicle, and only the state vectors of the n traffic participants closest to the vehicle are filled into the state matrix.

[0010] The road boundary field is modeled in the form of an exponential function with an adjustment factor and a scaling factor, as shown in formula (1):

[0011]

[0012] in, is the adjustment factor used to control the amplitude of the road boundary field, d is the horizontal coordinate in the Frenet coordinate system, and d center and d right They represent the horizontal coordinates of the road centerline and the road boundary respectively, and c1 is a positive constant used to adjust the rate at which the field strength changes with distance.

[0013] The lane line field uses a one-dimensional Gaussian distribution to generate a single lane line field, and superimposes them into a lane line synthetic field;

[0014] The calculation method of a single lane line is shown in formula (2):

[0015]

[0016] in, is the intensity adjustment coefficient, which is used to control the influence of the lane line field on the overall risk field, σ cl represents the standard deviation of the Gaussian function, which is used to adjust the rate at which the lane line field intensity changes with distance; d c，i Indicates the center position of the i-th lane line;

[0017] The strength of the lane line field C l is the field strength C of all single lane lines l，i The accumulation of is calculated as shown in formula (3):

[0018]

[0019] Among them, N lane -1 indicates the number of lane lines.

[0020] The expected lane field is modeled using a floor function, and the calculation process is shown in formula (4):

[0021]

[0022] Among them, d target represents the lateral position of the desired lane centerline, d l is the lane width, and It is an index used to adjust the field contribution.

[0023] The surrounding traffic participant field is modeled using the traffic participant characteristic axis model, and the risk field of each traffic participant is divided into the rear risk field transition area, the rear risk field dangerous area, the risk field inviolable area, the front risk field dangerous area, and the front risk field transition area;

[0024] When the position of the traffic participant is in front of the vehicle, the range of the dangerous area of ​​the risk field in front of it is calculated by the following formula (5); when the position of the traffic participant is behind the vehicle, the range of the dangerous area of ​​the risk field behind it is calculated by the following formula (6);

[0025]

[0026] Among them, s ob refers to the position of the traffic participant, and s e is the position of the vehicle, c2 and c3 are the adjustable parameters of the model, and v ob is the longitudinal speed of the traffic participant, TP ob The traffic participant type defined in Table 2;

[0027] When the position of the traffic participant is in front of the vehicle, the range of the dangerous area of ​​the risk field in front of it is calculated by the following formula (5); when the position of the traffic participant is behind the vehicle, the range of the dangerous area of ​​the risk field behind it is calculated by the following formula (6).

[0028] The risk field model of traffic participants is modeled using a one-dimensional Gaussian distribution within the range of the front and rear risk field danger areas and the risk field inviolable area, as shown in formula (7):

[0029]

[0030] Among them, A ob is the amplitude adjustment parameter of the risk field, c4 is the adjustable parameter of the model, d ob is the horizontal coordinate of the traffic participant in the Frenet coordinate system;

[0031] The risk field transition area behind the traffic participant is modeled using a two-dimensional Gaussian distribution model. The left half of the expected two-dimensional Gaussian distribution is used for risk field transition modeling. The specific mathematical expression is shown in formula (8), where the parameter K of the two-dimensional Gaussian distribution is determined by formula (9);

[0032]

[0033] Among them, d e is the horizontal coordinate of the vehicle in the Frenet coordinate system, s obr is the ordinate of the boundary of the dangerous area of ​​the risk field behind the traffic participant;

[0034] The risk field transition area in front of the traffic participant is modeled. The modeling process is similar to that of the risk field transition area behind the traffic participant. The specific mathematical expression is shown in formula (10), where the parameter K of the two-dimensional Gaussian distribution is determined by formula (11);

[0035]

[0036] Among them, s obf It is the ordinate of the boundary of the dangerous area of ​​the risk field in front of the traffic participant.

[0037] The road boundary field, lane line field, desired lane field and surrounding traffic participant field are weighted and superimposed to form a complete local bird's-eye view state representation, as shown in formula (12):

[0038] C=c7*C b +c8*C tl +c9*C l +c 10 *C ob (12);

[0039] Among them, c7, c8, c9 and c 10 Represent the road boundary field C b 、Expected lane field C tl 、Lane line field C l and surrounding traffic participants field C ob The weight factor of .

