A computer evaluation method for complexity of autonomous driving working environment of ground unmanned vehicle

Abstract

This invention discloses a computer-based method for evaluating the complexity of the working environment of unmanned ground vehicles (UGVs). First, it determines environmental complexity evaluation indicators and establishes an evaluation indicator system. Then, it acquires and normalizes multi-source environmental data using onboard sensors and a deep learning model. Next, it determines the weights of static environmental indicators using the Analytic Hierarchy Process (AHP) and calculates the static environmental complexity index using a nonlinear mapping function. Simultaneously, it establishes a potential energy function incorporating distance, speed, and predicted collision time, calculating the total environmental potential energy to output the dynamic environmental complexity index. For the state evolution complexity evaluation indicator, it constructs a state evolution complexity quantification model based on state change entropy, outputting the state evolution complexity index. Finally, it outputs the working environment complexity index by establishing a multi-dimensional coupled complexity model. This invention achieves real-time quantification and evaluation of working environment complexity, providing reliable data for autonomous decision-making, path planning, and safety control of UGVs.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically relating to a computer evaluation method for the complexity of the working environment of ground unmanned vehicles. Background Technology

[0002] With the rapid evolution of unmanned mobile platforms and intelligent autonomous maneuvering technologies, the environmental perception, behavioral decision-making, and motion control capabilities of ground mobile platforms in complex scenarios have become a key research direction. In actual autonomous driving environments, unmanned ground vehicles (UGVs) need to deal with multi-source heterogeneous environmental information, including static road structural elements, dynamic traffic participant behavior, and natural weather conditions. From driverless taxi services and ground cleaning operations in structured road scenarios to logistics transportation and autonomous delivery tasks in unstructured environments, the autonomous driving environments faced by UGVs exhibit diversity and uncertainty. Therefore, how to scientifically, in real-time, and objectively quantitatively assess the complexity of the autonomous driving environment of UGVs has become a crucial issue in ensuring the safety and stability of UGV operations.

[0003] Chinese patent application CN202411667369.8 discloses a technical solution for "a method for real-time evaluation of the complexity of continuous testing scenarios for autonomous vehicles". This method establishes a scenario evaluation framework that includes weather, static, and dynamic elements. It uses the extension hierarchical analysis method to determine the feature weights of scenario elements and establishes a complexity mapping relationship between scenario elements and the three subsystems of perception, decision-making, and execution by calculating the potential field model and the vehicle's reachability domain. Finally, it obtains the real-time complexity result of the continuous scenario through linear weighted coupling.

[0004] However, the aforementioned methods still suffer from the following problems when dealing with complex autonomous driving environments characterized by high dynamics and strong time variations: First, they lack a quantitative representation of the temporal evolution characteristics of the environmental state. Existing methods are mostly based on static scene assessments, making it difficult to effectively quantify the drastic changes in the environmental state over time. Second, existing assessment methods struggle to objectively quantify the complexity of dynamic environments through physical models, resulting in assessment results that are often lower than the actual complexity values, failing to provide ground-based unmanned vehicles with sufficient safety margins. Finally, existing methods often fail to balance offline pre-assessment with online real-time assessment, leading to insufficient coverage of the assessment system throughout the entire mission lifecycle. Summary of the Invention

[0005] The purpose of this invention is to provide a computer-based method for evaluating the complexity of the autonomous driving working environment of ground-based unmanned vehicles (UGVs), addressing the lack of representation of environmental state evolution in the evaluation of autonomous driving working environments for UGVs. By constructing a multi-dimensional evaluation index system, introducing a deep learning perception model to extract environmental features, and utilizing state change entropy to quantify the severity of environmental evolution, this method provides a real-time, robust, and interpretable quantitative evaluation of complex autonomous driving working environments. This offers a reliable quantitative basis for risk warning, safety decision-making, and planning of UGVs in complex environments.

[0006] The technical solution to achieve the objective of this invention is: a computer-based method for evaluating the complexity of the autonomous driving working environment of ground-based unmanned vehicles, comprising the following steps:

[0007] Step (1): Determine the evaluation index of the working environment complexity of the autonomous driving of the ground unmanned vehicle. Based on the determined index, the evaluation of the working environment complexity of the autonomous driving of the ground unmanned vehicle is divided into static environment complexity evaluation, dynamic environment complexity evaluation and state evolution complexity evaluation.

[0008] Step (2): Using the deep learning perception model deployed on the vehicle computing unit, the image and point cloud features are analyzed to obtain the index values ​​of meteorological, road and obstacle dimensions. At the same time, the vehicle sensors are used to obtain objective environmental data. Finally, the index values ​​and objective environmental data are fused through the normalization algorithm to output the index value vector containing the static environmental complexity, dynamic environmental complexity and state evolution complexity evaluation index values.

[0009] Step (3): For the static environment complexity assessment index in the numerical vector of assessment indexes, construct a static environment complexity quantification model, calculate the risk value of a single index by establishing a nonlinear risk mapping function, and determine the weight of the static environment complexity assessment index by using the AHP hierarchical analysis method. Calculate the static environment complexity index by weighted summation of risk value and weight.

[0010] Step (4): For the dynamic environment complexity assessment index in the evaluation index numerical vector, construct a dynamic environment complexity quantification model. By establishing a dynamic potential energy function that integrates relative speed, distance vector and predicted collision time, calculate the risk potential energy of a single obstacle relative to the vehicle. Accumulate and calculate the risk potential energy of all single obstacles within the perception range, and output the dynamic environment complexity index.

[0011] Step (5): For the state evolution complexity evaluation index in the evaluation index numerical vector, construct a state evolution complexity quantification model, use a time sliding window to map the continuous index numerical sequence to the discrete state space, calculate the state change entropy value of each index and sum them up by weight, and output the state evolution complexity index.

