Automatic driving vehicle highway suitability evaluation method based on ensemble learning

Abstract

The application relates to a kind of expressway adaptability evaluation methods of automatic driving vehicle based on ensemble learning, comprising the following steps: collecting the driving data and corresponding traffic data of automatic driving vehicle driving on expressway and carrying out pretreatment and class extraction, to obtain feature data;Utilize lane width detection data to judge lane detection failure type;Construct Boosting ensemble learning model, use feature data as input, use lane detection failure type as output, train the model;Based on the trained Boosting ensemble learning model, predict the lane detection failure type of the expressway to be evaluated, and obtain the automatic driving vehicle expressway adaptability evaluation conclusion.Compared with the prior art, the application realizes the construction of the nonlinear relationship between the expressway linear design, traffic marking attribute, traffic operation and environmental characteristics and lane detection performance, and solves the problem that the expressway adaptability of automatic driving vehicle is difficult to quantitatively evaluate.

Description

Technical Field

[0001] This invention relates to the field of safety assessment technology, and in particular to a method for assessing the drivability of autonomous vehicles on highways based on ensemble learning. Background Technology

[0002] Highways operate at high speeds and feature complex vehicle compositions, resulting in higher safety risks and greater accident severity. Current road design specifications regarding road alignment and traffic markings only consider the perception capabilities of drivers, neglecting the perception performance of autonomous vehicles. Due to differences in perception reaction time, distance, and altitude between drivers and autonomous vehicles, autonomous vehicles may face challenges under specific highway conditions, impacting their operational safety. Highway drivability assessment reflects the adaptability of highway infrastructure to existing autonomous driving technologies. Its evaluation includes road alignment, cross-section, pavement performance, traffic signs, and traffic markings. Conducting highway drivability assessments for autonomous vehicles and developing optimization strategies for road sections that fail the assessment can improve the operational safety of autonomous vehicles.

[0003] Existing technologies, focusing on lane departure warning and lane keeping functions of Advanced Driving Assistance Systems (ADAS) for autonomous vehicles, have collected data on function failure events based on field tests. These technologies analyze the influencing factors of these failure events by constructing generalized estimation equations, logistic regression models, or using scatter plot visualization. However, these methods cannot quantify the nonlinear relationship between function failure events and input characteristics. Furthermore, the models developed lack transferability in road drivability assessments, making it difficult to guide road design and the development of optimization strategies for onboard sensor performance. Summary of the Invention

[0004] The purpose of this invention is to provide a highway drivability assessment method for autonomous vehicles based on ensemble learning, which enables quantitative analysis of the impact of highway alignment design, traffic marking attributes, traffic operation, and environmental characteristics on the lane detection performance of autonomous vehicles, thus solving the problem of difficulty in quantitatively assessing the highway drivability of autonomous vehicles in existing technologies.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A highway drivability assessment method for autonomous vehicles based on ensemble learning includes the following steps:

[0007] Step 1) Collect driving data and corresponding traffic data of autonomous vehicles traveling on highways, wherein the driving data includes lane width detection data and vehicle operation data;

[0008] Step 2) Preprocess the raw data collected in Step 1) and extract the categories from the categorical variables to obtain feature data;

[0009] Step 3) For the lane width detection data preprocessed in Step 2), lane width detection failure events are used as an indicator to measure lane detection performance and determine the type of lane detection failure.

[0010] Step 4) Construct a Boosting ensemble learning model to learn the impact of input feature data on the lane detection performance of autonomous vehicles. Use the feature data as the model input and the lane detection failure type as the output to train the model.

[0011] Step 5) Based on the Boosting ensemble learning model trained in Step 4), predict the lane detection failure type of the highway to be evaluated, and draw a conclusion on the highway drivability assessment of autonomous vehicles.

[0012] Step 1) specifically involves: collecting lane width detection data and vehicle operation data using onboard sensors; collecting road alignment data based on road design drawings; and collecting traffic marking data, road segment feature data, and environmental data using onboard video.

