A method for evaluating driving ability of a driver

Abstract

A method for assessing driver ability is presented, belonging to the field of driving behavior detection and assessment technology. It addresses the technical problems of existing driving ability assessment technologies, such as limited assessment dimensions and the lack of specific models for individuals with impaired neurological function. By constructing a multimodal data acquisition platform involving humans, vehicles, and the environment, driving behavior data is collected under preset simulated driving conditions. Statistical feature values ​​are extracted, driving behavior distribution feature values ​​are constructed, and cluster analysis and classification neural networks are used to complete the driving ability level assessment. Corresponding driving suggestions and auxiliary prompts are automatically output, thereby achieving a scientific, systematic, and multidimensional assessment of the ability of individuals with impaired neurological function to regain driving ability.

Description

Technical Field

[0001] This invention belongs to the field of driving behavior detection and evaluation technology, specifically relating to a method for evaluating a driver's driving ability. Background Technology

[0002] Driving is a complex cognitive and motor skill that places high demands on a driver's neurological function, cognitive ability, reaction speed, visual attention, and physical coordination. Existing driving ability assessment technologies mainly include the following categories: (1) Identification method based on a single driving behavior parameter Existing research often uses single or limited vehicle dynamic parameters such as steering wheel angle, angular velocity, vehicle speed, and acceleration to build driving behavior recognition models for identifying behaviors such as lane keeping, lane changing, and braking. These methods are simple to implement, but they are mainly aimed at ordinary drivers and often rely on specific scenarios, making them difficult to apply to people with impaired neurological function.

[0003] (2) Classification method based on driving style quantification Some existing methods classify driving styles using indicators such as rapid acceleration, rapid deceleration, headway, and collision time, typically categorizing drivers as aggressive, moderate, or cautious. While these methods can reflect differences in driving habits, their indicator systems are difficult to tailor to specific rehabilitation populations.

[0004] (3) Driving ability assessment methods based on psychological or physiological tests Some studies use reaction time tests, selective attention tests, and eye-hand coordination tests to evaluate driving ability, while others incorporate physiological indicators such as eye movements and electroencephalograms (EEGs). These methods can reflect some driving-related functions, but most are not combined with complete driving task scenarios and lack joint modeling of the entire process of human-vehicle-environment interaction data.

[0005] The shortcomings of existing driving ability assessment technologies: (1) Single evaluation dimension Existing technologies are mostly based on neurological function tests, subjective judgments, or a limited number of driving parameters, which makes it difficult to fully reflect a driver's comprehensive ability in real driving tasks.

[0006] (2) Lack of specific models for people with impaired neurological function The driving behaviors of normal drivers and people with impaired neurological function differ, and existing models designed for ordinary drivers are unable to accurately identify abnormal driving patterns in the group of patients undergoing rehabilitation.

[0007] (3) Difficulty in handling the randomness and heterogeneity of driving behavior Due to differences in the location of injury, degree of recovery, and type of functional impairment, the driving behavior of people with impaired neurological function fluctuates greatly, and a single numerical feature is insufficient to fully characterize their behavioral patterns.

[0008] (4) Insufficient secure and controllable data acquisition system Real vehicle testing is characterized by high risks, uncontrollable scenarios, and difficulty in reproducibility, while existing simulation tests lack systematic test condition design and standardized data collection processes. Summary of the Invention

[0009] To address the aforementioned issues, this invention proposes a method for assessing a driver's driving ability. By constructing a multimodal data acquisition platform involving humans, vehicles, and the environment, driving behavior data is collected under preset simulated driving conditions. Statistical feature values ​​are extracted, driving behavior distribution feature values ​​are constructed, and cluster analysis and classification neural networks are used to complete the assessment of driving ability levels. Corresponding driving suggestions and auxiliary prompts are automatically output, thereby achieving a scientific, systematic, and multidimensional assessment of the ability of individuals with impaired neurological function to regain their driving ability.

