Long-term and short-term driving behavior recognition methods using deep learning and generative probabilistic models
By combining deep learning with generative probabilistic models, Gaussian mixture models, and KL divergence, this study addresses the problem of insufficient modeling of the correlation between short-term driving intentions and long-term driving styles in existing technologies. It achieves joint recognition of lateral and longitudinal driving behaviors and a comprehensive description of driving styles, thereby improving the accuracy and stability of driving behavior recognition.
Patent Information
- Application Number
- CN202611505364.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing driving behavior recognition methods struggle to effectively combine short-term driving intentions with long-term driving styles, and fail to fully consider the correlation between lateral and longitudinal driving behaviors and road environment factors, resulting in insufficient representation and incomplete description of driving styles in complex driving scenarios.
By employing deep learning and generative probabilistic models, and acquiring multidimensional driving data, we construct lateral and longitudinal driving intention recognition models. Combining Gaussian mixture models and KL divergence, we achieve multidimensional probability distribution representation of driving styles and use lateral and longitudinal driving behavior and environmental information for joint recognition and classification.
It improves the representation of driving intentions, enhances the comprehensiveness of driving style description, realizes the joint identification and classification of long-term and short-term driving behaviors, and improves the accuracy and stability of driving style classification.
Smart Images

Figure CN123115601A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of driving recognition, specifically to a method for recognizing long-term and short-term driving behaviors using deep learning and generative probabilistic models. Background Technology
[0002] With the development of intelligent vehicles and driver assistance technologies, driver behavior recognition is increasingly being applied to areas such as driving risk analysis, vehicle assistance control, and personalized driving services. Existing driver behavior recognition methods typically target the driver's behavioral characteristics at different time scales, separately identifying driving intention and classifying driving style. Driving intention mainly reflects the driver's short-term driving behavior in the current traffic scenario, while driving style mainly reflects the relatively stable driving habits formed by the driver over a longer period. These two methods are usually identified or modeled independently, making it difficult to effectively use short-term driving behavior information for long-term driving style representation.
[0003] On the other hand, existing driving intention recognition methods typically identify lateral and longitudinal driving intentions separately, with less consideration given to the correlation between lateral and longitudinal vehicle movements, as well as the influence of driving environment factors such as road slope on longitudinal driving behavior. This results in limitations in representing driving intentions in complex driving scenarios such as turning, lane changing, and sloping roads. Moreover, in terms of driving style recognition, existing methods usually use single statistical features such as mean and variance to describe driver behavior, which is insufficient to fully reflect the randomness, time-varying nature, and individual differences inherent in natural driving data, and it is also difficult to characterize the behavioral differences between different drivers from the perspective of the overall probability distribution of multidimensional driving data.
[0004] In summary, existing driving behavior recognition methods still have room for improvement in modeling the correlation between short-term driving intentions and long-term driving styles, jointly representing lateral and longitudinal driving intentions, and modeling the probability distribution of multidimensional natural driving data. Therefore, to address these issues, a short- and long-term driving behavior recognition method based on deep learning and generative probabilistic models is needed. This method should be able to utilize the correlation between lateral and longitudinal driving behaviors and road environment information to improve the representation of driving intentions and enhance the comprehensiveness of driving style description through multidimensional probability distributions. Summary of the Invention
[0005] In view of this, the purpose of this invention is to overcome the deficiencies in the prior art and provide a method for long-term and short-term driving behavior recognition based on deep learning and generative probabilistic models. This method can improve the representation of driving intentions by utilizing the correlation between lateral and longitudinal driving behaviors and road environment information, and enhance the comprehensiveness of driving style description through multidimensional probability distribution.
[0006] The method for long- and short-term driving behavior recognition based on deep learning and generative probabilistic models of the present invention includes:
[0007] Acquire natural driving data from multiple drivers during driving under different road and traffic conditions, and perform time synchronization and noise reduction on the natural driving data;
[0008] The vehicle's longitudinal and lateral positions in a preset coordinate system are calculated based on natural driving data to obtain the vehicle's driving trajectory; the lateral driving intention label is determined based on the vehicle's driving trajectory; and the longitudinal driving intention label is determined based on the vehicle's speed.
[0009] A lateral driving intention recognition model is constructed based on lateral driving intention labels and natural driving data. The natural driving data to be recognized is then input into the lateral driving intention recognition model to obtain the lateral driving intention recognition result of the current driver.
