CN-120783322-B - Driver behavior evaluation method and system based on driving scene

CN120783322BCN 120783322 BCN120783322 BCN 120783322BCN-120783322-B

Abstract

The invention relates to the technical field of driving safety, in particular to a driving scene-based driver behavior assessment method and a driving scene-based driver behavior assessment system, which comprise the steps of breaking through the limitation of a single data source and improving the comprehensiveness and accuracy of accident risk prediction by integrating multidimensional heterogeneous data such as weather, roads, facilities and the like; the method comprises the steps of dynamically calibrating driving scene labels, enabling a model to adapt to complex and changeable actual driving environments, enhancing generalization capability, reducing subjectivity of manual threshold setting through dividing and feature quantification of traffic states, helping to find hidden features ignored by a traditional method, generating driver behavior optimization suggestions under multiple constraints through optimizing a scene-specific driver behavior evaluation model, reducing traffic flow fluctuation caused by excessive speed limiting, reducing accident probability of high-risk scenes, and achieving cooperative optimization of safety and efficiency.

Inventors

PENG KUN
XU LIJIE
YAN JIANCAI
TANG WEI
LI CHUNLEI
XIE HAIPING
WANG XUAN

Assignees

深圳市综合交通与市政工程设计研究总院有限公司

Dates

Publication Date: 20260512
Application Date: 20250513

Claims (5)