[0013] Step 2) includes the following steps:

[0014] Step 2-1) Match lane width detection data, vehicle operation data, road alignment data, traffic marking data, road segment feature data, and environmental data to the same spatial coordinates based on timestamps, latitude and longitude, and highway station numbers;

[0015] Step 2-2) Based on the vehicle operation data, the instantaneous speed is taken as the vehicle speed; the distance between the vehicle and the nearest large vehicle in the current downstream lane or adjacent lane is calculated as the distance to the downstream large vehicle.

[0016] Steps 2-3) Based on the road alignment data and road design drawings, extract the average curvature and slope change rate of the downstream preset distance;

[0017] Steps 2-4) Extract the category of traffic marking data. Traffic marking data includes longitudinal marking types, special marking types, and marking defects types. Among them, longitudinal marking types are divided into double-sided dashed lines, left-side dashed line and right-side solid line, left-side solid line and right-side dashed line, double-sided solid lines, acceleration lane transition section extension line, and deceleration lane transition section extension line; special marking types are divided into ordinary markings, arrow markings, vehicle distance confirmation markings, lateral deceleration markings, and text markings; marking defects types are divided into ordinary markings, marking wear, old marking overlap, and traffic cone coverage.

[0018] Steps 2-5) Extract the category of road segment feature data. The road segment feature data includes road segment type and lane position. Specifically, the road segment type is divided into general main road segment, entrance main road segment, and exit main road segment; and the lane position is divided into leftmost lane, rightmost lane, and middle lane.

[0019] Steps 2-6) Extract the category of the environmental data. The environmental data includes lighting and weather. Lighting is divided into daytime, nighttime without streetlights, nighttime with streetlights, dawn, and dusk. Weather is divided into sunny, cloudy, light rain, moderate rain, heavy to torrential rain, light snow, moderate snow, heavy to torrential snow, and fog.

[0020] Step 3) includes the following steps:

[0021] Step 3-1) Define the lane width detection failure event as follows: Assuming the autonomous vehicle is located on the center line of the detection lane and one lane line of the detection lane overlaps with the lane line of the actual lane, when the lane width detection value is too wide or too narrow and the vehicle crosses the actual lane line, it is a critical safety situation for the autonomous vehicle to collide with the adjacent vehicle, and is regarded as a lane width detection failure event.

[0022] Step 3-2) Calculate the lane width detection threshold for lane width detection failure events, wherein the lane width detection threshold includes a lower threshold and an upper threshold;

[0023] Step 3-3) Determine the lane detection failure type based on the lane width detection threshold. When the lane width is greater than or equal to the lower threshold and less than or equal to the upper threshold, the lane detection failure type is not failed; otherwise, the lane detection failure type is failed.

[0024] The formula for calculating the lane width detection threshold is:

[0025] |W md -W a |+W v =W a

[0026] Among them, W md W represents the lane width detection threshold. aW represents the actual lane width. v This refers to the width of the vehicle.

[0027] Step 4) includes the following steps:

[0028] Step 4-1) Divide the preprocessed data into training and testing sets;

[0029] Step 4-2) Use the vehicle operation data, road alignment data, traffic marking data, road segment feature data, and environmental data from the training set as input features, and the lane detection failure type as output feature to train the Boosting ensemble learning model; evaluate the performance of the model training using the confusion matrix based on the test set data.

[0030] Step 5) includes the following steps:

[0031] Step 5-1) Take the vehicle operation data, road alignment data, traffic marking data, road segment feature data and environmental data of the highway to be evaluated as input features, input them into the trained Boosting ensemble learning model, and output the lane detection failure type prediction value of the autonomous vehicle in each road segment unit of the highway to be evaluated.

[0032] Step 5-2) Based on the lane detection failure type prediction value, determine the highway drivability assessment conclusion for autonomous vehicles. For road segment units with a prediction value of "failure", the assessment conclusion is "failure"; for road segment units with a prediction value of "not failed", the assessment conclusion is "pass".