[0010] The method is specifically as follows: S1. Construct a multimodal data acquisition platform for people, vehicles, and the environment: The platform includes a hardware system and a software system; S2. Set up simulated driving conditions and collect multimodal driving data: Construct multiple preset driving task scenarios, with subjects in each scenario including a driver test group and a healthy control group matched for age and gender; collect time-series data of driving behavior, physiological signal data and IMU inertial data for each scenario; preprocess the data and output a standardized dataset with uniform structure and time alignment; S3. Deep generative clustering for multimodal driving data: Improve the hyperspherical variational autoencoder, input the multimodal driving data of drivers into the improved encoder for deep clustering, and output the subtype label of each driver and its belonging probability as the input features of the subsequent driving ability classification and evaluation model. S4. Construct a driving ability classification and evaluation model: Improve the BBQ-Tree model to obtain a hierarchical decision model with multimodal constraints as the driving ability classification and evaluation model. Input the features obtained in step S3 into the driving ability classification and evaluation model. A continuous score value in the interval [0,1] represents the comprehensive score of driving ability. S5. Output evaluation results: Map the comprehensive score of driving ability to a driving ability level through a preset threshold and output it.

[0011] Furthermore, the hardware system includes driving control devices, data acquisition equipment, and computing processing units; The software system includes a vehicle dynamics simulation module, a virtual scene generation module, and a data management module; The driving control device includes a steering wheel, accelerator pedal, and brake pedal; the data acquisition equipment includes vehicle status sensors, eye trackers, physiological monitors, or other devices used to collect driver status information; the computing and processing unit is used to synchronously receive, process, and store various types of data.

[0012] Furthermore, several preset driving task scenarios include: On a one-way, two-lane straight section of a highway, a vehicle in the adjacent lane suddenly cuts into this lane; On a two-way, two-lane straight section of an urban road, the vehicle in front brakes suddenly. An obstacle has appeared ahead on a two-lane, two-way straight section of an urban road. The two-way, two-lane left-turn section of the suburban road has poor visibility and oncoming traffic. At a signalized intersection with four lanes in both directions on an urban road, a vehicle in the adjacent lane suddenly decelerates. Rural roads are two-way, two-lane roads without signalized intersections, with vehicles coming from the right in the transverse lanes.

[0013] Furthermore, the hyperspherical variational autoencoder is improved as follows: At the model structure level, a multi-branch coding network is added to extract features from the three types of data respectively, and then input them into the hyperspherical variational autoencoder. Subsequently, a cross-modal feature fusion mechanism is introduced to map the three types of features to a unified latent space, and an attention mechanism is used to adaptively assign importance weights for different modalities. In terms of training strategy, the original loss function is enhanced by constructing the following multi-objective optimization mechanism: Alignment loss: Forces the clustering latent variable Zc corresponding to two views of the same driver to be closer to each other on the sphere, making the same type of driving behavior more compact; Uniformity loss: Force the clustering latent variable Zc of different drivers to be uniformly distributed on a sphere, so as to maintain the separation between different types of driving behavior; Assigning consistency loss: forcing the clustering probability distributions of two views corresponding to the same driver to remain consistent; Consistency loss of probability distribution: ensures that the clustering probability distribution of the same driving sample is consistent in different modalities.

[0014] Furthermore, dynamic driving features are extracted from the temporal data of driving behavior through a temporal coding network; Physiological signal data are extracted using a frequency-time domain joint coding module to reveal features reflecting driving load and state. IMU data is used to extract vehicle motion stability and operational characteristics through an inertial feature encoder.

[0015] Furthermore, the BBQ-Tree model is improved by adding stability constraint split nodes to the internal nodes of the BBQ-Tree model, which are applied to IMU data to characterize the impact of vehicle operational stability on driving ability.

[0016] Furthermore, stability-constrained split nodes participate in the competitive selection of candidate split nodes. Their splitting criteria introduce a stability constraint term based on traditional information gain to characterize the vehicle motion stability features reflected by IMU data. When splitting a node, the classification discrimination ability corresponding to the candidate splitting method and the fluctuation degree of IMU data in each child node after splitting are calculated simultaneously. When the splitting can reduce the variance of intra-class IMU data while ensuring class discrimination, the stability-constrained split node is selected as the optimal splitting method. During the model pruning process, if the child node formed based on the stability-constrained split node decreases by less than a preset threshold compared to the original node, the branch is pruned.

