CN-121980151-A - Computer assessment method for complexity of autonomous running working environment of ground unmanned vehicle

CN121980151ACN 121980151 ACN121980151 ACN 121980151ACN-121980151-A

Abstract

The invention discloses a ground unmanned vehicle autonomous running working environment complexity computer evaluation method, which comprises the steps of firstly determining an environment complexity evaluation index, establishing an evaluation index system, acquiring and normalizing multi-source environment data by utilizing a vehicle-mounted sensor and a deep learning model, determining static environment index weight by utilizing an AHP method, calculating a static environment complexity index by utilizing a nonlinear mapping function, simultaneously establishing a potential energy function comprising distance, speed and predicted collision time, calculating environment total potential energy to output a dynamic environment complexity index, constructing a state evolution complexity quantization model based on state change entropy aiming at the state evolution complexity evaluation index, outputting the state evolution complexity index, and finally outputting the working environment complexity index by establishing a multidimensional coupling complexity model. The invention realizes real-time quantification and evaluation of the complexity of the working environment and provides reliable data basis for autonomous decision-making, path planning and safety control of ground unmanned vehicles.

Inventors

WANG ERLIE
LI JINGJING
OU YAYIN
WANG SAI
WANG MEICHEN
PI DAWEI
LI XIAOYIN
WANG HONGLIANG
Fan Yize
WANG XIANHUI

Assignees

南京理工大学

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (7)