[0033] The method further includes the following steps:

[0034] Step 6) Based on the Boosting ensemble learning model trained in Step 4), determine the global feature importance and feature dependencies;

[0035] Step 7) For highway segment units that failed the drivability assessment in Step 5), use the global feature importance and feature dependency relationships determined in Step 6) to formulate road design optimization strategies, determine the performance optimization priorities of on-board sensors, and improve the lane detection performance of autonomous vehicles on highways.

[0036] Step 6) includes the following steps:

[0037] Step 6-1) Using the SHAP method, based on the average absolute value of each input feature, calculate the magnitude of the influence of the input feature on the model's predicted output, i.e., the global feature importance.

[0038] Step 6-2) To reveal the impact of a single input feature on the model's predicted output, a SHAP dependency graph of the single input feature is plotted. The positive and negative trends and magnitudes of the impact of changes in the input feature on the predicted output are analyzed. Positive impact means the predicted value is "ineffective" and negative impact means the predicted value is "not ineffective".

[0039] Step 6-3) Considering the interactive effects between input features, in order to reveal the pairwise interactive effects of a single input feature and other features on the model's predicted output, draw the interaction SHAP dependency graph of two input features, and analyze the positive and negative trends and magnitude changes in the influence of changes in the combination of input features on the predicted output.

[0040] Step 7) includes the following steps:

[0041] Step 7-1) Based on the global feature importance calculated in Step 6), determine the feature priority for the optimization strategy.

[0042] Step 7-2) Based on the feature dependency relationship constructed in Step 6), along the feature value range optimization direction where the positive influence of input features on the model prediction output is weakened or the negative influence is increased, formulate countermeasures to improve the highway driving suitability of autonomous vehicles from three dimensions: road alignment and traffic marking design optimization, autonomous vehicle operation design domain management, and on-board sensor performance optimization, so as to ensure the safe operation of autonomous vehicles.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] (1) This invention provides a method for quantitatively evaluating drivability. Based on the detection data of on-board sensors of autonomous vehicles, lane width detection failure events are used as an indicator to measure lane detection performance, so as to quantitatively evaluate the highway drivability of autonomous vehicles.

[0045] (2) This invention utilizes the Boosting ensemble learning algorithm to construct a model of the influence of input feature data on the lane detection performance of autonomous vehicles. It can quantitatively reflect the nonlinear relationship between features and predicted output, comprehensively embody the advantages of gradient boosting, parallel processing, pruning and regularization techniques, process complex datasets, and provide excellent model prediction accuracy.

[0046] (3) This invention adopts the Shapley additive explanations (SHAP) method and utilizes the interpretability of machine learning models to realize the quantitative analysis of the impact of highway alignment design, traffic marking attributes, traffic operation and environmental characteristics on the lane detection performance of autonomous vehicles. This provides a scientific and evidence-based theoretical foundation for proposing countermeasures to improve the highway driving suitability of autonomous vehicles. Attached Figure Description

[0047] Figure 1 This is a flowchart of the method of the present invention;

[0048] Figure 2 One embodiment is a feature importance ranking based on the average absolute value of SHAP;

[0049] Figure 3 This is a SHAP dependency graph of the slope change rate feature in one embodiment;

[0050] Figure 4 This is an example of an interactive SHAP dependency graph between road segment type and lane location features. Detailed Implementation

[0051] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0052] A model based on ensemble learning algorithms is constructed to reflect the influence of input feature data on the lane detection performance of autonomous vehicles. This model can quantitatively reflect the nonlinear relationship between features and the predicted lane detection performance output, addressing the difficulty of quantitatively assessing the highway drivability of autonomous vehicles using existing technologies. Therefore, by utilizing lane detection performance and establishing a highway drivability assessment method based on ensemble learning algorithms, this approach better enables the quantitative analysis of the impact of highway alignment design, traffic marking attributes, traffic operation, and environmental characteristics on the lane detection performance of autonomous vehicles, thus more effectively improving their highway drivability.