[0012] Step (6): Based on the static environment complexity index, dynamic environment complexity index and state evolution complexity index, construct a multi-dimensional coupled environmental complexity quantification model, output a comprehensive environmental complexity index, and realize the complexity assessment of the working environment.

[0013] Furthermore, the evaluation index for the complexity of the autonomous driving working environment of the ground unmanned vehicle in step (1) specifically covers 5 evaluation dimensions and 19 objective environmental indicators belonging to each dimension, as follows:

[0014] The first dimension is weather conditions B1, including rain / snow intensity index x1, visibility index x2, temperature index x3, humidity index x4, atmospheric pressure index x5, cloud cover index x6, and illuminance index x7; the second dimension is road conditions B2, including road surface adhesion coefficient index x8, road curvature index x9, and road surface unevenness index x10; the third dimension is terrain conditions B3, including altitude index x11 and slope index x12; the fourth dimension is communication conditions B4, including electromagnetic interference index x13 and signal strength index x14; the fifth dimension is obstacles B5, including obstacle speed index x15, relative distance index x16, number of obstacles index x17, collision time index x18, and obstacle type index x19.

[0015] Furthermore, the acquisition of the numerical vector of the work environment complexity evaluation index in step (2) specifically includes the following steps:

[0016] Step (21): The vehicle-mounted computing unit controls the vehicle-mounted camera, lidar, and millimeter-wave radar to acquire real-time image sequences I(t) and point cloud data H(t), and constructs the perception input tensor D at time t. in (t); At the same time, the environmental physical parameters at the current moment are directly read using temperature sensors, pressure sensors, inertial measurement units and communication sensors as objective environmental data;

[0017] Step (22): Input the sensory input tensor D in (t) Input to the deep learning perception model Φ deployed on the vehicle computing unit DL Deep learning perception model Φ DL This is a multi-task convolutional neural network built on a residual network architecture. Through multi-scale feature extraction, it achieves parallel parsing of meteorological, road and obstacle features and outputs corresponding quantitative index values. Finally, a normalization algorithm is used to fuse the quantitative index values ​​with the objective environmental data and output the evaluation index value vector of each environmental feature.

[0018] Furthermore, the construction of the static environment complexity quantification model in step (3) specifically includes the following steps:

[0019] Step (31): Set static environment complexity assessment indicators including four dimensions: weather environment, road conditions, terrain conditions, and communication environment;

[0020] Step (32): Establish the AHP hierarchical analysis method structural model and determine the target layer, criterion layer and indicator layer;

[0021] Step (33): The vehicle-mounted computing unit reads the preset parameters of relative importance of indicators stored in the database, maps the parameters into pairwise comparison values ​​based on the 1-9 scale method logic, and constructs a judgment matrix A that reflects the weight relationship between each evaluation indicator;

[0022] Step (34): Calculate the largest eigenvalue and the corresponding eigenvector of the judgment matrix A, normalize the eigenvector, and pass the consistency test to obtain the weights w' of each evaluation index, thereby obtaining the weight vector W' composed of each evaluation index.

[0023] Step (35): For any static index x i Construct a nonlinear risk mapping function:

[0024] ,

[0025] In the formula S(x i ) represents the risk value of the i-th evaluation indicator, x i Let x represent the value of the i-th evaluation indicator. ith Let σ be the critical threshold for the i-th evaluation indicator to affect driving safety. i γ represents the risk sensitivity coefficient of the i-th indicator. i γ represents the directional factor of the i-th evaluation indicator, used to unify the evaluation logic of different attribute indicators. When it is a positive indicator, γ i When γ is 1, it is a contrarian indicator. i The value is -1. Positive indicators include rain and snow intensity x1, temperature x3, humidity x4, atmospheric pressure x5, cloud cover x6, road curvature x9, road surface unevenness x10, altitude x11, slope x12, and electromagnetic interference x13. Negative indicators include visibility x2, illuminance x7, road surface adhesion coefficient x8, and signal strength x14.

[0026] Step (36): By weighted summing of the risk values ​​of each static environmental assessment indicator, a quantitative model for static environmental complexity is obtained:

[0027] ,

[0028] In the formula C static w represents the static environment complexity index. i' represents the weight of the i-th evaluation indicator.

[0029] Furthermore, the construction of the dynamic environment complexity quantification model in step (4) specifically includes the following steps:

[0030] Step (41): Set the dynamic environment complexity evaluation index to the obstacle dimension, and use on-board sensors to obtain the motion state parameters of the vehicle and the i-th dynamic obstacle in the environment, specifically including: vehicle speed v e Vehicle speed and direction angle θ e The speed of the i-th obstacle, v i The velocity direction angle θ of the i-th obstacle i And the relative position vector from the center of the vehicle to the center of the i-th obstacle. ;

[0031] Step (42): Define the kinetic potential energy of the i-th obstacle risk as Q. i :

[0032] ,

[0033] In the formula, η represents the obstacle type gain coefficient, which is dynamically assigned by the system based on the category identified by deep learning. Let represent the relative velocity vector between the vehicle and obstacle i, calculated as follows:

[0034] ,

[0035] ,

[0036] ,

[0037] In the formula Represents the vehicle's velocity vector. This represents the velocity vector of the i-th obstacle. Additionally, a direction adjustment factor is defined to determine the velocity direction between the vehicle and the obstacle.

[0038] ,

[0039] ,

[0040] In the formula α i The angle between the relative velocity vector and the relative position vector between the vehicle and the i-th obstacle is represented by cos(α). i A value greater than 0 indicates that an obstacle is close to the vehicle, cos(α) i A value less than 0 indicates that the obstacle is moving away from the vehicle, and λ represents the attenuation suppression coefficient, which reflects the degree of potential energy attenuation when the obstacle moves away from the vehicle.