[0040] An information input device for an intelligent vehicle decision system in complex scenarios, comprising:

[0041] The state information encapsulation module is used to encapsulate the state information of the vehicle and surrounding traffic participants into multi-dimensional feature vectors, and represent the state of the vehicle and the environment in a numerical form;

[0042] A state matrix construction module is used to integrate the constructed feature vectors into a state matrix of fixed dimension, sorting and filling it according to the relative position and number of the ego vehicle and surrounding traffic participants;

[0043] A local bird's-eye view construction module is used to construct a local bird's-eye view to describe the dynamic and static environmental elements around the vehicle. The state representation in the form of a local bird's-eye view includes the road boundary field, lane line field, desired lane field, and surrounding traffic participant field.

[0044] The feature encoding and dimensionality upgrading processing module is used to perform feature encoding and dimensionality upgrading on the numerical state representation and risk field state representation as information input for intelligent vehicle behavior decision-making.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The numerical state space can accurately capture the dynamic quantitative relationship between the main vehicle and other surrounding traffic participants, while the local bird's-eye view state space can effectively capture the spatiotemporal relationship between the main vehicle and the surrounding dynamic and static traffic elements. These two state representation inputs provide a comprehensive, dynamic and accurate representation of the surrounding environment for intelligent vehicles. By comprehensively considering the type, speed, position of traffic participants and static factors of the road environment, it provides a relatively complete information input for the behavior decision-making of intelligent vehicles, thereby improving the decision-making ability and safety of intelligent vehicles in complex scenarios;

[0047] 2. By introducing parameters such as adjustment factor, scaling factor, and intensity adjustment coefficient, the model can be more flexibly adapted to different road environments and traffic conditions, thus improving the accuracy and reliability of the model;

[0048] 3. The surrounding traffic participant field is modeled using the traffic participant characteristic axis model, which divides the risk field of each traffic participant into multiple areas and takes into account the dynamic characteristics and relative position relationship of traffic participants, which more accurately reflects the complexity of the traffic environment;

[0049] 4. The road boundary field, lane line field, desired lane field and surrounding traffic participant field are weighted and superimposed to form a complete local bird's-eye view state representation, which can provide comprehensive and accurate information support for the behavioral decision-making of intelligent vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 A vehicle characteristic axis diagram in one embodiment of the present application;

[0052] Figure 2 This is a 3D visualization result diagram of the risk field of surrounding dynamic and static elements based on the risk field model in one embodiment of the present application. DETAILED DESCRIPTION

[0053] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] In a first aspect of the present application, a method for inputting information to an intelligent vehicle decision-making system in a complex scenario is provided, comprising the following steps:

[0055] S1. Encapsulate the status information of the vehicle and surrounding traffic participants into multi-dimensional feature vectors to represent the status of the vehicle and the environment in numerical form.

[0056] S2. Integrate the constructed feature vectors into a state matrix of fixed dimension, sort and fill it according to the relative position and number of the ego vehicle and surrounding traffic participants.

[0057] S3. Describe the dynamic and static environmental elements around the vehicle by constructing a local bird's-eye view. The state representation in the form of the local bird's-eye view includes the road boundary field, lane line field, desired lane field, and surrounding traffic participant field.

[0058] S4. Perform feature encoding and dimensionality upgrading on the numerical state representation and risk field state representation as information input for intelligent vehicle behavior decision-making.

[0059] Specifically, this embodiment encapsulates the state information of the vehicle and surrounding traffic participants into a multi-dimensional feature vector and integrates it into a fixed-dimensional state matrix to achieve a numerical representation of the vehicle and environmental state. At the same time, a local bird's-eye view is constructed to describe the dynamic and static environmental elements around the vehicle, including road boundaries, lane lines, desired lanes, and surrounding traffic participants. Finally, the numerical state representation and risk field state representation are feature encoded and dimensionally upgraded as information input for intelligent vehicle behavior decision-making.

[0060] The numerical state space can accurately capture the dynamic quantitative relationship between the main vehicle and other surrounding traffic participants, while the local bird's-eye view state space can effectively capture the spatiotemporal relationship between the main vehicle and the surrounding dynamic and static traffic elements. These two state representation inputs provide a comprehensive, dynamic and accurate representation of the surrounding environment for intelligent vehicles. By comprehensively considering the type, speed, position of traffic participants and static factors of the road environment, it provides a more complete information input for the behavioral decision-making of intelligent vehicles, thereby improving the decision-making ability and safety of intelligent vehicles in complex scenarios.