[0010] The longitudinal driving intention recognition model is used to perform hierarchical recognition of longitudinal driving intentions to obtain the current driver's longitudinal driving intention recognition result;
[0011] Natural driving data within a continuous time frame is combined with lateral driving intent recognition results and longitudinal driving intent recognition results to form driving style feature data;
[0012] Gaussian mixture models are established for the driving style feature data of different drivers, and the Gaussian mixture models are used to represent the joint probability distribution of the corresponding drivers in multiple driving feature dimensions.
[0013] The driving style differences between drivers are determined based on the KL divergence between Gaussian mixture models corresponding to different drivers, and the driving style of different drivers is classified according to the driving style differences.
[0014] Furthermore, the natural driving data includes at least one of the following: vehicle speed, longitudinal acceleration, lateral acceleration, steering wheel angle, yaw rate, accelerator pedal opening, brake fluid pressure, gear, engine speed, accelerator pedal opening change rate, and brake fluid pressure change rate.
[0015] The time synchronization includes aligning the natural driving data collected by different vehicle sensors according to a unified time reference, and resampling the natural driving data with different sampling frequencies using cubic spline interpolation.
[0016] The denoising process includes using wavelet denoising to suppress noise in the time-synchronized natural driving data.
[0017] Furthermore, the lateral driving intentions include going straight, turning left, turning right, changing lanes to the left, and changing lanes to the right; the lateral driving intention labels are obtained by clustering the vehicle's driving trajectory using the K-means clustering algorithm. During the trajectory clustering process, the Fréchet distance is used to calculate the distance between the trajectory to be clustered and the cluster center trajectory. The Fréchet distance is used to simultaneously characterize the spatial position difference and trajectory point order difference between the trajectory to be clustered and the cluster center trajectory, and the clustering results are respectively assigned to going straight, turning left, turning right, changing lanes to the left, and changing lanes to the right.
[0018] The vehicle's trajectory is determined according to the following formula:
[0019] ; ;
[0020] This represents the longitudinal position of the vehicle. The lateral position of the vehicle. For the turning angle of the wheel, The corrected longitudinal velocity, For lateral velocity; as well as The calculation method is as follows:
[0021] ; ; ;
[0022] The yaw rate is angular velocity. The measured vehicle speed value, For longitudinal acceleration, For lateral acceleration, For steering wheel angle, This is the steering gear ratio.
[0023] Furthermore, the longitudinal driving intentions include stopping, accelerating, decelerating, and cruising; wherein, acceleration, deceleration, and cruising are each divided into three driving intention levels;
[0024] The longitudinal driving intent label is determined using a combination of coarse and fine classification.
[0025] In the coarse classification, a vehicle is defined as being in a stopped state when its speed is 0; a vehicle is defined as being in a stopped state when the accelerator pedal opening is greater than 0 and the longitudinal acceleration is greater than... The acceleration state is defined as follows: when the accelerator pedal opening is greater than 0 and the absolute value of the longitudinal acceleration is not greater than... The state is determined to be cruise mode; data other than the aforementioned parking mode, acceleration mode, and cruise mode are determined to be deceleration mode.
[0026] In the detailed classification, K-means clustering is performed on acceleration, deceleration and cruise states respectively. Acceleration state is divided into slow acceleration, normal acceleration and rapid acceleration, deceleration state is divided into slow deceleration, normal deceleration and rapid deceleration, and cruise state is divided into low-speed cruise, medium-speed cruise and high-speed cruise.
[0027] Furthermore, when further classifying the acceleration state, longitudinal acceleration, accelerator pedal opening, and accelerator pedal opening change rate are used as clustering features.
[0028] When classifying the deceleration state, vehicle speed, longitudinal acceleration, and brake fluid pressure are used as clustering features;
[0029] When classifying the cruise states in detail, vehicle speed is used as the clustering feature;
[0030] It also performs slope compensation based on at least one of the vehicle speed, longitudinal acceleration, and brake oil pressure according to the road gradient.
[0031] Furthermore, the lateral driving intention recognition model and each longitudinal driving intention recognition model are all constructed using a CNN-BiLSTM network;
[0032] The CNN-BiLSTM network comprises, in sequence, a feature input layer, a convolutional feature extraction layer, a bidirectional long short-term memory network temporal feature extraction layer, and a classification output layer;
[0033] When training the lateral driving intention recognition model, lateral displacement, longitudinal displacement, and steering wheel angle are used as input features.
[0034] When training the acceleration driving intention recognition model, vehicle speed, longitudinal acceleration, accelerator pedal opening, gear position, and lateral driving intention recognition results are used as input features.