[0041] Finally, the spatiotemporal risk function G is defined.i Potential energy reflecting the distance between the vehicle and the obstacle and the time of collision:

[0042] ,

[0043] In the formula, d0 represents the distance potential smoothing factor, used to eliminate singularities at zero distance, TTC. i Let δ represent the predicted collision time between the i-th obstacle and the vehicle, and let δ represent a small positive number to prevent singularities where the denominator is zero.

[0044] Step (43): Construct a dynamic potential energy function using risk intensity, direction adjustment factor, and spatiotemporal risk function:

[0045] ,

[0046] In the formula U i This represents the individual dynamic potential energy of the i-th dynamic obstacle relative to the vehicle.

[0047] Step (44): Establish a quantitative model for dynamic environment complexity:

[0048] ,

[0049] In the formula C dynamic This represents the overall dynamic environment complexity index, where m is the total number of dynamic obstacles within the perception range. The complexity of the dynamic environment is quantified by calculating the sum of the potential energy of each obstacle within the perception range.

[0050] Furthermore, the construction of the state evolution complexity quantification model in step (5) specifically includes the following steps:

[0051] Step (51): Set the state evolution complexity evaluation index, which specifically includes five dimensions: weather environment, road conditions, terrain conditions, communication environment and obstacles. The vehicle computing unit maps the evaluation index to the discrete state space, divides the preset interval into L state levels, and determines the state level corresponding to each evaluation index at the current time.

[0052] Step (52): Within a sliding window of a preset time length K, calculate the state transition rate P based on the historical state level sequence of the evaluation index, as follows:

[0053] ,

[0054] In the formula, P (i) qs N represents the probability that index i changes from state level q to state level s. (i) qs This indicates the frequency with which index i transitions from state level q to state level s;

[0055] Step (53): Based on the transition probability P (i) qs Let STE be the entropy of the state change of the i-th evaluation index at time t. i The entropy of state change (t) is calculated using a weighted average method as follows:

[0056] ,

[0057] In the formula, L represents the total number of state levels, when the transition probability P (i) qs When the value is 0, the entropy value of the state change is 0 according to the limit value. By calculating the entropy of the state change of each evaluation index, the intensity of the state transition of each index over time is quantified, reflecting the degree of abrupt change of the environmental index.

[0058] Step (54): Construct a quantification model of state evolution complexity:

[0059] ,

[0060] In the formula C change STE represents the state evolution complexity index. i This represents the entropy value of the state change of the i-th evaluation index.

[0061] Furthermore, the construction of the multidimensional coupled environment complexity model in step (6) specifically includes the following steps:

[0062] Step (61): Based on the static environment complexity index, dynamic environment complexity index, and state evolution complexity index, construct a multi-dimensional coupled environment complexity model:

[0063] ,

[0064] In the formula, C represents the working environment complexity index. This model is based on the idea of ​​joint probability distribution and can characterize the nonlinear superposition effect of multidimensional environmental features on the driving complexity of unmanned vehicles.

[0065] Step (62): Output the working environment complexity index through the environmental complexity model, and transmit the working environment complexity index C to the vehicle decision control unit through the bus; the vehicle decision control unit compares the index with the preset risk threshold, and the decision module adaptively adjusts the driving strategy according to the index level; at the same time, the index is uploaded to the remote terminal for synchronous visual risk warning to assist manual supervision.

[0066] The significant advantages of this invention compared to the prior art are as follows:

[0067] The present invention discloses a computer evaluation method for the complexity of the autonomous driving working environment of ground unmanned vehicles. It establishes a state change entropy quantification model oriented towards time-domain evolution characteristics, which characterizes the intensity of environmental evolution by statistically analyzing the state change probability of scene elements. Compared with the traditional entropy weight method that only relies on numerical dispersion, it avoids evaluation failure caused by low data dispersion and significantly improves the perception sensitivity and evaluation accuracy of the algorithm under non-stationary or extreme conditions.

[0068] The present invention discloses a computer-based method for evaluating the complexity of the autonomous driving working environment of ground unmanned vehicles. This method establishes a quantitative model of physical potential energy field complexity, which includes distance, relative speed, and collision time. Dynamic obstacles are transformed into physical entities with "potential energy." At the same time, by combining the nonlinear risk mapping of the static environment, the static environment evaluation index is mapped to a risk value to quantify its complexity. This improves the objectivity and physical interpretability of the evaluation and provides a reliable basis for the adaptive decision-making of unmanned vehicles in complex scenarios, which conforms to physical laws.

[0069] This invention discloses a computer-based method for evaluating the complexity of the working environment of an unmanned ground vehicle (UGV). This method has dual functions: offline evaluation and online real-time evaluation. In offline evaluation, the complexity is pre-quantified using pre-stored data or prior information of the target environment to improve the accuracy of environmental evaluation and provide an objective basis for task dispatch, vehicle selection, and task planning. In real-time online evaluation, the complexity of the working environment can be evaluated in real time during vehicle operation, thereby improving the safety of UGV operation and the reliability of decision-making. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the processing method of the present invention;

[0071] Figure 2 This is a graph showing the static environment complexity index in the driving scenario of this invention.

[0072] Figure 3 This is a curve showing the dynamic environment complexity index in the driving scenario of this invention.

[0073] Figure 4 This is a graph showing the exponential curve of state evolution complexity in a driving scenario according to an example of the present invention.

[0074] Figure 5 This is a diagram showing the real-time complexity evaluation results in the driving scenario of this invention. Detailed Implementation

[0075] The present invention will now be described in further detail with reference to the accompanying drawings.

[0076] See appendix Figure 1A computer-based method for evaluating the complexity of the autonomous driving environment of ground-based unmanned vehicles includes the following steps:

[0077] Step (1): Determine the evaluation index of the working environment complexity of the autonomous driving of the ground unmanned vehicle. Based on the determined index, the evaluation of the working environment complexity of the autonomous driving of the ground unmanned vehicle is divided into static environment complexity evaluation, dynamic environment complexity evaluation and state evolution complexity evaluation.