1. A driver behavior evaluation method based on driving scenes is characterized by comprising the following steps of: S1, under a free flow state, fusing multi-source heterogeneous basic data, generating an accident risk prediction value, calibrating a dynamic driving scene tag, and constructing a scene-specific driver behavior evaluation model by combining with the driver safety vehicle speed preference obtained by SP investigation, wherein the multi-source heterogeneous basic data comprises road geometry linear parameters, traffic facility type configuration, real-time environment meteorological data and dynamic traffic flow state which are synchronously obtained from a vehicle-mounted sensor, a road side unit and meteorological monitoring equipment; The multi-source heterogeneous basic data are fused, an accident risk predicted value is generated, a dynamic driving scene label is calibrated, and the process of constructing a scene-specific driver behavior evaluation model by combining the driver safety vehicle speed preference obtained by SP investigation comprises the following steps: Synchronously acquiring four types of multi-source heterogeneous basic data, namely road geometric linear parameters, traffic facility type configuration, real-time environmental meteorological data and dynamic traffic flow states, from a vehicle-mounted sensor, a road side unit and meteorological monitoring equipment; The method comprises the steps of preprocessing the four acquired multi-source heterogeneous basic data, carrying out normalization processing on road geometric line parameters, carrying out classified coding on traffic facility type configuration according to functions, namely, forbidden mark=1, warning mark=2 and indication mark=3, grading and weighting real-time environmental meteorological data according to risk levels, and carrying out normalization processing on dynamic traffic flow states; The method comprises the steps of establishing association between a dynamic traffic flow state and three other types of heterogeneous basic data, associating the dynamic traffic flow state with road geometric linear parameters by calculating linear sensitivity coefficients, associating the linear sensitivity coefficients with traffic facility type configuration by constructing facility influence matrixes, associating the linear sensitivity coefficients with real-time environment weather data by introducing weather correction factors, wherein the linear sensitivity coefficients are obtained by calculating the product of the inverse of the road geometric linear parameters and dynamic traffic flow state normalization values, obtaining facility influence matrixes by training historical data, reflecting the suppression coefficients of traffic facility types on the dynamic traffic flow state normalization values, and obtaining the weather correction factors by calculating the product of risk classification weights of the real-time environment weather data and the dynamic traffic flow state normalization values; adopting a graph neural network to fusion analyze four types of multi-source heterogeneous basic data and generating an accident risk prediction value; Calibrating dynamic driving scene labels, namely a safety scene, a general scene, a risk scene and a dangerous scene, based on accident risk prediction values obtained by the graph neural network; Based on the generated dynamic driving scene label, acquiring safety vehicle speed preference data of the experienced drivers in different driving scenes through SP investigation; based on the generated dynamic driving scene tag, the safety vehicle speed preference data of the experienced drivers in different driving scenes is obtained through SP investigation, and the process of establishing the scene-specific driver behavior evaluation model based on the investigation result of the safety vehicle speed comprises the following steps: The method comprises the steps of carrying out hierarchical sampling according to accident risk prediction values in dynamic driving scene labels, removing abnormal responses contradicting the dynamic driving scene labels, encoding the dynamic driving scene labels into multidimensional vectors, fusing the multidimensional vectors with driver safety vehicle speed preference data, converting discrete dynamic driving scene labels into four-dimensional binary vectors, namely, security scenes [1, 0], general scenes [0,1, 0], risk scenes [0,1, 0], dangerous scenes [0,1], adding the accident risk prediction values into the four-dimensional binary vectors as additional features to generate five-dimensional feature vectors, digitizing the five-dimensional feature vectors after the SP investigation to form expansion feature vectors, carrying out standardized or embedded encoding on individual attributes of a driver, splicing the additional features with the expansion vectors to generate a complete feature matrix, and forming a fused feature data set; The method comprises the steps of adopting a random forest regression model to fit the relation between dynamic driving scene characteristics and safety speed of a driver in a corresponding scene, initializing the random forest regression model, setting basic parameters including the number of decision trees, the maximum depth and the minimum sample splitting number, traversing the candidate values of the number of the decision trees through grid search on a training set, taking the mean square error of a verification set as an evaluation index, selecting the number of trees with the lowest mean square error as the optimal number of trees, using the mean square error as a tree node splitting standard, defining a random forest regression model parameter space, namely the maximum feature number candidate value, the minimum sample leaf number candidate value and the maximum depth candidate value, using Bayesian optimization or random search to perform joint optimization on the parameter space, setting the optimal target as the minimum mean square error of the verification set, setting the iteration number as A times, using the optimal parameter combination to train the random forest model, starting parallel calculation acceleration training, sorting the fused feature data sets based on the mean value of the reduction of the coefficient, and screening the first B key features; S2, in a non-free flow state, realizing accurate division and characteristic quantification of the traffic state through density clustering and dispersion statistical analysis of multi-characteristic traffic flow samples, wherein the traffic state is divided into smooth, smoother, crowded and blocked through a K-MEANS clustering algorithm; S3, optimizing the constructed scene-specific driver behavior evaluation model according to the divided traffic state labels, wherein the process for generating the driver behavior optimization suggestion under multiple constraints comprises the steps of constructing a multidimensional feature matrix containing static and dynamic features, dynamically distributing feature weights based on the traffic states, namely, improving the weights of the real-time dispersion and the following distance related features in the congestion and blocking states, enhancing the weights of the historical sliding window mean value and the speed fluctuation rate features in the clear and smoother states, and activating corresponding sub-models aiming at different states so as to generate the multidimensional driver behavior optimization suggestion; Optimizing the constructed scene-specific driver behavior evaluation model according to the divided traffic state labels, wherein the process for generating the driver behavior optimization suggestion under multiple constraints comprises the following steps: The method comprises the steps of extracting statistical features and real-time vehicle speed dispersion of four traffic states from a database, integrating real-time data streams to construct a multi-dimensional feature matrix comprising static features and dynamic features, wherein the static features comprise density, flow, occupancy and clustering labels, and the dynamic features comprise the real-time vehicle speed dispersion and a sliding time window mean value; dynamically allocating feature weights based on traffic state classification results; The four traffic states are input as independent submodels, and corresponding prediction frames are respectively constructed, wherein for smooth submodels and smoother submodels, a historical sliding window mean value and a real-time sliding window mean value are taken as core inputs, and the trend of the emphasis speed is predicted; according to the real-time traffic state label, automatically activating a corresponding sub-model, wherein the vehicle speed is dynamically tracked by adopting Kalman filtering aiming at a smooth sub-model and a smoother sub-model, and a historical sliding window mean value and a real-time sliding window mean value are fused to generate a smooth safe vehicle speed; Defining double constraints of the safe vehicle speed, including safety constraint embedding and efficiency constraint layering adaptation; and integrating the outputs of all sub-models to generate multidimensional driver behavior optimization suggestions including safe vehicle speed values, applicable traffic states and constraint condition satisfaction marks, thereby realizing global optimal behavior strategies.
2. The driving scene-based driver behavior assessment method according to claim 1, wherein the process of generating the accident risk prediction value by fusion analysis of four types of heterogeneous basic data by using a graph neural network comprises the following steps: constructing road nodes, edges, facility attributes and meteorological weights into super-dimensional association heterograms, and extracting global features through a message transmission mechanism; The method comprises the steps of obtaining a linear sensitivity coefficient, a facility influence matrix and a weather correction factor, embedding the linear sensitivity coefficient, the facility influence matrix and the weather correction factor into a graph neural network, and further refining characteristic association in a super-dimensional association heterogram, wherein the specific steps are as follows, taking a normalized value of a road geometric linear parameter as a road node characteristic, taking a normalized value of a dynamic traffic flow state as an edge characteristic, taking traffic facility classification coding and weather classification weight as node attributes, splicing association results of the dynamic traffic flow state and other three types of heterogeneous basic data as characteristic vectors, and taking the characteristic vectors as initial nodes and edge characteristics of the graph neural network, so as to construct a characteristic association matrix; Dividing node types, including road nodes, facility nodes and weather nodes, wherein the road nodes are used for storing geometric parameter normalization values, the facility nodes are used for storing traffic facility classification codes, and the weather nodes are used for storing real-time weather classification weights; dividing dynamic traffic flow edge types, including geometric-traffic edges, facility-traffic edges and weather-traffic edges, wherein the geometric-traffic edges are used for connecting road nodes and dynamic traffic flow edges, represent the influence of linearity on the speed of a vehicle, the facility-traffic edges are used for connecting facility nodes and road nodes, represent the constraint of facilities on the flow, and the weather-traffic edges are used for connecting weather nodes and dynamic traffic flow edges, represent the attenuation of weather on the speed of the vehicle; Establishing node updating rules, including road node updating rules and dynamic traffic flow edge updating rules; Generating accident risk prediction values through a global feature aggregation layer of the graph neural network Wherein Aggregating all node and edge characteristics by using a reading function, namely summing the characteristics of all nodes and edges; and training and optimizing by adopting a loss function, wherein the optimizing target is to predict accident risk.
3. The driving scene-based driver behavior evaluation method according to claim 1, wherein the process of calibrating dynamic driving scene tags, namely a safety scene, a general scene, a risk scene and a danger scene, based on accident risk prediction values acquired by a graph neural network comprises the following steps: Based on accident risk prediction value Making driving scene label mapping rule if If the accident risk is extremely low, the vehicle is stable to run and a safety scene label is generated, if Indicating the risk of slight accident and generating general scene labels if Indicating that the accident risk is higher and generating a risk scene label, if And the accident risk is extremely high, and a dangerous scene label is generated.
4. The driving scene-based driver behavior assessment method according to claim 1, wherein the process of achieving accurate division and feature quantification of traffic states through density clustering and dispersion statistical analysis of multi-feature traffic flow samples comprises the following steps: The method comprises the steps of obtaining traffic flow samples, carrying out standardization processing on the traffic flow samples by using a minimum-maximum standardization method, dividing the traffic flow samples into a plurality of subintervals by taking density as a division standard, dividing the density into R as intervals, guaranteeing that the number of traffic flow sample points randomly selected in each interval is consistent, carrying out parameter statistics on absolute vehicle speed dispersion and relative vehicle speed dispersion under four types of traffic conditions respectively, and storing a clustering result and a statistical feature into a database after obtaining a clustering center matrix through a K-MEANS clustering algorithm, carrying out standardization processing on all traffic flow data, and then calculating the Euclidean distance between each traffic flow sample point and a clustering center vector, wherein the clustering center corresponding to the obtained minimum value is the traffic condition of the traffic flow sample point.
5. A driving scenario-based driver behavior assessment system, characterized in that it is applied to a driving scenario-based driver behavior assessment method according to any one of claims 1 to 4, the system comprising: The free flow data risk analysis module is used for fusing multi-source heterogeneous basic data in a free flow state, generating an accident risk prediction value, calibrating a dynamic driving scene tag and constructing a scene-specific driver behavior evaluation model by combining the driver safety vehicle speed preference obtained by SP investigation; The non-free flow clustering and quantifying module is used for realizing accurate division and characteristic quantification of the traffic state through density clustering and dispersion statistical analysis of the multi-characteristic traffic flow samples in the non-free flow state; And the polymorphic constraint optimization module optimizes the constructed scene-specific driver behavior evaluation model according to the divided traffic state labels to generate a driver behavior optimization suggestion under multiple constraints.