[0017] Furthermore, the overall score of driving ability is mapped to a driving ability level as follows: Rating ≥ 0.7: Suitable for driving; 0.4 ≤ Score < 0.7: Driver assistance required; Rating <0.4: Not suitable for driving.

[0018] The beneficial effects of the method are as follows: (1) A multimodal data acquisition platform integrating "human-vehicle-environment" was constructed, which can simultaneously collect vehicle dynamic parameters, driver physiological signals and driving environment information, overcoming the limitations of traditional assessment methods that rely solely on single neurological function examinations or subjective judgments, and significantly improving the scientificity, comprehensiveness and objectivity of the assessment results.

[0019] (2) A deep generative clustering method based on an improved superspherical variational autoencoder (ADS-VAE) is adopted. In view of the heterogeneity of driving behavior of patients with impaired neurological function, the method achieves accurate clustering of driving behavior patterns by decoupling latent variables, modeling spherical distribution and geometric regularization. The number of clusters is automatically determined without manual preset. Compared with traditional clustering algorithms, it can better capture the behavioral patterns of this group.

[0020] (3) The traditional BBQ-Tree model cannot effectively model the differences and synergistic relationships between different modal data, and fails to make full use of the driving behavior clustering results obtained in deep generative clustering. Therefore, this invention improves upon the BBQ-Tree by adding stability constraint split nodes to characterize the impact of vehicle operation stability on driving ability. Traditional BBQ-Tree methods primarily target a single feature space for decision partitioning, failing to effectively address the differences in data dimensionality, statistical distribution, and time scale among driving behavior data, physiological signal data, and IMU data. This invention employs a modality-aware splitting mechanism, enabling different modalities to utilize differentiated splitting strategies. This avoids the weakening of important features caused by traditional uniform splitting methods, thereby improving the overall utilization efficiency of multi-source driving data. By introducing the latent variable Zc for driving behavior clustering into the high-level structure of the decision tree, different driving behavior patterns correspond to different subtree structures. Compared with the traditional BBQ-Tree method that uses a uniform decision path for all samples, this approach can dynamically adjust the classification logic based on differences in driving style, thereby reducing feature overlap between different driving behaviors and improving the relevance and consistency of driving ability classification results. This invention introduces a modality-weighted information gain metric during node splitting and combines it with an uncertainty-based pruning mechanism to adaptively prune low-contribution branches. Compared to the traditional BBQ-Tree method that prunes solely based on classification error, this invention reduces excessive splitting caused by local noise or outliers, improving the model's generalization ability across different driver groups and testing scenarios. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall process of the method described in this embodiment of the invention. Figure 2 This is a flowchart of the deep generative clustering process in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the driving ability classification and evaluation model in an embodiment of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0023] This embodiment provides a method for assessing a driver's driving ability. By constructing a multimodal data acquisition platform for people, vehicles, and the environment, driving behavior data is collected under preset simulated driving conditions. Statistical feature values ​​are extracted, driving behavior distribution feature values ​​are constructed, and cluster analysis and classification neural networks are used to complete the assessment of driving ability levels. This enables a scientific, systematic, and multidimensional assessment of the ability to return to driving for people recovering from neurological impairment (such as stroke survivors).

[0024] like Figure 1 As shown, a general flowchart of the method described in this embodiment is presented, combined with... Figure 1 This method will be described in further detail.

[0025] Step 1: Construct a multimodal data acquisition platform for people, vehicles, and the environment. The platform includes a hardware system and a software system.

[0026] The hardware system includes: driving control devices, data acquisition equipment, and computing processing units; The software system includes: a vehicle dynamics simulation module, a virtual scene generation module, and a data management module.

[0027] The driving control devices include a steering wheel, accelerator pedal, and brake pedal; the data acquisition equipment includes vehicle status sensors, eye trackers, physiological monitors, or other devices used to collect driver status information; and the computing and processing unit is used to synchronously receive, process, and store various types of data.

[0028] The hardware components of the driving simulator include a high-fidelity force feedback system, the SensoWheel torque steering wheel, and a Kistler MSW (Measurement Steering Wheel) steering wheel sensor to measure steering wheel angle, speed, torque, and other signals. Potentiometers were also used to modify the accelerator and brake pedals, providing longitudinal control for the driver. A Peak PCAN-PCI Express FD board and an NI PCIe-6363 board were installed in an industrial computer to achieve CAN communication and analog signal acquisition.