[0035] When training the deceleration driving intention recognition model, vehicle speed, longitudinal acceleration, brake fluid pressure, gear position, and lateral driving intention recognition results are used as input features.
[0036] When training the cruise driving intention recognition model, vehicle speed, gear position, engine speed, and lateral driving intention recognition results are used as input features.
[0037] Furthermore, a longitudinal driving intention recognition model is used to perform hierarchical recognition of longitudinal driving intentions, specifically including:
[0038] Select the corresponding longitudinal driving intent recognition model based on the vehicle's current state of parking, acceleration, deceleration, or cruising;
[0039] When the vehicle is parked, the parking intention is output directly. When the vehicle is accelerating, the acceleration driving intention recognition model is called to identify slow acceleration, normal acceleration or rapid acceleration. When the vehicle is decelerating, the deceleration driving intention recognition model is called to identify slow deceleration, normal deceleration or rapid deceleration. When the vehicle is cruising, the cruise driving intention recognition model is called to identify low-speed cruise, medium-speed cruise or high-speed cruise.
[0040] Each longitudinal driving intention recognition model uses the lateral driving intention recognition result as input feature, so that the current lateral driving behavior state participates in the temporal feature extraction and classification of longitudinal driving intention.
[0041] Furthermore, the driving style feature data consists of natural driving data within a continuous time range and the corresponding driving intention recognition results. The driving style feature data includes vehicle speed, longitudinal acceleration, accelerator pedal opening, accelerator pedal opening change rate, brake fluid pressure, brake fluid pressure change rate, and driving intention.
[0042] A seven-dimensional Gaussian mixture model is established for each driver's driving style feature data. The model parameters of the seven-dimensional Gaussian mixture model are iteratively estimated using the expectation-maximization (EM) algorithm to obtain the driving style probability model of the corresponding driver. The driving style probability model is used to characterize the joint probability distribution of the corresponding driver on the seven-dimensional driving style features.
[0043] Furthermore, the seven-dimensional Gaussian mixture models corresponding to different drivers are used as the probability models to be classified, and the number of driving style categories is predetermined, and representative Gaussian mixture models corresponding to the number of driving style categories are initialized.
[0044] Calculate the KL divergence between each probability model to be classified and each representative Gaussian mixture model, and determine the driving style category corresponding to the representative Gaussian mixture model with the smallest KL divergence as the driving style category of the driver corresponding to the probability model to be classified.
[0045] After classifying the driving style categories of all drivers, the driving style feature data of drivers belonging to the same driving style category are merged, and a representative Gaussian mixture model corresponding to that driving style category is rebuilt.
[0046] Repeat the process of classifying driver driving style categories and updating the representative Gaussian mixture model until the driving style category for each driver remains unchanged.
[0047] Furthermore, an improved Metropolis-Hastings method is used to sample the probability distribution of the Gaussian mixture model to obtain probability samples for KL divergence calculation, specifically including:
[0048] Initialize the state transition probability distribution, initial samples, and preset number of samplings, and generate candidate samples from the current samples based on the state transition probability distribution;
[0049] The acceptance probability judgment value is generated from the uniform distribution, and the acceptance of the candidate sample is determined based on the probability value of the Gaussian mixture model corresponding to the candidate sample and the current sample, as well as the state transition probability.
[0050] After completing a preset number of pre-sampling operations, a preset number of probability samples are continuously collected.