[0053] Specifically, such as Figure 1 As shown, this embodiment provides a method for evaluating the drivability of autonomous vehicles on highways based on ensemble learning, including the following steps:

[0054] Step 1) Collect driving data and corresponding traffic data of autonomous vehicles traveling on highways.

[0055] Specifically, vehicle-mounted sensors are used to collect lane width detection data and vehicle operation data; road alignment data is collected based on road design drawings; and traffic marking data, road segment feature data, and environmental data are collected using vehicle-mounted video.

[0056] Step 2) Preprocess the raw data collected in Step 1) and extract the categories from the categorical variables to obtain feature data;

[0057] Specifically, step 2) includes the following steps:

[0058] Step 2-1) Match lane width detection data, vehicle operation data, road alignment data, traffic marking data, road segment feature data, and environmental data to the same spatial coordinates based on timestamps, latitude and longitude, and highway station numbers;

[0059] Step 2-2) Based on the vehicle operation data, the instantaneous speed is taken as the vehicle speed; the distance between the vehicle and the nearest large vehicle in the current downstream lane or adjacent lane is calculated as the distance to the downstream large vehicle.

[0060] Steps 2-3) Based on the road alignment data and road design drawings, extract the average curvature and slope change rate of the downstream 200m.

[0061] Steps 2-4) Extract the category of traffic marking data. Traffic marking data includes longitudinal marking types, special marking types, and marking defects types. Among them, longitudinal marking types are divided into double-sided dashed lines, left-side dashed line and right-side solid line, left-side solid line and right-side dashed line, double-sided solid lines, acceleration lane transition section extension line, and deceleration lane transition section extension line; special marking types are divided into ordinary markings, arrow markings, vehicle distance confirmation markings, lateral deceleration markings, and text markings; marking defects types are divided into ordinary markings, marking wear, old marking overlap, and traffic cone coverage.

[0062] Steps 2-5) Extract the category of road segment feature data. The road segment feature data includes road segment type and lane position. Specifically, the road segment type is divided into general main road segment, entrance main road segment, and exit main road segment; and the lane position is divided into leftmost lane, rightmost lane, and middle lane.

[0063] Steps 2-6) Extract the category of the environmental data. The environmental data includes lighting and weather. Lighting is divided into daytime, nighttime without streetlights, nighttime with streetlights, dawn, and dusk. Weather is divided into sunny, cloudy, light rain, moderate rain, heavy to torrential rain, light snow, moderate snow, heavy to torrential snow, and fog.

[0064] Step 3) For the lane width detection data preprocessed in Step 2), lane width detection failure events are used as an indicator to measure lane detection performance and determine the type of lane detection failure.

[0065] Specifically, step 3) includes the following steps:

[0066] Step 3-1) Define the lane width detection failure event as follows: Assuming the autonomous vehicle is located on the center line of the detection lane and one lane line of the detection lane overlaps with the lane line of the actual lane, when the lane width detection value is too wide or too narrow and the vehicle crosses the actual lane line, it is a safety threshold situation for the autonomous vehicle to collide with the adjacent vehicle and is regarded as a lane width detection failure event.

[0067] Step 3-2) Calculate the lane width detection threshold for lane width detection failure events according to formula (1):

[0068] |W md -W a |+W v =W a (1)

[0069] Among them, W md W represents the lane width detection threshold. a W represents the actual lane width. v This refers to the width of the vehicle.

[0070] Step 3-3) Based on the lane width detection threshold W md Determine the type of lane detection failure.

[0071] According to formula (1), the lane width detection threshold W md The calculation result should include two values: the lower threshold and the upper threshold, which are denoted as W in this embodiment. md1 (Lower threshold) and W md2 (Upper threshold), where W md1 <W md2 Then, the lane detection failure type is determined according to the following formula (2):

[0072]

[0073] Among them, W d This refers to the lane width.