[0078] Step (2): Using the deep learning perception model deployed on the vehicle computing unit, the image and point cloud features are analyzed to obtain the index values ​​of meteorological, road and obstacle dimensions. At the same time, the vehicle sensors are used to obtain objective environmental data. Finally, the index values ​​and objective environmental data are fused through the normalization algorithm to output the index value vector containing the static environmental complexity, dynamic environmental complexity and state evolution complexity evaluation index values.

[0079] Step (3): For the static environment complexity assessment index in the numerical vector of assessment indexes, construct a static environment complexity quantification model, calculate the risk value of a single index by establishing a nonlinear risk mapping function, and determine the weight of the static environment complexity assessment index by using the AHP hierarchical analysis method. Calculate the static environment complexity index by weighted summation of risk value and weight.

[0080] Step (4): For the dynamic environment complexity assessment index in the evaluation index numerical vector, construct a dynamic environment complexity quantification model. By establishing a dynamic potential energy function that integrates relative speed, distance vector and predicted collision time, calculate the risk potential energy of a single obstacle relative to the vehicle. Accumulate and calculate the risk potential energy of all single obstacles within the perception range, and output the dynamic environment complexity index.

[0081] Step (5): For the state evolution complexity evaluation index in the evaluation index numerical vector, construct a state evolution complexity quantification model, use a time sliding window to map the continuous index numerical sequence to the discrete state space, calculate the state change entropy value of each index and sum them up by weight, and output the state evolution complexity index.

[0082] Step (6): Based on the static environment complexity index, dynamic environment complexity index and state evolution complexity index, construct a multi-dimensional coupled environmental complexity quantification model, output a comprehensive environmental complexity index, and realize the complexity assessment of the working environment.

[0083] The evaluation index for the complexity of the autonomous driving working environment of the ground unmanned vehicle mentioned in step (1) specifically covers 5 evaluation dimensions and 19 objective environmental indicators belonging to each dimension, as follows:

[0084] The first dimension is weather conditions B1, including rain / snow intensity index x1, visibility index x2, temperature index x3, humidity index x4, atmospheric pressure index x5, cloud cover index x6, and illuminance index x7; the second dimension is road conditions B2, including road surface adhesion coefficient index x8, road curvature index x9, and road surface unevenness index x10; the third dimension is terrain conditions B3, including altitude index x11 and slope index x12; the fourth dimension is communication conditions B4, including electromagnetic interference index x13 and signal strength index x14; the fifth dimension is obstacles B5, including obstacle speed index x15, relative distance index x16, number of obstacles index x17, collision time index x18, and obstacle type index x19.

[0085] Furthermore, the acquisition of the numerical vector of the work environment complexity evaluation index in step (2) specifically includes the following steps:

[0086] Step (21): The vehicle-mounted computing unit controls the vehicle-mounted camera, lidar, and millimeter-wave radar to acquire real-time image sequences I(t) and point cloud data H(t), and constructs the perception input tensor D at time t. in (t); At the same time, the environmental physical parameters at the current moment are directly read using temperature sensors, pressure sensors, inertial measurement units and communication sensors as objective environmental data;

[0087] Step (22): Convert the sensory input tensor D in (t) Input to the deep learning perception model Φ deployed on the vehicle computing unit DL Deep learning perception model Φ DL This is a multi-task convolutional neural network built on a residual network architecture. It achieves parallel parsing and outputs corresponding quantitative index values ​​for meteorological, road, and obstacle features through multi-scale feature extraction. Finally, a normalization algorithm is used to fuse the quantitative index values ​​with the objective environmental data, outputting an evaluation index vector for each environmental feature. The network specifically includes the following three parallel branches:

[0088] ① Global Environment Classification Branch: Utilizes convolutional neural networks to extract global semantic features of images, and outputs scene classification probabilities of weather environment through fully connected layers, which are used to quantify meteorological dimension indicators such as rain / snow intensity x1, visibility x2, and cloud cover x6;

[0089] ② Road semantic segmentation branch: The road area is segmented at the pixel level using an encoder-decoder structure, and the geometric features of the drivable area are calculated to obtain the road curvature x9; at the same time, the road surface material and condition are identified, and the road surface adhesion coefficient x8 and road surface roughness x10 are obtained by looking up a table.

[0090] ③ Dynamic Target Detection Branch: The target detection sub-network is used to identify dynamic obstacles in the field of view, and the obstacle bounding box and category information are output to determine the obstacle type x19 and number x17; the visual detection box is spatiotemporally matched with the point cloud depth information of laser / millimeter-wave radar to accurately calculate the relative speed of the obstacle x15, the relative distance x16 and the predicted collision time x18.

[0091] Furthermore, the construction of the static environment complexity quantification model described in step (3) specifically includes the following steps:

[0092] Step (31): Set the indicators for static environment complexity, including four dimensions: weather environment, road conditions, terrain conditions, and communication environment;

[0093] Step (32): Establish a hierarchical structure model: the target layer is scene complexity, the criterion layer is weather environment B1, road conditions B2, terrain conditions B3, communication environment B4, and the index layer is rain and snow intensity index x1, visibility index x2, temperature index x3, humidity index x4, atmospheric pressure index x5, cloud cover index x6, illuminance index x7, road surface adhesion coefficient index x8, road curvature index x9, road surface unevenness index x10, altitude index x11, slope index x12, electromagnetic interference index x13, and signal strength index x14.