[0029] The software system primarily consists of a co-simulation platform comprised of MATLAB / Simulink, Mechanical Simulation CarSim, and SIEMENS Simcenter PreScan. This platform runs in real-time on a host computer using the Simulink Desktop Real-Time toolkit, providing drivers with a real-time visual traffic environment while simultaneously collecting experimental data. Specifically, it comprises the following three parts: Scene Generation and Management: PreScan software was used to create a shared virtual city environment, including complex road layouts, dynamic traffic flow, variable weather conditions, and various obstacles. PreScan was also used to create the driver's view and the left and right rearview mirror views, which were then displayed on the monitor using the VisViewer plugin.

[0030] Vehicle Dynamics and Behavior Simulation: Using CarSim, the physical behavior of each vehicle is simulated, providing accurate simulations and dynamic responses of vehicle dynamic performance and driving characteristics. During modeling, multiple key factors are considered, including the suspension system, steering system, vehicle mass distribution, and tire characteristics, making the simulation results more realistic and reliable, ensuring that vehicle responses are consistent with the real world. Simultaneously, each vehicle in the simulator can interact in real time based on driver input, including complex behaviors such as lane changing, overtaking, and emergency braking.

[0031] Collaborative Control and Communication Platform: A central control and coordination system developed based on MATLAB / Simulink is responsible for managing the synchronized operation of all simulators, processing driver input, simulated vehicle output, and environmental change information, and achieving precise time synchronization and data exchange among multiple driving simulators. Specifically, Simulink is used to model the driving pedal and steering wheel mechanisms of the driving simulators, converting hardware input signals into the main vehicle's steering angle, throttle opening, and braking pressure input signals from the vehicle dynamics model, enabling real-time data interaction.

[0032] Step 2: Set up simulated driving conditions and collect test data Construct multiple preset driving mission scenarios in the driving simulator, including at least: 1) On a one-way two-lane straight section of a highway, a vehicle in the adjacent lane suddenly cuts into this lane; 2) On a two-way, two-lane straight section of an urban road, the vehicle in front brakes suddenly; 3) An obstacle appears ahead on a two-lane, two-way straight section of an urban road; 4) Two-way, two-lane left-turn sections of suburban roads with poor visibility and oncoming traffic; 5) At a signalized intersection with two lanes in both directions on an urban road, a vehicle in the adjacent lane suddenly decelerates, and its forward view is obstructed by parked vehicles, buildings or road facilities, causing pedestrians or non-motorized vehicles to suddenly enter the vehicle's path from the obstructed area. 6) Rural roads with two lanes in both directions and no signalized intersections, with vehicles coming from the right in the transverse lane.

[0033] The subjects included a stroke survivor trial group and an age- and sex-matched healthy control group.

[0034] The collected data includes: driving behavior time-series data, physiological signal data, and IMU inertial data. Step 3: Data Preprocessing First, the raw data undergoes format conversion. Data stored in .mat files in Matlab's timesseries structure is parsed into standard numerical arrays and re-saved as structured .mat files. This data is then exported to the universal .csv format to achieve cross-platform data compatibility and subsequent analysis. During the conversion process, the one-to-one correspondence between timestamps and signals from each channel is maintained to ensure data integrity. Next, to address the issue of fluctuating sampling frequencies across different experimental scenarios, a unified time benchmark is implemented. This is achieved by extracting the original sampling time interval sequence, calculating its mean and standard deviation, and removing outlier sampling intervals based on the 3σ criterion. This determines the stable sampling time step for each scenario and further synthesizes a unified time benchmark applicable to multiple scenarios. Finally, a scenario anchor point positioning mechanism is introduced. Based on the trigger time information marked in external record files (.xlsx files), the occurrence times of key events in each scenario are read and mapped to corresponding positions in the time series data using time matching or interpolation methods, achieving alignment and fusion of multi-source data. Centered on the trigger moment, a fixed time window is set with a unified time step. The window is expanded forward and backward by a preset duration, and the corresponding number of data points is calculated. Complete and consistent-length scene segments are extracted from the original data, ensuring consistency in time scale across different samples. Finally, the extracted data segments are standardized. This includes downsampling the approximately 1000 Hz original high-frequency data to approximately 200 Hz after anti-aliasing filtering to reduce data redundancy and improve processing efficiency. Outliers, missing values, and noise are cleaned from the data. For example, outliers are removed using statistical methods, missing data is filled using interpolation, and noise is smoothed and denoised using filtering methods. Data is also normalized or standardized as needed to eliminate the influence of different units of measurement. After these steps, a standardized dataset with a unified structure, time alignment, and reliable quality is output, providing a stable data foundation for subsequent feature extraction, model training, and driving behavior analysis.