[0061] In some embodiments, the state information of the vehicle and surrounding traffic participants is respectively encapsulated into 9-dimensional feature vectors, as shown in Table 1 below:

[0062]

[0063] Table 1 State vectors of the ego vehicle and surrounding vehicles

[0064] The 9-dimensional feature vector specifically includes: existence, used to indicate whether the vehicle or traffic participant exists in the current scene; traffic participant type, used to indicate the type of vehicle or traffic participant; lane number, used to indicate the lane position of the vehicle or traffic participant; longitudinal distance from the vehicle, used to indicate the relative distance between the vehicle or traffic participant and the vehicle in the longitudinal direction; lateral distance from the vehicle, used to indicate the relative distance between the vehicle or traffic participant and the vehicle in the lateral direction; cosine value of the heading angle, used to indicate the cosine value of the heading angle of the vehicle or traffic participant; sine value of the heading angle, used to indicate the sine value of the heading angle of the vehicle or traffic participant; longitudinal speed of the vehicle, used to indicate the longitudinal speed of the vehicle or traffic participant; lateral speed of the vehicle, used to indicate the lateral speed of the vehicle or traffic participant. Through the design of the 9-dimensional feature vector, the dynamic and static relationship between the vehicle and the surrounding traffic participants can be fully captured, the accuracy of environmental perception is improved, and more abundant and comprehensive information input is provided for the behavior decision of intelligent vehicles.

[0065] In some embodiments, the constructed feature vectors are integrated into a state matrix of fixed dimension, and are sorted and filled according to the relative position and number of the own vehicle and the surrounding traffic participants, including: filling the feature vectors into a fixed dimension numerical state matrix, wherein the first row is used to fill in the state information of the own vehicle, and the state vectors of the surrounding traffic participants existing in the current frame are filled in sequence starting from the second row; when the number of traffic participants in the decision interest area is less than a predetermined value of 14, the remaining rows at the bottom of the state matrix are filled with 0; if the number of traffic participants exceeds the predetermined value of 14, they are sorted according to their distance from the own vehicle, and only the state vectors of the 14 traffic participants closest to the own vehicle are filled into the state matrix.

[0066] Specifically, this embodiment provides a structured way to organize and represent the state information of the vehicle and surrounding traffic participants. Through a fixed-dimensional state matrix, the information input into the intelligent vehicle decision-making system is ensured to be consistent. At the same time, the restriction and sorting of the number of traffic participants can highlight the traffic participants that have the greatest impact on the vehicle's decision, thereby improving the effectiveness and efficiency of the decision.

[0067] In some embodiments, the road boundary field is modeled in the form of an exponential function with an adjustment factor and a scaling factor, as shown in formula (1):

[0068]

[0069] in, is the adjustment factor used to control the amplitude of the road boundary field, d is the horizontal coordinate in the Frenet coordinate system, and d center and d right They represent the horizontal coordinates of the road centerline and the road boundary respectively, and c1 is a positive constant used to adjust the rate at which the field strength changes with distance.

[0070] The amplitude and rate of change of the road boundary field can be flexibly adjusted by adjusting factors and scaling factors to adapt to different road environments and traffic conditions. Modeling in the form of exponential functions can better simulate the nonlinear characteristics of the impact of road boundaries on vehicles and improve the accuracy and reliability of the model.

[0071] In some embodiments, the lane line field uses a one-dimensional Gaussian distribution to generate a single lane line field, and superimposes the fields into a lane line composite field;

[0072] The calculation method of a single lane line is shown in formula (2):

[0073]

[0074] in, is the intensity adjustment coefficient, which is used to control the influence of the lane line field on the overall risk field, σ cl represents the standard deviation of the Gaussian function, which is used to adjust the rate at which the lane line field intensity changes with distance; d c，i Indicates the center position of the i-th lane line, which determines the peak position of the lane line field strength.

[0075] The strength of the lane line field C l is the field strength C of all single lane lines l，i The accumulation of is calculated as shown in formula (3):

[0076]

[0077] Among them, N lane -1 indicates the number of lane lines.