1. A computer evaluation method for the complexity of an autonomous running working environment of a ground unmanned vehicle is characterized by comprising the following steps: Determining an autonomous running working environment complexity evaluation index of the ground unmanned vehicle, and dividing the autonomous running working environment complexity evaluation of the ground unmanned vehicle into a static environment complexity evaluation, a dynamic environment complexity evaluation and a state evolution complexity evaluation according to the determined index; Analyzing the image and the point cloud characteristics by using a deep learning perception model deployed in a vehicle-mounted computing unit to obtain index values of weather, roads and obstacle dimensions, simultaneously obtaining objective environment data by using a vehicle-mounted sensor, and finally carrying out data fusion processing on the index values and the objective environment data by using a normalization algorithm to output an evaluation index value vector containing static environment complexity, dynamic environment complexity and state evolution complexity; Aiming at the static environment complexity evaluation index in the evaluation index numerical vector, constructing a static environment complexity quantization model, calculating a single index risk value by establishing a nonlinear risk mapping function, determining a static environment complexity evaluation index weight by using an AHP analytic hierarchy process, and outputting a static environment complexity index by calculating the weighted sum of the risk value and the weight; aiming at the dynamic environment complexity evaluation index in the evaluation index numerical vector, constructing a dynamic environment complexity quantization model, calculating the risk potential of the monomer obstacle relative to the vehicle by establishing a dynamic potential energy function fusing the relative speed, the distance vector and the predicted collision time, accumulating and calculating all the monomer risk potential in the perception range, and outputting a dynamic environment complexity index; Step (5), constructing a state evolution complexity quantization model aiming at state evolution complexity evaluation indexes in the evaluation index numerical value vector, mapping a continuous index numerical value sequence to a discrete state space by utilizing a time sliding window, calculating state change entropy values of all indexes, weighting and summing, and outputting a state evolution complexity index; And (6) constructing a multidimensional coupled environment complexity quantization model based on the static environment complexity index, the dynamic environment complexity index and the state evolution complexity index, outputting the comprehensive environment complexity index, and realizing the complexity evaluation of the working environment.
2. The method according to claim 1, wherein the complexity evaluation index of the autonomous driving environment of the ground unmanned vehicle in step (1) specifically covers 5 evaluation dimensions and 19 objective environment indexes belonging to each dimension, and specifically comprises the following steps: The first dimension is weather condition B1 comprising a rain and snow intensity index x1, a visibility index x2, a temperature index x3, a humidity index x4, an atmospheric pressure index x5, a cloud amount index x6 and an illuminance index x7, the second dimension is road condition B2 comprising a road adhesion coefficient index x8, a road curvature index x9 and a road unevenness index x10, the third dimension is terrain condition B3 comprising an altitude index x11 and a gradient index x12, the fourth dimension is communication condition B4 comprising an electromagnetic interference degree index x13 and a signal intensity index x14, and the fifth dimension is obstacle B5 comprising an obstacle speed index x15, a relative distance index x16, an obstacle quantity index x17, a collision time index x18 and an obstacle type index x19.
3. The method according to claim 2, wherein the obtaining of the working environment complexity evaluation index value vector in step (2) specifically comprises the steps of: The method comprises the steps of (21) controlling a vehicle-mounted camera, a laser radar and a millimeter wave radar to acquire a real-time image sequence I (t) and point cloud data H (t) by a vehicle-mounted computing unit, constructing a perception input tensor D in (t) at the moment t, and simultaneously directly reading environmental physical parameters at the current moment by using a temperature sensor, a pressure sensor, an inertia measuring unit and a communication sensor to serve as objective environmental data; The method comprises the steps of (22) inputting a perception input tensor D in (t) into a deep learning perception model phi DL deployed in a vehicle-mounted computing unit, wherein the deep learning perception model phi DL is a multi-task convolutional neural network constructed based on a residual network architecture, parallel analysis and output of corresponding quantization index values of meteorological, road and obstacle characteristics are achieved through multi-scale feature extraction, and finally, data fusion is carried out on the quantization index values and objective environment data through a normalization algorithm, and an evaluation index value vector of each environment characteristic is output.
4. A method according to claim 3, wherein the construction of the static environment complexity quantization model in step (3) specifically comprises the following steps: Setting static environment complexity evaluation indexes including weather environment, road condition, topography condition and communication environment; Establishing an AHP analytic hierarchy process structural model, and determining a target layer, a criterion layer and an index layer; The vehicle-mounted computing unit reads the index relative importance preset parameters stored in the database, maps the parameters into pairwise comparison values based on 1-9 scale logic, and constructs a judgment matrix A reflecting the weight relation among all evaluation indexes; Calculating the maximum eigenvalue and corresponding eigenvector of the judgment matrix A, normalizing the eigenvector, and obtaining the weight W 'of each evaluation index through consistency test, thereby obtaining the weight vector W' formed by each evaluation index; step (35), constructing a nonlinear risk mapping function for any static index x i : , Wherein S (x i ) represents a risk value of an ith evaluation index, x i represents an ith evaluation index numerical value, x ith represents a critical threshold value of the ith evaluation index for influencing driving safety, sigma i represents a risk sensitivity coefficient of the ith index, gamma i represents a direction factor of the ith evaluation index, the method is used for unifying evaluation logic of different attribute indexes, gamma i is 1 when the method is a forward index, gamma i is-1 when the method is a reverse index, the forward index specifically comprises a rain and snow intensity index x1, a temperature index x3, a humidity index x4, an atmospheric pressure index x5, a cloud amount index x6, a road curvature index x9, a road surface unevenness index x10, an altitude index x11, a gradient index x12 and an electromagnetic interference index x13, and the reverse index specifically comprises a visibility index x2, an illuminance index x7, a road surface adhesion coefficient index x8 and a signal intensity index x14; and (36) obtaining a static environment complexity quantization model by carrying out weighted summation on the risk values of all the static environment evaluation indexes: , Where C static represents the static environment complexity index and w i ' represents the i-th assessment index weight.
5. The method according to claim 4, wherein the construction of the dynamic environment complexity quantization model in step (4) specifically comprises the following steps: Setting a dynamic environment complexity evaluation index as an obstacle dimension, and acquiring motion state parameters of an i-th dynamic obstacle in the environment by using an on-vehicle sensor, wherein the motion state parameters comprise a vehicle speed v e , a vehicle speed direction angle theta e , an i-th obstacle speed v i , a speed direction angle theta i of the i-th obstacle and a relative position vector pointing to the i-th obstacle center direction from the vehicle center ; Defining the ith obstacle risk potential energy as Q i : , wherein eta represents the gain coefficient of the type of the obstacle, the system carries out dynamic assignment according to the deep learning identification category, The relative velocity vector representing the vehicle and the obstacle i is calculated as follows: , , , In the middle of A vehicle speed vector is represented as a self-velocity vector, Representing the ith obstacle velocity vector, and further defining a direction adjustment factor for determining the speed direction of the vehicle to the obstacle: , , Wherein α i represents an included angle between a relative velocity vector and a relative position vector between the vehicle and the ith obstacle, cos (α i ) is greater than 0 represents that the obstacle is close to the vehicle, cos (α i ) is less than 0 represents that the obstacle is far away from the vehicle, λ represents an attenuation suppression coefficient, and reflects the degree of potential energy attenuation of the obstacle when the obstacle is far away from the vehicle; Finally, defining potential energy of space-time risk function G i reflecting distance between the self-vehicle and obstacle and collision time: , Wherein d 0 represents a distance potential energy smoothing factor for eliminating singular points at zero distance, TTC i represents predicted collision time of an ith obstacle and a vehicle, delta represents a tiny positive number, and singular points with zero denominator are prevented from occurring; Step (43), constructing a dynamic potential energy function through the risk intensity, the direction adjustment factor and the space-time risk function: , wherein U i represents the monomer dynamic potential energy of the ith dynamic barrier relative to the vehicle; step (44), establishing a dynamic environment complexity quantization model: , Wherein C dynamic represents the total dynamic environment complexity index, m is the total number of dynamic obstacles in the perception range, and the complexity of the dynamic environment is quantified by calculating the potential energy sum of the obstacles in the perception range.
6. The method according to claim 5, wherein the construction of the state evolution complexity quantization model in step (5) specifically comprises the following steps: Setting state evolution complexity evaluation indexes which specifically comprise five dimensions of weather environment, road condition, topography condition, communication environment and barrier, mapping the evaluation indexes into a discrete state space by a vehicle computing unit, dividing a preset interval into L state grades, and determining the state grade corresponding to each evaluation index at the current moment; And (52) calculating a state transition rate P based on the historical state grade sequence of the evaluation index in a sliding window with the preset time length of K, wherein the calculation is as follows: , Where P (i) qs represents the probability that the index i changes from state level q to state level s, and N (i) qs represents the frequency of transition of the index i from state level q to state level s; Based on the transition probability P (i) qs , defining the state change entropy of the ith evaluation index at the time t as STE i (t), wherein the state change entropy is calculated in a weighted calculation mode and is calculated as follows: , Wherein L represents the total number of state grades, when the transition probability P (i) qs is 0, the state change entropy value is 0 according to the limit value, and the state change entropy of each evaluation index is calculated to quantify the state transition severity of each index evolving along with time and reflect the mutation degree of the environmental index; Constructing a state evolution complexity quantization model: , Wherein C change represents a state evolution complexity index, and STE i represents a state change entropy value of the i-th evaluation index.
7. The method of claim 6, wherein the constructing of the multidimensional coupled environmental complexity model in step (6) comprises the steps of: Step (61), constructing a multidimensional coupled environment complexity model based on the static environment complexity index, the dynamic environment complexity index and the state evolution complexity index: , C represents a complexity index of the working environment, and the model can represent the nonlinear superposition effect of the multidimensional environment characteristics on the running complexity of the unmanned vehicle based on the probability joint distribution thought; the method comprises the steps of (62) outputting a working environment complexity index through an environment complexity model, transmitting the working environment complexity index C to a vehicle decision control unit through a bus, comparing the index with a preset risk threshold by the vehicle decision control unit, adaptively adjusting a driving strategy by a decision module according to an index grade, and uploading the index to a remote terminal to synchronously carry out visual risk warning to assist manual supervision.