[0035] Step 4: Deep Generative Clustering Method for Multimodal Driving Data Given the multi-source heterogeneity (driving behavior time-series data, physiological signal data, and IMU inertial data), strong time dependence, and significant individual differences in the data processed in this invention, traditional deep clustering methods based on hyperspherical variational autoencoders (ADS-VAE) are difficult to apply directly. Specifically, the original method mainly focuses on modeling single feature distributions, lacking the ability to effectively model structural differences and cross-modal correlations between multimodal data, and also failing to consider the impact of temporal dynamic features on clustering stability. Therefore, this invention makes targeted improvements to ADS-VAE, proposing a multimodal time-aware deep generative clustering method, the process of which is as follows: Figure 2 As shown.

[0036] At the model structure level, a multi-branch coding network is constructed to extract features from the three types of data respectively: (1) Dynamic driving features are extracted from time-series data of driving behavior through a time-series coding network; (2) Physiological signal data are extracted using a frequency-time domain joint coding module to reflect the characteristics of driving load and state; (3) The IMU data is used to extract the vehicle's motion stability and operational characteristics through an inertial feature encoder.

[0037] Subsequently, a cross-modal feature fusion mechanism is introduced to map the three types of features to a unified latent space, and an attention mechanism is used to adaptively allocate the importance weights of different modalities, thereby improving the ability to express the essential features of driving behavior.

[0038] An encoder network consisting of fully connected layers is constructed to map the input features to two latent variables located on a hypersphere: a clustering latent variable Zc and a generative latent variable Zg. The clustering latent variable Zc is used to capture the driver's behavior type, and its distribution is a von Mises-Fisher (vMF) distribution, which is a statistical model that specifically describes the clustering of data points on a sphere around a central direction. The generative latent variable Zg is used to capture individual differences unrelated to driving type, and its distribution is a Power Spherical (PS) distribution, which is a statistical model suitable for stable sampling on a sphere.

[0039] In terms of training strategy, the original loss function is enhanced by constructing the following multi-objective optimization mechanism: Alignment loss: Forces the Zc corresponding to two views of the same driver to be close to each other on the sphere, making the same type of driving behavior more compact; Uniformity loss: Force different drivers' Zc to be uniformly distributed on the spherical surface to maintain separation between different types of driving behavior; Assigning consistency loss: This forces the clustering probability distributions of two views corresponding to the same driver to remain consistent, thereby enhancing the stability of clustering.

[0040] Probability distribution consistency loss: ensures that the cluster probability distribution of the same driving sample is consistent across different modalities; The model's total loss function consists of two parts: first, the variational lower bound (ELBO), which ensures the model can accurately reconstruct the original data; and second, the weighted sum of the four geometric regularization terms (multi-objective optimization mechanisms) mentioned above, which optimizes the clustering structure. By minimizing the total loss, the model can simultaneously learn an effective representation of driving data and clustering.

[0041] After training, the model performs a posterior entropy pruning on a separate validation dataset. This algorithm progressively removes the vMF mixture components with the smallest contribution and calculates the normalized entropy (a metric for clustering uncertainty) of the remaining components. When the entropy value reaches its minimum, the corresponding number of components is the final determined number of driving behavior subtypes K, thus achieving automatic determination of the number of clusters without manual pre-setting. Finally, the model outputs the subtype label and its probability for each driver, serving as input features for subsequent driving ability classification and evaluation models.