[0078] Specifically, in order to obtain the entire lane line composite field, the field strengths of all single lane lines need to be accumulated. The accumulation process is achieved by adding the field strength values ​​of each lane line at the corresponding position. The final lane line composite field strength can reflect the comprehensive impact of all lane lines around the vehicle on the vehicle's behavior decision. The Gaussian distribution is used to simulate the impact of lane lines on vehicles. The introduction of intensity adjustment coefficients and standard deviations enables the model to adapt more flexibly to different road environments and traffic conditions. The calculation of the lane line composite field further considers the comprehensive impact of multiple lane lines on vehicle behavior decisions, improving the accuracy and reliability of the model.

[0079] In some embodiments, the desired lane field is modeled using a floor function, and the calculation process is shown in formula (4):

[0080]

[0081] Among them, d target represents the lateral position of the desired lane centerline, which determines the spatial position of the desired lane field. l is the lane width, affecting the extent of the desired lane field, and It is an index used to adjust the field contribution, which determines the contribution of the expected lane field to the overall risk field.

[0082] Specifically, by rounding down the function, the lateral position of the centerline of the desired lane can be mapped to a discrete field strength value, thereby obtaining the distribution of the desired lane field. This modeling method provides an efficient way to simulate the impact of the desired lane on vehicle behavior decisions. At the same time, by adjusting the field contribution index, the weight of the desired lane field in the overall risk field can be flexibly controlled, thereby achieving accurate modeling of different driving scenarios and desired lane selection.

[0083] In some embodiments, the surrounding traffic participant field is modeled using a traffic participant characteristic axis model, and the risk field of each traffic participant is divided into a rear risk field transition area, a rear risk field dangerous area, a risk field inviolable area, a front risk field dangerous area, and a front risk field transition area, such as Figure 1 shown.

[0084] This modeling method provides a refined approach to describe the impact of surrounding traffic participants on the vehicle's behavioral decisions. By dividing the risk field into different areas and considering the dynamic characteristics and relative position relationships of traffic participants, the model can more accurately reflect the complexity of the traffic environment and provide more comprehensive and accurate information for the behavioral decisions of intelligent vehicles.

[0085] When the position of the traffic participant is in front of the vehicle, the range of the dangerous area of ​​the risk field in front of it is calculated by the following formula (5); when the position of the traffic participant is behind the vehicle, the range of the dangerous area of ​​the risk field behind it is calculated by the following formula (6):

[0086]

[0087] Among them, s ob refers to the position of the traffic participant, and s e is the position of the vehicle, c2 and c3 are the adjustable parameters of the model, and v ob is the longitudinal speed of the traffic participant, TP ob is the type of traffic participant defined in Table 2 below.

[0088]

[0089] Table 2 Definition of traffic participant types

[0090] The calculation of these two formulas takes into account the dynamic characteristics and relative position relationship of traffic participants, so that the model can more accurately reflect the impact of traffic participants on the vehicle's behavior decision. By adjusting the model parameters, it can adapt to different traffic environments and driving scenarios, improving the flexibility and applicability of the model.

[0091] In some embodiments, a one-dimensional Gaussian distribution is used to model the risk field model of traffic participants within the range of the front and rear risk field danger areas and the risk field inviolable area, as shown in formula (7):

[0092]

[0093] Among them, A ob is the amplitude adjustment parameter of the risk field, C4 is the adjustable parameter of the model, d ob is the horizontal coordinate of the traffic participant in the Frenet coordinate system.

[0094] Specifically, this modeling method can well describe the impact of traffic participants on the risk of the vehicle in these areas as the lateral distance changes. Among them, the amplitude adjustment parameters of the risk field and the adjustable parameters of the model are used to control the shape and intensity of the risk field to adapt to different traffic environments and driving scenarios.

[0095] The risk field transition area behind the traffic participant is modeled using a two-dimensional Gaussian distribution model. The left half of the expected two-dimensional Gaussian distribution is used for risk field transition modeling. The specific mathematical expression is shown in formula (8), where the parameter K of the two-dimensional Gaussian distribution is determined by formula (9);

[0096]

[0097] Among them, d e is the horizontal coordinate of the vehicle in the Frenet coordinate system, s obr It is the ordinate of the boundary of the dangerous area of ​​the risk field behind the traffic participant.