Description

Driver behavior evaluation method and system based on driving scene Technical Field The invention relates to the technical field of driving safety, in particular to a driving scene-based driver behavior assessment method and system. Background In the current driving safety field, a driver is used as a core decision body, the behavior evaluation and guidance of the driver faces multiple challenges that the potential influence characterization and risk quantification of the current driving scene on the driver behavior are insufficient, the existing research is mainly dependent on a single data source (such as map data or accident record) to indirectly associate the driver behavior, or static scene labels (such as simple distinction of high-speed/urban roads) are used for general inference, complex and changeable dynamic scenes (such as complex risk avoidance, vehicle speed adjustment and other behavior modes required by the driver under a 'rainy day + curved road + front construction' complex risk scene) are difficult to be described, the evaluation and optimization of the real-time response capability of the driver under the traffic flow state change are seriously lagged, the traditional model is not adaptive to the rapid fluctuation of traffic flow under a non-free flow scene (such as sudden congestion and traffic flow disturbance caused by accident chain reaction) because the traffic flow state is assumed to be stable, so that the behavior guidance strategies such as vehicle speed recommendation generated based on the models have obvious lagging or complete failure, and the safety requirements of the driver in the complex dynamic scenes are difficult to be satisfied. Therefore, the invention provides a driving scene-based driver behavior evaluation method and system. Disclosure of Invention The invention aims to solve the problems in the background art, and provides a driving scene-based driver behavior evaluation method and system. In order to achieve the above purpose, the present invention adopts the following technical scheme: a driving scenario-based driver behavior assessment method, comprising: S1, under a free flow state, fusing multi-source heterogeneous basic data, generating an accident risk predicted value, calibrating a dynamic driving scene tag, and constructing a scene-specific driver behavior evaluation model by combining the driver safety vehicle speed preference obtained by SP investigation; S2, under a non-free flow state, through density clustering and dispersion statistical analysis of multi-feature traffic flow samples, accurate division and feature quantification of traffic states are achieved; And S3, optimizing the constructed scene-specific driver behavior evaluation model according to the divided traffic state labels, and generating a driver behavior optimization suggestion under multiple constraints. Further, the process of fusing multisource heterogeneous basic data, generating accident risk prediction values, calibrating dynamic driving scene labels, combining the driver safety vehicle speed preference obtained by SP investigation, and constructing a scene-specific driver behavior evaluation model comprises the following steps: Synchronously acquiring four types of multi-source heterogeneous basic data, namely road geometric linear parameters, traffic facility type configuration, real-time environmental meteorological data and dynamic traffic flow states, from a vehicle-mounted sensor, a road side unit and meteorological monitoring equipment; The method comprises the steps of preprocessing the four acquired multi-source heterogeneous basic data, carrying out normalization processing on road geometric line parameters, carrying out classified coding on traffic facility type configuration according to functions, namely, forbidden mark=1, warning mark=2 and indication mark=3, grading and weighting real-time environmental meteorological data according to risk levels, and carrying out normalization processing on dynamic traffic flow states; the dynamic traffic flow state is related to other three types of heterogeneous basic data by calculating linear sensitivity coefficients and road geometric linear parameters, by constructing a facility influence matrix and traffic facility type configuration, by introducing weather correction factors and real-time environment weather data; adopting a graph neural network to fusion analyze four types of multi-source heterogeneous basic data and generating an accident risk prediction value; Calibrating dynamic driving scene labels, namely a safety scene, a general scene, a risk scene and a dangerous scene, based on accident risk prediction values obtained by the graph neural network; based on the generated dynamic driving scene label, the SP surveys to obtain the safe vehicle speed preference data of the experienced drivers in different driving scenes, and based on the investigation result of the safe vehicle speed, builds a scene-specific driver