[0042] Step 5: Construct a driving ability classification and assessment model Given the multimodal heterogeneity (driving behavior, physiological signals, and IMU data), the presence of prior clustering information, and the significant physiological-behavioral coupling characteristics of stroke patients, the traditional BBQ-Tree model cannot effectively model the differences and synergistic relationships between different modalities of data, and it fails to fully utilize the driving behavior clustering results obtained in step 4. Therefore, this invention improves upon the BBQ-Tree model by proposing a multimodal constrained hierarchical decision-making model, such as... Figure 3 As shown, this is used to grade and assess the driving ability of stroke survivors.

[0043] First, the driving behavior clustering latent variable Zc obtained in step 4 is used as the high-level structure guiding variable to initially group the samples, and corresponding sub-decision tree structures are constructed based on different driving behavior patterns, thereby realizing the transformation from a "unified decision model" to a "differentiated decision model".

[0044] Construct a BBQ-Tree model. This model is a tree structure, and its internal nodes are divided into four types: Quantum logic splitting (Q-split): These nodes do not perform simple threshold checks, but directly output the input feature value itself, representing a "continuous" membership relationship. Its physical meaning is: the higher the feature value, the greater its contribution to the final driving ability score; Boolean logic split (B-split): This type of node is the same as the traditional decision tree, comparing the feature value with a learnable threshold and outputting 0 or 1, which is used to achieve fine-grained decision boundary division; Fixed-threshold Boolean split (B¹ / ²-split): This is a special case of B-split, with a fixed threshold of 0.5, used to simplify the model structure.

[0045] Stability Constraint Split Node (S-split): Primarily used with IMU data to characterize the impact of vehicle operational stability on driving ability.

[0046] BBQ-Tree is trained by selecting the optimal splitting method layer by layer to construct the tree structure: Phase 1: Prioritizes the use of Q-split and B¹ / ²-split to construct the trunk of the tree. This phase focuses on interpretability, revealing the trend relationship between features and driving ability through Q-split and achieving basic class distinctions through B¹ / ²-split. The maximum depth of this phase is controlled by the hyperparameter depth_BQ, typically set to 3 to 5 layers to ensure a concise and easily understandable trunk structure. The second stage involves adding several B-split layers at the bottom of the main tree for fine-tuning accuracy. The maximum depth of this stage is controlled by the hyperparameter `depth_B`, typically set to 2 to 3 layers. Since B-splits are located at the lower levels of the tree, their decision-making logic is relatively complex, but they are only used to optimize classification accuracy and do not affect the interpretability of the upper-level main tree. Stability-constrained split nodes (S-splits) participate in the competitive selection of candidate split nodes. Their splitting criterion introduces a stability constraint term based on traditional information gain to characterize the vehicle motion stability features reflected by the IMU data. During node splitting, the classification discrimination ability corresponding to the candidate splitting method and the fluctuation degree of IMU features within each child node after splitting are calculated simultaneously. When the splitting can reduce the variance of intra-class IMU features while maintaining class discrimination, the S-split is preferentially selected as the optimal splitting method. Furthermore, during model pruning, if the percentage decrease of the child node formed based on the S-split compared to the original node is less than a preset threshold (5%-10%), the branch is pruned to control model complexity and improve generalization ability.

[0047] Step 6: Output evaluation results The output of step 5 is a continuous score value within the range [0,1], representing the overall driving ability score. This score is then mapped to a driving ability level using a preset threshold. Rating ≥ 0.7: Suitable for driving; 0.4 ≤ Score < 0.7: Driver assistance required; Rating <0.4: Not suitable for driving.

Claims

1. A method for assessing a driver's driving ability, characterized in that, The method includes the following steps: S1. Construct a multimodal data acquisition platform for people, vehicles, and the environment: The platform includes a hardware system and a software system; S2. Set up simulated driving conditions and collect multimodal driving data: Construct multiple preset driving task scenarios, with subjects in each scenario including a driver test group and a healthy control group matched for age and gender; collect time-series data of driving behavior, physiological signal data and IMU inertial data for each scenario; preprocess the data and output a standardized dataset with uniform structure and time alignment; S3. Deep generative clustering for multimodal driving data: Improve the hyperspherical variational autoencoder, input the multimodal driving data of drivers into the improved encoder for deep clustering, and output the subtype label of each driver and its belonging probability as the input features of the subsequent driving ability classification and evaluation model. S4. Construct a driving ability classification and evaluation model: Improve the BBQ-Tree model to obtain a hierarchical decision model with multimodal constraints as the driving ability classification and evaluation model. Input the features obtained in step S3 into the driving ability classification and evaluation model. A continuous score value in the interval [0,1] represents the comprehensive score of driving ability. S5. Output evaluation results: Map the comprehensive score of driving ability to a driving ability level through a preset threshold and output it.