[0094] Step (33): The on-board computing unit reads the preset parameters of relative importance of indicators stored in the database, maps the parameters into pairwise comparison values ​​based on the 1-9 scaling method, and constructs a judgment matrix reflecting the weight relationship between each evaluation indicator:

[0095]

[0096] In the formula, A is the judgment matrix, representing the importance of each evaluation indicator, where a ij This indicates the importance of the indicator in the i-th row relative to the indicator in the j-th column, and n represents the number of indicators.

[0097] Step (34): Calculate the weight vector: First, calculate the geometric mean of the elements in each row:

[0098]

[0099] In the formula w ai This indicates that the weight of the i-th evaluation indicator is unnormalized. The weights of each indicator are then normalized to obtain the final weights.

[0100]

[0101] In the formula w i' represents the weight of evaluation indicator i, and k is the index of evaluation indicator. The weight vector is obtained based on the calculated weights of each evaluation indicator. =( , ,…, For each obtained weight vector W' of the evaluation index, calculate the consistency ratio CR. If CR < 0.1, the judgment matrix is ​​considered to be consistent, and W' is confirmed as an effective weight vector.

[0102] Step (35): For any static index x i Construct its nonlinear risk mapping function:

[0103]

[0104] In the formula S(x i ) represents the risk value of the i-th evaluation indicator, x i Let x represent the value of the i-th evaluation indicator. ith Let σ be the critical threshold for the i-th evaluation indicator to affect driving safety. i Represents the risk sensitivity coefficient; γ i γ represents the directional factor of the i-th evaluation indicator, used to unify the evaluation logic of different attribute indicators. When it is a positive indicator, γ i When γ is 1, it is a contrarian indicator. i The value is -1. Positive indicators include rain and snow intensity x1, temperature x3, humidity x4, atmospheric pressure x5, cloud cover x6, road curvature x9, road surface unevenness x10, altitude x11, slope x12, and electromagnetic interference x13. Negative indicators include visibility x2, illuminance x7, road surface adhesion coefficient x8, and signal strength x14.

[0105] Step (36): By weighted summing of the risk values ​​of each static environmental assessment indicator, a quantitative model for static environmental complexity is obtained:

[0106]

[0107] In the formula C static w represents the static environment complexity index. i ' represents the weight of the i-th evaluation indicator.

[0108] Furthermore, the construction of the dynamic environment complexity quantification model described in step (4) specifically includes the following steps:

[0109] Step (41): Set the dynamic environment complexity evaluation index to the obstacle dimension, and use on-board sensors to obtain the motion state parameters of the vehicle and the i-th dynamic obstacle in the environment, specifically including: vehicle speed ve Vehicle speed and direction angle θ e The speed of the i-th obstacle, v i The velocity direction angle θ of the i-th obstacle i And the relative position vector from the center of the vehicle to the center of the i-th obstacle. ;

[0110] Step (42): Define the kinetic potential energy of the i-th obstacle risk as Q. i :

[0111]

[0112] In the formula, η represents the obstacle type gain coefficient, which is dynamically assigned by the system based on the deep learning-based classification: pedestrians and non-motorized vehicles are set to 1.8, large engineering vehicles and buses are set to 1.5, ordinary standard vehicles are set to a baseline value of 1.0, and other unclassified obstacles are set to 1.2, which is used to adjust the environmental potential energy intensity. Let represent the relative velocity vector between the vehicle and obstacle i, which is calculated as follows:

[0113] , ,

[0114] In the formula Represents the vehicle's velocity vector. This represents the velocity vector of the i-th obstacle. Additionally, a direction adjustment factor is defined to determine the velocity direction between the vehicle and the obstacle.

[0115] ,

[0116] In the formula α i The angle between the relative velocity vector and the relative position vector between the vehicle and the i-th obstacle is represented by cos(α). i A value greater than 0 indicates that an obstacle is close to the vehicle, cos(α) i A value less than 0 indicates that the obstacle is moving away from the vehicle, and λ represents the attenuation suppression coefficient, which reflects the degree of potential energy attenuation when the obstacle moves away from the vehicle.

[0117] Finally, the spatiotemporal risk function G is defined. i Potential energy reflecting the distance between the vehicle and the obstacle and the time of collision:

[0118]

[0119] In the formula, d0 represents the distance potential smoothing factor, used to eliminate singularities at zero distance, TTC. i Let δ represent the predicted collision time between the i-th obstacle and the vehicle, and let δ represent a small positive number to prevent singularities where the denominator is zero.

[0120] Step (43): Construct a dynamic potential energy function using risk intensity, direction adjustment factor, and spatiotemporal risk function:

[0121]

[0122] In the formula U i This represents the individual dynamic potential energy of the i-th dynamic obstacle relative to the vehicle.

[0123] Step (44): Establish a quantitative model for dynamic environment complexity:

[0124]

[0125] In the formula C dynamic This represents the overall dynamic environment complexity index, where m is the total number of dynamic obstacles within the perception range. The complexity of the dynamic environment is quantified by calculating the sum of the potential energy of each obstacle within the perception range.

[0126] Furthermore, the construction of the state evolution complexity quantification model in step (5) specifically includes the following steps:

[0127] Step (51): Set the state evolution complexity evaluation index, which specifically includes five dimensions: weather environment, road conditions, terrain conditions, communication environment, and obstacles. Map each normalized evaluation index to a discrete state space, set L state levels, and divide the interval [0, 1] into L sub-intervals. For index i, its state level S at time t is... i (t)∈{1, 2, 3,…, L};

[0128] Step (52): Introduce a time sliding window of length K. At the current time t, backtrack the data sequence {S} of the past K frames. i (t-K+1),…,S i Within the window, calculate the state transition probability P of the i-th evaluation index. (i) qs The calculation is as follows:

[0129]

[0130] In the formula, P (i) qs N represents the probability that index i changes from state level q to state level s. (i) qs This indicates the frequency with which index i transitions from state level q to state level s;

[0131] Step (53): Based on the transition probability P (i) qsLet STE be the entropy of the state change of the i-th evaluation index at time t. i (t), to reflect the "transfer uncertainty" and "variation range" of each evaluation indicator, is calculated using a weighted average method as follows:

[0132]

[0133] In the formula, L represents the total number of state levels, when the transition probability P (i) qs When the value is 0, the entropy value of the state change is 0 according to the limit value. By calculating the entropy of the state change of each evaluation index, the intensity of the state transition of each index over time is quantified, reflecting the degree of abrupt change of the environmental index.