[0074] Step 4) Construct a Boosting ensemble learning model to learn the impact of input feature data on the lane detection performance of autonomous vehicles. Use the feature data as the model input and the lane detection failure type as the output to train the model.

[0075] Specifically, step 4) includes the following steps:

[0076] Step 4-1) Randomly divide 70% of the preprocessed data into the training set and 30% into the test set;

[0077] Step 4-2) Use the vehicle operation data, road alignment data, traffic marking data, road segment feature data, and environmental data from the training set as input features, and the lane detection failure type as output feature to train the Boosting ensemble learning model; evaluate the performance of the model training using the confusion matrix based on the test set data.

[0078] Step 5) Based on the Boosting ensemble learning model trained in Step 4), predict the lane detection failure type of the highway to be evaluated, and draw a conclusion on the highway drivability assessment of autonomous vehicles.

[0079] Specifically, step 5) includes the following steps:

[0080] Step 5-1) Take the vehicle operation data, road alignment data, traffic marking data, road segment feature data and environmental data of the highway to be evaluated as input features, input them into the trained Boosting ensemble learning model, and output the predicted value of lane detection failure type of autonomous vehicle in each road segment unit of the highway to be evaluated.

[0081] In order to ensure good predictive performance of the model, the highway to be evaluated should have similar characteristics in terms of terrain, climate and lane width to the highway used to train the Boosting ensemble learning model, and the road alignment and traffic markings should be consistent with the standards and specifications used in the design phase.

[0082] Step 5-2) Based on the lane detection failure type prediction value, determine the highway drivability assessment conclusion for autonomous vehicles. For road segment units with a prediction value of "failure", the assessment conclusion is "failure"; for road segment units with a prediction value of "not failed", the assessment conclusion is "pass".

[0083] In this embodiment, the method further includes an optimization process based on the evaluation results, specifically the following steps:

[0084] Step 6) Based on the Boosting ensemble learning model trained in Step 4), determine the global feature importance and feature dependencies.

[0085] Specifically, step 6) includes the following steps:

[0086] Step 6-1) Using the SHAP (Shapley additive explanations) method, based on the average absolute value of the SHAP of each input feature, we calculate the magnitude of the influence of the input feature on the model's predicted output, i.e., the global feature importance.

[0087] Step 6-2) To reveal the impact of a single input feature on the model's predicted output, a SHAP dependency graph of the single input feature is plotted. The positive and negative trends and magnitudes of the impact of changes in the input feature on the predicted output are analyzed. Positive impact means the predicted value is "ineffective" and negative impact means the predicted value is "not ineffective".

[0088] Step 6-3) Considering the interactive effects between input features, in order to reveal the pairwise interactive effects of a single input feature and other features on the model's predicted output, draw the interaction SHAP dependency graph of two input features, and analyze the positive and negative trends and magnitude changes in the influence of changes in the combination of input features on the predicted output.

[0089] Step 7) For highway segment units whose evaluation conclusion in Step 5) is "not passed", use the global feature importance and feature dependency relationship determined in Step 6) to formulate road design optimization strategies, determine the performance optimization focus of on-board sensors, and improve the lane detection performance of autonomous vehicles on highways.

[0090] Specifically, step 7) includes the following steps:

[0091] Step 7-1) Based on the global feature importance calculated in Step 6), determine the feature priority for the optimization strategy.

[0092] Step 7-2) Based on the feature dependency relationship constructed in Step 6), along the feature value range optimization direction where the positive influence of input features on the model prediction output is weakened or the negative influence is increased (i.e., the failure probability is reduced), from three dimensions—road alignment and traffic marking design optimization, autonomous vehicle operation design domain management, and on-board sensor performance optimization—to formulate countermeasures to improve the highway driving suitability of autonomous vehicles and ensure the safe operation of autonomous vehicles.