[0098] Specifically, this modeling approach can more accurately describe how the impact of traffic participants on the risk of the ego vehicle changes with lateral and longitudinal distances in the transition area.

[0099] The risk field transition area in front of the traffic participant is modeled. The modeling process is similar to that of the risk field transition area behind the traffic participant, but the model parameters are adjusted accordingly according to the relative positions of the traffic participants. The difference is that when the traffic participant is in front of the vehicle, the effect of the risk field transition area in front is weakened, and the effect of the risk field transition area behind is enhanced. The specific mathematical expression is shown in formula (10), where the parameter K of the two-dimensional Gaussian distribution is determined by formula (11);

[0100]

[0101] Among them, S obf It is the ordinate of the boundary of the dangerous area of ​​the risk field in front of the traffic participant.

[0102] Specifically, when a traffic participant is in front of the ego vehicle, the effect of the transition zone of the risk field in front of the participant will be relatively weakened, while the effect of the transition zone of the risk field behind the participant will be correspondingly enhanced. This is because in actual driving scenarios, the impact of the traffic participant in front on the ego vehicle is usually smaller than that of the traffic participant behind the participant. In order to describe the change of this impact, a two-dimensional Gaussian distribution model is adopted, and the parameter K is adjusted. The formula (11) of the parameter K takes into account the ordinate of the boundary of the dangerous area of ​​the risk field in front of the traffic participant, ensuring that the model can accurately reflect the impact of the traffic participant on the risk of the ego vehicle at different positions.

[0103] In some embodiments, the road boundary field, the lane line field, the desired lane field, and the surrounding traffic participant field are weighted and superimposed to form a complete local bird's-eye view state representation, as shown in formula (12):

[0104] C=c7*C b +c8*C tl +c9*C l +c 10 *C ob (12);

[0105] Among them, c7, c8, c9 and c 10 Represent the road boundary field C b 、Expected lane field C tl 、Lane line field C l and surrounding traffic participants field C ob The weight factor of .

[0106] The introduction of these weight factors is to be able to flexibly adjust the combination of different fields according to their importance to the vehicle's behavior decision. By weighted superposition of these different fields, a complete local bird's-eye view state representation that integrates the road environment, lane information, and traffic participant status can be obtained, which can provide comprehensive and accurate information support for the behavior decision of intelligent vehicles.

[0107] The 3D visualization results of the risk field of surrounding dynamic and static elements based on the risk field model are as follows: Figure 2 shown.

[0108] In some embodiments, the numerical state representation and the risk field state representation are feature encoded and dimensionally upgraded as information input for intelligent vehicle behavior decision making, including:

[0109] Encode the numerical vector using a fully connected layer that maps the 9-dimensional input features to 32-dimensional output features;

[0110] A CNN with three convolutional layers is used to encode the risk field image. Each convolutional layer is followed by a ReLU activation function and a maximum pooling layer, and finally a 128-dimensional feature vector is output;

[0111] Concatenate these two feature vectors to get a 160-dimensional feature vector;

[0112] Add a fully connected layer to increase the dimension of the 160-dimensional feature vector to 256 dimensions, and add a ReLU activation function after it;

[0113] This 256-dimensional feature vector is used as the information input of the intelligent vehicle behavior decision-making system.

[0114] In a second aspect of the present application, an information input device for an intelligent vehicle decision-making system in a complex scenario is provided, comprising:

[0115] The state information encapsulation module is used to encapsulate the state information of the vehicle and surrounding traffic participants into multi-dimensional feature vectors, and to represent the state of the vehicle and the environment in a numerical form;

[0116] A state matrix construction module is used to integrate the constructed feature vectors into a state matrix of fixed dimension, sorting and filling it according to the relative position and number of the ego vehicle and surrounding traffic participants;

[0117] A local bird's-eye view construction module is used to construct a local bird's-eye view to describe the dynamic and static environmental elements around the vehicle. The state representation in the form of a local bird's-eye view includes the road boundary field, lane line field, desired lane field, and surrounding traffic participant field.

[0118] The feature encoding and dimensionality upgrading processing module is used to perform feature encoding and dimensionality upgrading on the numerical state representation and risk field state representation as information input for intelligent vehicle behavior decision-making.