[0134] Step (54): Construct a quantification model of state evolution complexity:

[0135]

[0136] In the formula C change It represents the state evolution complexity index.

[0137] Furthermore, the construction of the multidimensional coupled environment complexity model in step (6) specifically includes the following steps:

[0138] Step (61): Based on the static environment complexity index, dynamic environment complexity index, and state evolution complexity index, construct a multi-dimensional coupled environment complexity model:

[0139]

[0140] In the formula, C represents the working environment complexity index. This model is based on the idea of ​​joint probability distribution and can characterize the nonlinear superposition effect of multidimensional environmental features on the driving complexity of unmanned vehicles.

[0141] Step (62): Output the working environment complexity index through the environmental complexity model, and transmit the working environment complexity index C to the vehicle decision control unit through the bus; the vehicle decision control unit compares the index with the preset risk threshold, and the decision module adaptively adjusts the driving strategy according to the index level; at the same time, the index is uploaded to the remote terminal for synchronous visual risk warning to assist manual supervision.

[0142] Example

[0143] This embodiment specifically illustrates the actual deployment and operation process of a computer-based evaluation method for the complexity of the autonomous driving working environment of a ground-based unmanned vehicle on an experimental vehicle, verifying the effectiveness of the invention in a mixed test road scenario. In terms of hardware implementation, the experimental vehicle is equipped with a multimodal perception sensor group consisting of a solid-state LiDAR, a high-definition camera, a millimeter-wave radar, and a micro-meteorological sensor, used to capture raw data on weather, road conditions, terrain conditions, communication environment, and obstacles from all directions. The vehicle computing platform adopts an automotive-grade high-performance autonomous driving domain control unit. Based on this hardware, the system first constructs an evaluation system according to the present invention, including 5 evaluation dimensions and 19 objective environmental indicators belonging to each dimension, namely, weather environment B1, road conditions B2, terrain conditions B3, communication environment B4, and obstacles B5. Before the vehicle is put into operation, the system pre-sets a static environmental indicator weight vector calculated based on the AHP (Analytic Hierarchy Process). The onboard computing unit scores the importance of each of the 14 static indicators in the weather, road, terrain, and communication dimensions pairwise, constructing a judgment matrix A. After calculation and normalization, the weight vector W' of each static index was established as (0.05, 0.12, 0.08, 0.08, 0.04, 0.06, 0.07, 0.15, 0.10, 0.05, 0.05, 0.05, 0.05, 0.05). This vector was fixed as a static benchmark parameter in the configuration file of the on-board computing unit.

[0144] The autonomous vehicle entered autonomous cruising mode at a speed of 20 km / h in the test site, with data flow following the logic of "perception-processing-quantization-fusion-output". In the perception phase, onboard sensors synchronously acquired real-time image sequences, point cloud data, and objective environmental parameters at a frequency of 10 Hz. After time synchronization and spatial alignment, the original input tensor at time t was constructed and input into the convolutional neural network model Φ. DL The original values ​​of each indicator output by the model are then normalized by the onboard computing unit. For static environment complexity, the system uses a nonlinear risk mapping function to calculate any static indicator, and for the safety critical threshold x... ith and risk sensitivity σ i The core indicators are set as follows: rain / snow intensity x1, visibility x2, road surface adhesion coefficient x8, and signal strength x14. ith σ is 0.25. i The index is 12; illuminance x7, road curvature x9, unevenness x10, slope x12, and electromagnetic interference x13. ith σ is 0.40. i The value is 8.0; the indicators are temperature x3, humidity x4, atmospheric pressure x5, cloud cover x6, and altitude x11. ith It is 0.75, σ iThe value is 2.5. Based on this, the risk value of each static indicator is calculated, and combined with the weight vector W', the static environment complexity index C is output. static ,See Figure 2 ,from Figure 2 As can be seen, the curve exhibits a smooth, single-peaked trend, representing the entire process of the autonomous vehicle entering and leaving a low-adhesion coefficient road section. This result demonstrates that the nonlinear risk mapping function can effectively characterize the gradual changes in static environmental elements with spatial location. Simultaneously, for the dynamic obstacle dimension, a dynamic potential energy model is constructed, setting the attenuation suppression coefficient λ to 0.6 and the small positive number δ (0.001) to prevent the denominator from being zero. The system calculates in real-time the risk intensity, direction adjustment factor, and spatiotemporal risk function including TTC for each obstacle within the perception range, summing them to obtain the total dynamic environment complexity index C. dynamic ,See Figure 3 ,from Figure 3 As can be seen, the curve exhibits a significant multi-peak characteristic, which precisely corresponds to the three dynamic obstacle interaction events in the experiment; in particular, the exponent reaches a saturation value of 1.0 during the period from t=8s to t=12s, indicating that both extremely close distance and high relative velocity are present at this time, verifying the sensitivity of the dynamic potential energy model in quantifying high-risk collision risks.