[0051] The beneficial effects of this invention are as follows: This invention discloses a method for recognizing short- and long-term driving behaviors using deep learning and generative probabilistic models. By collecting and preprocessing natural driving data, it labels lateral driving intentions based on vehicle trajectories and utilizes the lateral driving intention recognition results and vehicle motion state information to perform hierarchical recognition of longitudinal driving intentions. Furthermore, it combines the driving intention recognition results with multidimensional natural driving data, employs a Gaussian mixture model to establish a joint probability distribution for drivers, and uses KL divergence to measure differences in driving styles among drivers, achieving joint recognition of short- and long-term driving behaviors and driving style classification. This invention can improve the representation of driving intentions by utilizing the correlation between lateral and longitudinal driving behaviors and road environment information, and enhance the comprehensiveness of driving style description through multidimensional probability distribution. Attached Figure Description
[0052] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0053] Figure 1 This is a schematic diagram of the long-term and short-term driving behavior recognition method of the present invention;
[0054] Figure 2 This is a schematic diagram of the long-term and short-term driving behavior recognition framework of the present invention;
[0055] Figure 3 This is a partial schematic diagram of the lateral driving intention label of the present invention;
[0056] Figure 4 This is a schematic diagram of a portion of the longitudinal driving intent label of the present invention;
[0057] Figure 5 This is a schematic diagram of the CNN-BiLSTM neural network model of the present invention;
[0058] Figure 6 This is a schematic diagram of the short-term driving intention recognition model of the present invention;
[0059] Figure 7 This is a schematic diagram of the lateral driving intent recognition result of the present invention;
[0060] Figure 8This is a schematic diagram of the longitudinal driving intent recognition result of the present invention;
[0061] Figure 9 This is a schematic diagram of the driver classification model based on GMM and KL divergence of the present invention;
[0062] Figure 10 This is a schematic diagram of the joint probability distribution of velocity and acceleration according to the present invention;
[0063] Figure 11 This is a schematic diagram of the joint probability distribution of speed and throttle opening according to the present invention;
[0064] Figure 12 This is a schematic diagram of the combined probability distribution of speed and brake oil pressure according to the present invention;
[0065] Figure 13 This is a schematic diagram of the joint probability distribution of speed and intent tags in this invention. Detailed Implementation
[0066] The present invention will be further described below with reference to the accompanying drawings, as shown in the figures:
[0067] This embodiment discloses a method for identifying short- and long-term driving behaviors using deep learning and generative probabilistic models. It defines short-term driving behaviors in the current traffic scenario as driving intentions, and relatively stable driving habits exhibited by the driver over a longer driving time range as driving style. Specifically, it includes the following steps:
[0068] By acquiring natural driving data from multiple drivers under different road types, traffic conditions, and driving conditions using onboard sensors, and performing time synchronization and noise reduction on the data collected from different sensors, a natural driving data sequence with a consistent time reference is formed.
[0069] Based on the preprocessed natural driving data, the longitudinal and lateral positions of the vehicle in a preset coordinate system are obtained, resulting in the vehicle's driving trajectory. A lateral driving intention label is then determined based on the vehicle's driving trajectory. Simultaneously, a longitudinal driving intention label is determined based on vehicle speed, longitudinal acceleration, and driver operation data. The driver operation data includes at least one of the following: accelerator pedal opening, brake fluid pressure, steering wheel angle, and gear position.
[0070] After obtaining the driving intention labels, the lateral driving intention recognition model and the longitudinal driving intention recognition model are constructed using the labeled natural driving data. The natural driving data to be recognized is then input into the corresponding models to obtain the lateral driving intention recognition results and the longitudinal driving intention recognition results of the current driver.
[0071] By combining vehicle status, driver operation information, and corresponding driving intention recognition results within a certain continuous time range, driving style feature data is formed, enabling short-term driving behavior to participate in driving style modeling over a longer time scale.
[0072] Gaussian mixture models are established for different drivers using their driving style feature data. The joint probability distribution of multiple driving feature dimensions represents the driving behavior characteristics of the corresponding drivers. The KL divergence between the probability models of different drivers is calculated. The degree of difference between the probability distributions represents the driving style differences between drivers. Based on the differences, drivers are classified into driving styles. This forms a process from natural driving data collection and short-term driving intention recognition to long-term driving style modeling and classification.
[0073] In this embodiment, natural driving data is acquired by onboard sensors, a vehicle controller, or a vehicle information acquisition system installed on the vehicle. The natural driving data includes at least one of the following: vehicle speed, longitudinal acceleration, lateral acceleration, steering wheel angle, yaw rate, accelerator pedal opening, brake fluid pressure, gear position, engine speed, accelerator pedal opening rate of change, and brake fluid pressure rate of change. Specifically, vehicle speed, longitudinal acceleration, and lateral acceleration describe the vehicle's motion state; steering wheel angle and yaw rate describe the vehicle's lateral movement and steering state; accelerator pedal opening and brake fluid pressure describe the driver's longitudinal operation; gear position and engine speed assist in characterizing the vehicle's power state; and accelerator pedal opening rate of change and brake fluid pressure rate of change characterize the driver's operational changes.
[0074] Since different sensors have different sampling frequencies and sampling times, a time correspondence between various types of natural driving data is first established according to a unified time reference. For data with inconsistent sampling times, a unified time series is used as the resampling time point, and cubic spline interpolation is used to calculate the data values at the corresponding time, thereby obtaining multi-source driving data with a consistent time axis. After time synchronization is completed, wavelet denoising is performed on the natural driving data. Specifically, the appropriate wavelet basis function and decomposition level are selected according to the frequency characteristics of the natural driving data, and the original signal is decomposed into low-frequency and high-frequency components at different scales. Thresholding is performed on the high-frequency components containing random measurement noise, and then wavelet reconstruction is used to obtain the denoised natural driving data. Through the above processing, data from different sensors can correspond at the same time position, and the impact of sensor noise and random interference on subsequent driving intention recognition and driving style modeling is reduced.