[0119] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0120] In the description of the present specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish descriptions and cannot be understood as indicating or suggesting relative importance. It should also be understood that although the terms "first", "second", etc. are used to describe various elements in some embodiments of the present application in the text, these elements should not be limited by these terms. These terms are only used to distinguish one element from another element. For example, the first table can be named as the second table, and similarly, the second table can be named as the first table without departing from the scope of the various described embodiments. The first table and the second table are both tables, but they are not the same table.

[0121] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0122] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An information input method for intelligent vehicle decision-making system in complex scenarios, It is characterized in that include: The state information of the vehicle and surrounding traffic participants is encapsulated into multi-dimensional feature vectors to represent the state of the vehicle and the environment in numerical form; The constructed feature vectors are integrated into a state matrix of fixed dimension, which is sorted and filled according to the relative position and number of the ego vehicle and surrounding traffic participants; The dynamic and static environment elements around the vehicle are described by constructing a local bird's-eye view. The state representation in the form of the local bird's-eye view includes the road boundary field, lane line field, desired lane field, and surrounding traffic participant field. The numerical state representation and risk field state representation are feature encoded and dimensionally upgraded as information input for intelligent vehicle behavior decision-making.

2. According to claim 1, the information input method of the intelligent vehicle decision system in complex scenarios, It is characterized in that The state information of the vehicle and surrounding traffic participants is respectively encapsulated into a 9-dimensional feature vector, and the 9-dimensional feature vector specifically includes: Whether exists, used to indicate whether the vehicle or traffic participant exists in the current scene; Traffic participant type, used to indicate the type of vehicle or traffic participant; The lane number is used to indicate the lane position of the vehicle or traffic participant; Longitudinal distance to the ego vehicle, used to indicate the relative distance between the vehicle or traffic participant and the ego vehicle in the longitudinal direction; The lateral distance from the vehicle is used to indicate the relative distance between the vehicle or traffic participant and the vehicle in the lateral direction. The cosine value of the heading angle is used to represent the cosine value of the heading angle of a vehicle or traffic participant; The sine value of the heading angle is used to represent the sine value of the heading angle of a vehicle or traffic participant; Vehicle longitudinal speed is used to indicate the speed of a vehicle or traffic participant in the longitudinal direction; Vehicle lateral speed is used to indicate the speed of a vehicle or traffic participant in the lateral direction.

3. The information input method of the intelligent vehicle decision system in complex scenarios according to claim 1, It is characterized in that The constructed feature vectors are integrated into a state matrix of fixed dimension, and are sorted and filled according to the relative positions and numbers of the vehicle and surrounding traffic participants, including: Fill the feature vector into a fixed-dimensional numerical state matrix, where the first row is used to fill in the state information of the vehicle, and the state vectors of the surrounding traffic participants in the current frame are filled in from the second row onwards; When the number of traffic participants in the decision-making area of ​​interest is less than the preset value n, the remaining rows at the bottom of the state matrix are filled with 0; if the number of traffic participants exceeds the preset value n, they are sorted according to their distance from the vehicle, and only the state vectors of the n traffic participants closest to the vehicle are filled into the state matrix.

4. The information input method of the intelligent vehicle decision system in complex scenarios according to claim 1, It is characterized in that The road boundary field is modeled in the form of an exponential function with an adjustment factor and a scaling factor, as shown in formula (1): Among them, θ b is the adjustment factor used to control the amplitude of the road boundary field, d is the horizontal coordinate in the Frenet coordinate system, and d center and d right Represents the horizontal coordinates of the road centerline and road boundary, c 1 is a positive constant that adjusts the rate at which field strength changes with distance.

5. According to claim 1, the information input method of the intelligent vehicle decision system in complex scenarios, It is characterized in that The lane line field uses a one-dimensional Gaussian distribution to generate a single lane line field, and superimposes them into a lane line synthetic field; The calculation method of a single lane line is shown in formula (2): Among them, θ l is the intensity adjustment coefficient, which is used to control the influence of the lane line field on the overall risk field, σ cl represents the standard deviation of the Gaussian function, which is used to adjust the rate at which the lane line field intensity changes with distance; d c，i Indicates the center position of the i-th lane line; The strength of the lane line field C l is the field strength C of all single lane lines l，i The accumulation of is calculated as shown in formula (3): Among them, N lane -1 indicates the number of lane lines.