[0145] To quantify the drastic nature of environmental evolution, the system sets the time sliding window length K to 20 frames (i.e., backtracking through 2 seconds of data) and maps the normalized indicators to a discrete state space, setting the number of state levels L to 10. Within the sliding window of K=20, the system counts the transition frequency of each indicator between different state levels and calculates the state transition probability P. Based on the probability P, the system calculates the state change entropy STE of each indicator and then uses a weighted average to calculate the final output state evolution complexity index C. change ,See Figure 4 ,from Figure 4 As can be seen, the curve exhibits high-frequency pulse oscillation characteristics, with high values ​​mainly concentrated at operational condition switching nodes (such as the transition from a safe to a dangerous state). This result indicates that the state change entropy model can effectively capture the uncertainty and severity of changes during environmental state transitions. Finally, the system fuses the three complexity indices through a multi-dimensional coupling model, outputting the final working environment complexity index C, as shown in [reference needed]. Figure 5 ,from Figure 5 As can be seen, the total index curve integrates the characteristics of static, dynamic, and state evolution environments. It reaches its highest value in the interval from t=8s to t=12s due to the superposition of high static risk and strong dynamic potential energy. This verifies that the multi-dimensional coupling model can objectively reflect the superposition effect of multiple sources of risk, providing a quantitative basis with sufficient safety margin for vehicle adaptive decision-making. The index C is transmitted to the decision-making and planning unit via the bus, where the driving strategy is adaptively adjusted based on threshold comparison, and simultaneously uploaded to a remote terminal for visualized monitoring.

Claims

1. A computer-based method for evaluating the complexity of the working environment of an unmanned ground vehicle, characterized in that, Includes the following steps: Step (1): Determine the evaluation index of the working environment complexity of the autonomous driving of the ground unmanned vehicle. Based on the determined index, the evaluation of the working environment complexity of the autonomous driving of the ground unmanned vehicle is divided into static environment complexity evaluation, dynamic environment complexity evaluation and state evolution complexity evaluation. Step (2): Using the deep learning perception model deployed on the vehicle computing unit, the image and point cloud features are analyzed to obtain the index values ​​of meteorological, road and obstacle dimensions. At the same time, the vehicle sensors are used to obtain objective environmental data. Finally, the index values ​​and objective environmental data are fused through the normalization algorithm to output the index value vector containing the static environmental complexity, dynamic environmental complexity and state evolution complexity evaluation index values. Step (3): For the static environment complexity assessment index in the numerical vector of assessment indexes, construct a static environment complexity quantification model, calculate the risk value of a single index by establishing a nonlinear risk mapping function, and determine the weight of the static environment complexity assessment index by using the AHP hierarchical analysis method. Calculate the static environment complexity index by weighted summation of risk value and weight. Step (4): For the dynamic environment complexity assessment index in the evaluation index numerical vector, construct a dynamic environment complexity quantification model. By establishing a dynamic potential energy function that integrates relative speed, distance vector and predicted collision time, calculate the risk potential energy of a single obstacle relative to the vehicle. Accumulate and calculate the risk potential energy of all single obstacles within the perception range, and output the dynamic environment complexity index. Step (5): For the state evolution complexity evaluation index in the evaluation index numerical vector, construct a state evolution complexity quantification model, use a time sliding window to map the continuous index numerical sequence to the discrete state space, calculate the state change entropy value of each index and sum them up by weight, and output the state evolution complexity index. Step (6): Based on the static environment complexity index, dynamic environment complexity index and state evolution complexity index, construct a multi-dimensional coupled environmental complexity quantification model, output a comprehensive environmental complexity index, and realize the complexity assessment of the working environment.

2. The method according to claim 1, characterized in that, The evaluation index for the complexity of the autonomous driving working environment of the ground unmanned vehicle in step (1) specifically covers 5 evaluation dimensions and 19 objective environmental indicators belonging to each dimension, as follows: The first dimension is weather conditions B1, including rain / snow intensity index x1, visibility index x2, temperature index x3, humidity index x4, atmospheric pressure index x5, cloud cover index x6, and illuminance index x7; the second dimension is road conditions B2, including road surface adhesion coefficient index x8, road curvature index x9, and road surface unevenness index x10; the third dimension is terrain conditions B3, including altitude index x11 and slope index x12; the fourth dimension is communication conditions B4, including electromagnetic interference index x13 and signal strength index x14; the fifth dimension is obstacles B5, including obstacle speed index x15, relative distance index x16, number of obstacles index x17, collision time index x18, and obstacle type index x19.

3. The method according to claim 2, characterized in that, The acquisition of the numerical vector of the work environment complexity evaluation index in step (2) specifically includes the following steps: Step (21): The vehicle-mounted computing unit controls the vehicle-mounted camera, lidar, and millimeter-wave radar to acquire real-time image sequences I(t) and point cloud data H(t), and constructs the perception input tensor D at time t. in (t); At the same time, the environmental physical parameters at the current moment are directly read using temperature sensors, pressure sensors, inertial measurement units and communication sensors as objective environmental data; Step (22): Input the sensory input tensor D in (t) Input to the deep learning perception model Φ deployed on the vehicle computing unit DL Deep learning perception model Φ DL This is a multi-task convolutional neural network built on a residual network architecture. Through multi-scale feature extraction, it achieves parallel parsing of meteorological, road and obstacle features and outputs corresponding quantitative index values. Finally, a normalization algorithm is used to fuse the quantitative index values ​​with the objective environmental data and output the evaluation index value vector of each environmental feature.