Description

Computer assessment method for complexity of autonomous running working environment of ground unmanned vehicle Technical Field The invention belongs to the technical field of automatic driving, and particularly relates to a computer evaluation method for the complexity of an autonomous running working environment of a ground unmanned vehicle. Background With the rapid evolution of unmanned mobile platforms and intelligent autonomous maneuver technologies, the environment perception, behavior decision-making and motion control capabilities of ground mobile platforms in complex scenes have become key research directions. In the actual autonomous running working environment, the ground unmanned vehicle needs to face the multi-source heterogeneous environment information such as road static structural elements, dynamic traffic participant behaviors, natural weather conditions and the like. From unmanned renting service and ground cleaning operation in a structured road scene to logistics transportation and autonomous distribution tasks in an unstructured environment, the autonomous running working environment faced by the ground unmanned vehicle presents diversity and uncertainty. Therefore, how to carry out scientific, real-time and objective quantitative evaluation on the complexity of the autonomous running working environment of the ground unmanned aerial vehicle becomes a key problem for guaranteeing the operation safety and stability of the ground unmanned aerial vehicle. The Chinese patent application CN202411667369.8 discloses a technical scheme of a method for evaluating the complexity of continuous test scenes of an automatic driving automobile in real time. The method establishes a scene assessment framework containing weather, static and dynamic elements, determines the feature weights of the scene elements by using an extension hierarchical analysis method, establishes the complexity mapping relation between the scene elements and three subsystems of perception, decision and execution by calculating a potential field model and a vehicle reachable domain, and finally obtains the real-time complexity result of the continuous scene by linear weighted coupling. However, the above method still has the problem of firstly lacking quantitative characterization of the time domain evolution characteristics of the environmental state when dealing with a highly dynamic and strongly time-varying complex autonomous driving working environment. The existing method is mainly based on static scene assessment, and is difficult to effectively quantify the severity of the environment state evolving along with time, secondly, the existing assessment method is difficult to objectively quantify the dynamic environment complexity through a physical model, the assessment result is often lower than an actual complexity value, a reference basis with enough safety margin cannot be provided for a ground unmanned vehicle, and finally, the existing method is often difficult to combine off-line pre-assessment and on-line real-time assessment, so that the coverage capacity of an assessment system in the whole period of a task is insufficient. Disclosure of Invention The invention aims to provide a computer evaluation method for the complexity of an autonomous running working environment of a ground unmanned vehicle, which solves the problem that the environment state evolution characterization is lacking in the evaluation of the autonomous running working environment of the ground unmanned vehicle. The deep learning perception model is introduced to extract environmental features by constructing a multi-dimensional evaluation index system, and the state change entropy is utilized to quantify environmental evolution intensity, so that real-time, robust and interpretable quantitative evaluation is carried out on a complex autonomous driving working environment, and a reliable quantification basis is provided for risk early warning, safety decision and planning of a ground unmanned vehicle in the complex environment. The technical scheme for realizing the purpose of the invention is that the computer assessment method for the complexity of the autonomous running working environment of the ground unmanned vehicle comprises the following steps: Determining an autonomous running working environment complexity evaluation index of the ground unmanned vehicle, and dividing the autonomous running working environment complexity evaluation of the ground unmanned vehicle into a static environment complexity evaluation, a dynamic environment complexity evaluation and a state evolution complexity evaluation according to the determined index; Analyzing the image and the point cloud characteristics by using a deep learning perception model deployed in a vehicle-mounted computing unit to obtain index values of weather, roads and obstacle dimensions, simultaneously obtaining objective environment data by using a vehicle-mounted sensor, and finally c