6. The information input method of the intelligent vehicle decision system in complex scenarios according to claim 1, It is characterized in that The expected lane field is modeled using a floor function, and the calculation process is shown in formula (4): Among them, d target represents the lateral position of the desired lane centerline, d l is the lane width, and θ tl It is an index used to adjust the field contribution.

7. The information input method of the intelligent vehicle decision system in complex scenarios according to claim 1, It is characterized in that The surrounding traffic participant field is modeled using the traffic participant characteristic axis model, and the risk field of each traffic participant is divided into the rear risk field transition area, the rear risk field dangerous area, the risk field inviolable area, the front risk field dangerous area, and the front risk field transition area; When the position of the traffic participant is in front of the vehicle, the range of the dangerous area of ​​the risk field in front of it is calculated by the following formula (5); when the position of the traffic participant is behind the vehicle, the range of the dangerous area of ​​the risk field behind it is calculated by the following formula (6); Among them, s ob refers to the location of traffic participants, while S e is the position of the vehicle, c 2 、c 3 is the adjustable parameter of the model, v ob is the longitudinal speed of the traffic participant, TP ob The traffic participant type defined in Table 2; When the position of the traffic participant is in front of the vehicle, the range of the dangerous area of ​​the risk field in front of it is calculated by the following formula (5); when the position of the traffic participant is behind the vehicle, the range of the dangerous area of ​​the risk field behind it is calculated by the following formula (6).

8. The information input method of the intelligent vehicle decision system in complex scenarios according to claim 7, It is characterized in that The risk field model of traffic participants is modeled using a one-dimensional Gaussian distribution within the range of the front and rear risk field danger areas and the risk field inviolable area, as shown in formula (7): Among them, A ob is the amplitude adjustment parameter of the risk field, c 4 is the adjustable parameter of the model, d ob is the horizontal coordinate of the traffic participant in the Frenet coordinate system; The risk field transition area behind the traffic participant is modeled using a two-dimensional Gaussian distribution model. The left half of the expected two-dimensional Gaussian distribution is used for risk field transition modeling. The specific mathematical expression is shown in formula (8), where the parameter K of the two-dimensional Gaussian distribution is determined by formula (9); Among them, d e is the horizontal coordinate of the vehicle in the Frenet coordinate system, s obr is the ordinate of the boundary of the dangerous area of ​​the risk field behind the traffic participant; The risk field transition area in front of the traffic participant is modeled. The modeling process is similar to that of the risk field transition area behind the traffic participant. The specific mathematical expression is shown in formula (10), where the parameter K of the two-dimensional Gaussian distribution is determined by formula (11); Among them, s obf It is the ordinate of the boundary of the dangerous area of ​​the risk field in front of the traffic participant.

9. The information input method of the intelligent vehicle decision system in complex scenarios according to claim 1, It is characterized in that The road boundary field, lane line field, desired lane field and surrounding traffic participant field are weighted and superimposed to form a complete local bird's-eye view state representation, as shown in formula (12): C=c 7 *C b +c 8 *C tl +c 9 *C l +c 10 *C ob (12); Among them, c 7 、c 8 、c 9 and c 10 Represent the road boundary field C b 、Expected lane field C tl , Lane line field C l and surrounding traffic participants field C ob The weight factor of .

10. The information input device of the intelligent vehicle decision-making system in complex scenarios according to claim 1, It is characterized in that include: The state information encapsulation module is used to encapsulate the state information of the vehicle and surrounding traffic participants into multi-dimensional feature vectors, and to represent the state of the vehicle and the environment in a numerical form; A state matrix construction module is used to integrate the constructed feature vectors into a state matrix of fixed dimension, sorting and filling it according to the relative position and number of the ego vehicle and surrounding traffic participants; A local bird's-eye view construction module is used to construct a local bird's-eye view to describe the dynamic and static environmental elements around the vehicle. The state representation in the form of a local bird's-eye view includes the road boundary field, lane line field, desired lane field, and surrounding traffic participant field. The feature encoding and dimensionality upgrading processing module is used to perform feature encoding and dimensionality upgrading on the numerical state representation and risk field state representation as information input for intelligent vehicle behavior decision-making.

Images