4. The method according to claim 3, characterized in that, The construction of the static environment complex quantification model in step (3) specifically includes the following steps: Step (31): Set static environment complexity assessment indicators including four dimensions: weather environment, road conditions, terrain conditions, and communication environment; Step (32): Establish the AHP hierarchical analysis method structural model and determine the target layer, criterion layer and indicator layer; Step (33): The vehicle-mounted computing unit reads the preset parameters of relative importance of indicators stored in the database, maps the parameters into pairwise comparison values ​​based on the 1-9 scale method logic, and constructs a judgment matrix A that reflects the weight relationship between each evaluation indicator; Step (34): Calculate the largest eigenvalue and the corresponding eigenvector of the judgment matrix A, normalize the eigenvector, and pass the consistency test to obtain the weights w' of each evaluation index, thereby obtaining the weight vector W' composed of each evaluation index. Step (35): For any static index x i Construct a nonlinear risk mapping function: ， In the formula S(x i ) represents the risk value of the i-th evaluation indicator, x i Let x represent the value of the i-th evaluation indicator. ith Let σ be the critical threshold that the i-th evaluation index affects driving safety. i γ represents the risk sensitivity coefficient of the i-th indicator. i γ represents the directional factor of the i-th evaluation indicator, used to unify the evaluation logic of different attribute indicators. When it is a positive indicator, γ i When γ is 1, it is a contrarian indicator. i The value is -1. Positive indicators include rain and snow intensity x1, temperature x3, humidity x4, atmospheric pressure x5, cloud cover x6, road curvature x9, road surface unevenness x10, altitude x11, slope x12, and electromagnetic interference x13. Negative indicators include visibility x2, illuminance x7, road surface adhesion coefficient x8, and signal strength x14. Step (36): By weighted summing of the risk values ​​of each static environmental assessment indicator, a quantitative model for static environmental complexity is obtained: ， In the formula C static w represents the static environment complexity index. i ' represents the weight of the i-th evaluation indicator.

5. The method according to claim 4, characterized in that, The construction of the dynamic environment complexity quantification model in step (4) specifically includes the following steps: Step (41): Set the dynamic environment complexity evaluation index to the obstacle dimension, and use on-board sensors to obtain the motion state parameters of the vehicle and the i-th dynamic obstacle in the environment, specifically including: vehicle speed v e Vehicle speed and direction angle θ e The speed of the i-th obstacle, v i The velocity direction angle θ of the i-th obstacle i And the relative position vector from the center of the vehicle to the center of the i-th obstacle. ; Step (42): Define the kinetic potential energy of the i-th obstacle risk as Q. i : ， In the formula, η represents the obstacle type gain coefficient, which is dynamically assigned by the system based on the category identified by deep learning. Let represent the relative velocity vector between the vehicle and obstacle i, calculated as follows: ， ， ， In the formula Represents the vehicle's velocity vector. This represents the velocity vector of the i-th obstacle. Additionally, a direction adjustment factor is defined to determine the velocity direction between the vehicle and the obstacle. ， ， In the formula α i The angle between the relative velocity vector and the relative position vector between the vehicle and the i-th obstacle is represented by cos(α). i A value greater than 0 indicates that an obstacle is close to the vehicle, cos(α) i A value less than 0 indicates that the obstacle is moving away from the vehicle, and λ represents the attenuation suppression coefficient, which reflects the degree of potential energy attenuation when the obstacle moves away from the vehicle. Finally, the spatiotemporal risk function G is defined. i Potential energy reflecting the distance between the vehicle and the obstacle and the time of collision: ， In the formula, d0 represents the distance potential smoothing factor, used to eliminate singularities at zero distance, TTC. i Let δ represent the predicted collision time between the i-th obstacle and the vehicle, and let δ represent a small positive number to prevent singularities where the denominator is zero. Step (43): Construct a dynamic potential energy function using risk intensity, direction adjustment factor, and spatiotemporal risk function: ， In the formula U i This represents the individual dynamic potential energy of the i-th dynamic obstacle relative to the vehicle. Step (44): Establish a quantitative model for dynamic environment complexity: ， In the formula C dynamic This represents the overall dynamic environment complexity index, where m is the total number of dynamic obstacles within the perception range. The complexity of the dynamic environment is quantified by calculating the sum of the potential energy of each obstacle within the perception range.

6. The method according to claim 5, characterized in that, The construction of the state evolution complexity quantification model in step (5) specifically includes the following steps: Step (51): Set the state evolution complexity evaluation index, which specifically includes five dimensions: weather environment, road conditions, terrain conditions, communication environment and obstacles. The vehicle computing unit maps the evaluation index to the discrete state space, divides the preset interval into L state levels, and determines the state level corresponding to each evaluation index at the current time. Step (52): Within a sliding window of a preset time length K, calculate the state transition rate P based on the historical state level sequence of the evaluation index, as follows: ， In the formula, P (i) qs N represents the probability that index i changes from state level q to state level s. (i) qs This indicates the frequency with which index i transitions from state level q to state level s; Step (53): Based on the transition probability P (i) qs Let STE be the entropy of the state change of the i-th evaluation index at time t. i The entropy of state change (t) is calculated using a weighted average method as follows: ， In the formula, L represents the total number of state levels, when the transition probability P (i) qs When the value is 0, the entropy value of the state change is 0 according to the limit value. By calculating the entropy of the state change of each evaluation index, the intensity of the state transition of each index over time is quantified, reflecting the degree of abrupt change of the environmental index. Step (54): Construct a quantification model of state evolution complexity: ， In the formula C change STE represents the state evolution complexity index. i This represents the entropy value of the state change of the i-th evaluation index.

7. The method according to claim 6, characterized in that, The construction of the multidimensional coupled environment complexity model in step (6) specifically includes the following steps: Step (61): Based on the static environment complexity index, dynamic environment complexity index, and state evolution complexity index, construct a multi-dimensional coupled environment complexity model: ， In the formula, C represents the working environment complexity index. This model is based on the idea of ​​joint probability distribution and can characterize the nonlinear superposition effect of multidimensional environmental features on the driving complexity of unmanned vehicles. Step (62): Output the working environment complexity index through the environmental complexity model, and transmit the working environment complexity index C to the vehicle decision control unit through the bus. The vehicle decision control unit compares the index with a preset risk threshold, and the decision module adaptively adjusts the driving strategy according to the index level; at the same time, the index is uploaded to a remote terminal for synchronous visual risk warning to assist manual supervision.

Images