[0093] In summary, the method steps of this invention are briefly described as follows: First, acquire lane width detection data and vehicle operation data from the onboard sensors of an autonomous vehicle while it is driving on a highway. Second, determine the lane detection failure type based on the lane width detection threshold. Third, use vehicle operation data, road alignment data, traffic marking data, road segment feature data, and environmental data as input features, and the lane detection failure type as the output feature. Employ the Boosting ensemble learning algorithm to construct a model of the influence of the input feature data on the lane detection performance of the autonomous vehicle. Fourth, input the features of the highway to be evaluated into the trained model to predict the lane detection failure type and determine the autonomous vehicle's highway drivability assessment conclusion. Fifth, determine the global feature importance and feature dependencies based on the Shapley Additive Explanations (SHAP) method, formulate road design optimization strategies, and determine the key areas for onboard sensor performance optimization.

[0094] This embodiment conducted a 380-kilometer field test on a certain section of the main road of a certain expressway, collecting and acquiring the raw data described in step 1). To ensure a uniform distribution of traffic markings and road segment feature data samples, vehicles completed a single test journey in fixed lanes, with the total test journey covering the leftmost, rightmost, and middle lanes of the expressway.

[0095] After preprocessing in step 2), the raw data is used to obtain the input feature data required for modeling, including vehicle operation data (driving speed, distance to downstream trucks), road alignment data (average curvature of 200 meters downstream, slope change rate), traffic marking data (marking type, marker type, marking defects), road segment feature data (road segment type, lane position), and environmental data (lighting conditions).

[0096] According to step 3 of the present invention, determine the type of lane width detection failure event. Based on the road design drawings, test the actual value of the lane width (W) of the main road section. a All are 3.75 meters; to broaden the applicability of the assessment conclusions, the vehicle type is set as a single-row van, and the vehicle width (W) in formula (1) in step 3) is 3.75 meters. v The value is taken as 2.44 meters; therefore, the lane width detection threshold (W) md The widths are 5.06 meters and 2.44 meters respectively, as shown in formula (3). After processing in step 3), out of a total of 148,520 frames of lane width detection data, 2,398 frames were judged to be invalid, accounting for 1.6% of the total.

[0097]

[0098] According to step 4) of this invention, the XGBoost ensemble learning algorithm is used to construct a model of the influence of input feature data on the lane detection performance of autonomous vehicles. The model performance is evaluated, and its recall, precision, F1 score, and accuracy are 0.9906, 0.9916, 0.9911, and 0.9997 on the training set, and 0.9016, 0.9524, 0.9263, and 0.9975 on the test set, indicating excellent model performance.

[0099] According to step 5 of the present invention, the Beijing-Shanghai Expressway (Shanghai section) is taken as the expressway to be evaluated. Its vehicle operation data, road alignment data, traffic marking data, road segment feature data and environmental data are input into the trained model, and the lane width detection failure type prediction value of its road segment unit is output. A total of 626 frames are predicted to be failure.

[0100] According to steps 6)-7) of the present invention, the global features affecting lane detection performance, in descending order of importance, are: driving speed, road segment type, distance to downstream trucks, lane position, marking type, average curvature over 200 meters downstream, slope change rate, marking type, marking defects, and lighting conditions. Figure 2 This embodiment prioritizes feature importance based on the SHAP mean absolute value, focusing on design optimization and management constraints for the top-ranked features. Taking the impact of the single feature of slope change rate as an example... Figure 3This embodiment uses the SHAP dependency map of slope change rate features. When the slope change rate is greater than 10% / km (convex curve) or 20% / km (concave curve), it has a positive impact on the probability of lane width detection failure, and this probability gradually increases. Therefore, the alignment design should be optimized towards the range where the failure probability decreases. Taking the interaction between road segment type and lane position features as an example... Figure 4 This is an interactive SHAP dependency graph of road segment type and lane position features. When a vehicle is driving in the rightmost lane of the main road segment at the entrance and exit, the positive impact of lane width detection failure probability is the highest. This can constrain the operating design domain of autonomous vehicles that do not need to enter or exit the main road segment, and prevent them from driving in the rightmost lane of the main road segment at the entrance and exit.