2. The method for assessing a driver's driving ability according to claim 1, characterized in that, The hardware system includes driving control devices, data acquisition equipment, and computing processing units; The software system includes a vehicle dynamics simulation module, a virtual scene generation module, and a data management module; The driving control device includes a steering wheel, accelerator pedal, and brake pedal; the data acquisition device includes vehicle status sensors, eye trackers, physiological monitors, or other devices used to collect driver status information. The computing processing unit is used to synchronously receive, process, and store various types of data.

3. The method for assessing a driver's driving ability according to claim 2, characterized in that, Several preset driving task scenarios include: On a one-way, two-lane straight section of a highway, a vehicle in the adjacent lane suddenly cuts into this lane; On a two-way, two-lane straight section of an urban road, the vehicle in front brakes suddenly. An obstacle has appeared ahead on a two-lane, two-way straight section of an urban road. The two-way, two-lane left-turn section of the suburban road has poor visibility and oncoming traffic. At a signalized intersection with four lanes in both directions on an urban road, a vehicle in the adjacent lane suddenly decelerates. Rural roads are two-way, two-lane roads without signalized intersections, with vehicles coming from the right in the transverse lanes.

4. The method for assessing a driver's driving ability according to claim 3, characterized in that, The specific improvements to the hyperspherical variational autoencoder are as follows: At the model structure level, a multi-branch coding network is added to extract features from the three types of data respectively, and then input them into the hyperspherical variational autoencoder. Subsequently, a cross-modal feature fusion mechanism is introduced to map the three types of features to a unified latent space, and an attention mechanism is used to adaptively assign importance weights for different modalities. In terms of training strategy, the original loss function is enhanced by constructing the following multi-objective optimization mechanism: Alignment loss: Forces the clustering latent variable Zc corresponding to two views of the same driver to be closer to each other on the sphere, making the same type of driving behavior more compact; Uniformity loss: Force the clustering latent variable Zc of different drivers to be uniformly distributed on a sphere, so as to maintain the separation between different types of driving behavior; Assigning consistency loss: forcing the clustering probability distributions of two views corresponding to the same driver to remain consistent; Consistency loss of probability distribution: ensures that the clustering probability distribution of the same driving sample is consistent in different modalities.

5. The method for assessing a driver's driving ability according to claim 4, characterized in that, Dynamic driving features are extracted from time-series driving behavior data using a time-series coding network. Physiological signal data are extracted using a frequency-time domain joint coding module to reveal features reflecting driving load and state. IMU data is used to extract vehicle motion stability and operational characteristics through an inertial feature encoder.

6. The method for assessing a driver's driving ability according to claim 5, characterized in that, The improvement to the BBQ-Tree model is as follows: a stability constraint split node is added to the internal nodes of the BBQ-Tree model, which acts on the IMU data to characterize the impact of vehicle operational stability on driving ability.

7. The method for assessing a driver's driving ability according to claim 6, characterized in that, Stability-constrained split nodes participate in the competitive selection of candidate split nodes. The splitting criterion introduces a stability constraint term on the basis of traditional information gain to characterize the vehicle motion stability features reflected by IMU data. When splitting a node, the classification discrimination ability corresponding to the candidate splitting method and the fluctuation degree of IMU data in each child node after splitting are calculated simultaneously. When the splitting can reduce the variance of intra-class IMU data while ensuring class discrimination, the stability-constrained split node is selected as the optimal splitting method. During the model pruning process, if the child node formed based on the stability-constrained split node decreases by less than a preset threshold compared to the original node, the branch is pruned.

8. The method for assessing a driver's driving ability according to claim 7, characterized in that, The overall driving ability score is mapped to a driving ability level as follows: Rating ≥ 0.7: Suitable for driving; 0.4 ≤ Score < 0.7: Driver assistance required; Rating <0.4: Not suitable for driving.

Images