[0101] Therefore, the embodiments of the present invention achieve a quantitative analysis of the impact of highway alignment design, traffic marking attributes, traffic operation and environmental characteristics on the lane detection performance of autonomous vehicles. Based on the global feature importance and the range of values ​​for the reduction of failure probability in the feature dependency relationship, optimization strategies for road design and on-board sensor performance are proposed, which will help improve the drivability of road segment units that have not passed the evaluation and solve the problem that the existing technology is difficult to quantitatively evaluate the drivability of autonomous vehicles on highways.

[0102] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for evaluating the drivability of autonomous vehicles on highways based on ensemble learning, characterized in that, Includes the following steps: Step 1) Collect driving data and corresponding traffic data of autonomous vehicles traveling on highways, wherein the driving data includes lane width detection data and vehicle operation data; Step 2) Preprocess the raw data collected in Step 1) and extract the categories from the categorical variables to obtain feature data; Step 3) For the lane width detection data preprocessed in Step 2), lane width detection failure events are used as an indicator to measure lane detection performance and determine the type of lane detection failure. Step 4) Construct a Boosting ensemble learning model to learn the impact of input feature data on the lane detection performance of autonomous vehicles. Use the feature data as the model input and the lane detection failure type as the output to train the model. Step 5) Based on the Boosting ensemble learning model trained in Step 4), predict the lane detection failure type of the highway to be evaluated, and draw a conclusion on the highway drivability assessment of autonomous vehicles. Step 6) Based on the Boosting ensemble learning model trained in Step 4), determine the global feature importance and feature dependencies; specifically, Step 6) includes the following steps: Step 6-1) Using the SHAP method, based on the average absolute value of each input feature, calculate the magnitude of the influence of the input feature on the model's predicted output, i.e., the global feature importance. Step 6-2) To reveal the impact of a single input feature on the model's predicted output, a SHAP dependency graph of the single input feature is plotted. The positive and negative trends and magnitudes of the impact of changes in the input feature on the predicted output are analyzed. Positive impact means the predicted value is "ineffective" and negative impact means the predicted value is "not ineffective". Step 6-3) Considering the interactive effects between input features, in order to reveal the pairwise interactive effects of a single input feature and other features on the model's predicted output, draw the interaction SHAP dependency graph of two input features, and analyze the positive and negative trends and magnitude of the influence of changes in the combination of input features on the predicted output. Step 7) For highway segment units that failed the drivability assessment in Step 5), use the global feature importance and feature dependency relationships determined in Step 6) to formulate road design optimization strategies, determine the key areas for performance optimization of onboard sensors, and improve the lane detection performance of autonomous vehicles on highways.

2. The method for evaluating the drivability of autonomous vehicles on highways based on ensemble learning according to claim 1, characterized in that, Step 1) specifically involves: collecting lane width detection data and vehicle operation data using onboard sensors; collecting road alignment data based on road design drawings; and collecting traffic marking data, road segment feature data, and environmental data using onboard video.

3. The method for evaluating the drivability of autonomous vehicles on highways based on ensemble learning according to claim 2, characterized in that, Step 2) includes the following steps: Step 2-1) Match lane width detection data, vehicle operation data, road alignment data, traffic marking data, road segment feature data, and environmental data into the same spatial coordinates based on timestamps, latitude and longitude, and highway station numbers; Step 2-2) Based on the vehicle operation data, the instantaneous speed is taken as the vehicle speed; the distance between the vehicle and the nearest large vehicle in the current downstream lane or adjacent lane is calculated as the distance to the downstream large vehicle. Steps 2-3) Based on the road alignment data and road design drawings, extract the average curvature and slope change rate of the downstream preset distance; Steps 2-4) Extract the category of traffic marking data. Traffic marking data includes longitudinal marking types, special marking types, and marking defects types. Among them, longitudinal marking types are divided into double-sided dashed lines, left-side dashed line and right-side solid line, left-side solid line and right-side dashed line, double-sided solid lines, acceleration lane transition section extension line, and deceleration lane transition section extension line; special marking types are divided into ordinary markings, arrow markings, vehicle distance confirmation markings, lateral deceleration markings, and text markings; marking defects types are divided into ordinary markings, marking wear, old marking overlap, and traffic cone coverage. Steps 2-5) Extract the category of the road segment feature data. The road segment feature data includes road segment type and lane position. The road segment type is divided into general main road segment, entrance main road segment, and exit main road segment; the lane position is divided into leftmost lane, rightmost lane, and middle lane. Steps 2-6) Extract the category of the environmental data. The environmental data includes lighting and weather. Lighting is divided into daytime, nighttime without streetlights, nighttime with streetlights, dawn, and dusk. Weather is divided into sunny, cloudy, light rain, moderate rain, heavy to torrential rain, light snow, moderate snow, heavy to torrential snow, and fog.

4. The method for evaluating the drivability of autonomous vehicles on highways based on ensemble learning according to claim 1, characterized in that, Step 3) includes the following steps: Step 3-1) Define the lane width detection failure event as follows: Assuming the autonomous vehicle is located on the center line of the detection lane and one lane line of the detection lane overlaps with the lane line of the actual lane, when the lane width detection value is too wide or too narrow and the vehicle crosses the actual lane line, it is a critical safety situation for the autonomous vehicle to collide with the adjacent vehicle, and is regarded as a lane width detection failure event. Step 3-2) Calculate the lane width detection threshold for lane width detection failure events, wherein the lane width detection threshold includes a lower threshold and an upper threshold; Step 3-3) Determine the lane detection failure type based on the lane width detection threshold. When the lane width is greater than or equal to the lower threshold and less than or equal to the upper threshold, the lane detection failure type is not failed; otherwise, the lane detection failure type is failed.

5. The method for evaluating the drivability of autonomous vehicles on highways based on ensemble learning according to claim 4, characterized in that, The formula for calculating the lane width detection threshold is: in, W md This indicates the lane width detection threshold. W a This represents the actual lane width. W v This refers to the width of the vehicle.

6. The method for evaluating the drivability of autonomous vehicles on highways based on ensemble learning according to claim 2, characterized in that, Step 4) includes the following steps: Step 4-1) Divide the preprocessed data into training and testing sets; Step 4-2) Use the vehicle operation data, road alignment data, traffic marking data, road segment feature data, and environmental data of the training set as input features, and the lane detection failure type as output features to train the Boosting ensemble learning model; use the confusion matrix to evaluate the performance of the model training based on the test set data.

7. The method for evaluating the drivability of autonomous vehicles on highways based on ensemble learning according to claim 2, characterized in that, Step 5) includes the following steps: Step 5-1) Take the vehicle operation data, road alignment data, traffic marking data, road segment feature data and environmental data of the highway to be evaluated as input features, input them into the trained Boosting ensemble learning model, and output the predicted value of lane detection failure type of autonomous vehicle in each road segment unit of the highway to be evaluated. Step 5-2) Based on the lane detection failure type prediction value, determine the highway drivability assessment conclusion for autonomous vehicles. For road segment units with a prediction value of "failure", the assessment conclusion is "failure"; for road segment units with a prediction value of "not failed", the assessment conclusion is "pass".

8. The method for evaluating the drivability of autonomous vehicles on highways based on ensemble learning according to claim 1, characterized in that, Step 7) includes the following steps: Step 7-1) Based on the global feature importance calculated in Step 6), determine the feature priority for the optimization strategy. Step 7-2) Based on the feature dependency relationship constructed in Step 6), along the feature value range optimization direction where the positive influence of input features on the model prediction output is weakened or the negative influence is increased, formulate countermeasures to improve the highway driving suitability of autonomous vehicles from three dimensions: road alignment and traffic marking design optimization, autonomous vehicle operation design domain management, and on-board sensor performance optimization, so as to ensure the safe operation of autonomous vehicles.

Images