[0075] In this embodiment, lateral driving intentions include going straight, turning left, turning right, changing lanes to the left, and changing lanes to the right. The lateral driving intention label is obtained using the following method:
[0076] First, based on information such as vehicle speed, longitudinal acceleration, lateral acceleration, yaw rate, and steering wheel angle from natural driving data, the vehicle's position change in a preset coordinate system is calculated, and the vehicle's trajectory is extracted according to a continuous time window. The preset coordinate system can use the vehicle's initial position, road reference lines, or pre-defined geographical coordinates as reference points to uniformly represent the spatial motion process of vehicles driven by different drivers. The vehicle trajectory is determined according to the following formula:
[0077] ; ;
[0078] This represents the longitudinal position of the vehicle. The lateral position of the vehicle. For the turning angle of the wheel, The corrected longitudinal velocity, For lateral velocity; as well as The calculation method is as follows:
[0079] ; ; ;
[0080] The yaw rate is angular velocity. The measured vehicle speed value, For longitudinal acceleration, For lateral acceleration, For steering wheel angle, This is the steering gear ratio.
[0081] For each segment of the vehicle's driving trajectory to be labeled, the trajectory is represented as an ordered sequence of trajectory points according to a preset sampling interval, and the distance between the trajectory to be labeled and the candidate cluster center trajectory is calculated using the Fréchet distance. Unlike the method of comparing only the Euclidean distance between trajectory points at corresponding times, the Fréchet distance can preserve the chronological order between trajectory points while considering the differences in the spatial location of the trajectory, thus reflecting the overall trajectory shape formed during vehicle turning, lane changing, and straight-line driving.
[0082] Based on the calculated trajectory distances, the K-means clustering algorithm is used to iteratively cluster the trajectories: First, five cluster centers are initialized randomly or according to preset rules; then, the Fréchet distance between each trajectory and the five cluster centers is calculated, and the trajectories are assigned to the cluster centers with the smallest distances; the cluster centers are then redefined based on the trajectories in each cluster, and the above process is repeated until the change in cluster centers or the change in trajectory categories meets the preset convergence condition. After clustering, based on the trajectory shape corresponding to each cluster center, the five clusters are labeled as straight ahead, left turn, right turn, left lane change, and right lane change, respectively, thus forming the label data for training the lateral driving intention recognition model. Some of the obtained lateral driving intention labels are shown below. Figure 3 As shown.
[0083] In this embodiment, longitudinal driving intent is used to describe the driver's control behavior on the longitudinal speed of the vehicle, including four basic states: stopping, accelerating, decelerating, and cruising. Accelerating, decelerating, and cruising are further divided into three driving intent levels.
[0084] The vertical driving intent labels are obtained through a two-level processing method of coarse classification and fine classification, which reduces the impact of overlapping feature distributions between different driving states when directly clustering all subcategories.
[0085] In the coarse classification stage, whether the vehicle speed is 0 is used as the criterion for determining the stationary state. For data where the vehicle is in motion, the acceleration, cruising, or deceleration state is determined by combining the accelerator pedal opening and longitudinal acceleration. When the accelerator pedal opening is greater than 0 and the longitudinal acceleration is greater than 0.2 m / s², the corresponding data is classified as acceleration; when the accelerator pedal opening is greater than 0 and the absolute value of the longitudinal acceleration is not greater than 0.2 m / s², the corresponding data is classified as cruising; data that does not belong to the above states are classified as deceleration. The aforementioned 0.2 m / s² is a preset acceleration threshold used in this embodiment to distinguish the longitudinal motion state of the vehicle, which can be set according to the vehicle type, data collection characteristics, and experimental conditions.
[0086] In the fine-classification stage, K-means clustering was performed on acceleration, deceleration, and cruising data respectively. For acceleration data, it was divided into gradual acceleration, normal acceleration, and rapid acceleration based on driving characteristics; for deceleration data, it was divided into gradual deceleration, normal deceleration, and rapid deceleration; and for cruising data, it was divided into low-speed cruising, medium-speed cruising, and high-speed cruising. This resulted in 10 longitudinal driving intention labels, including parking and the aforementioned 9 sub-states. By first determining the basic longitudinal motion states and then performing fine-classification within each state, different longitudinal behavioral states can be classified according to their behavioral characteristics, providing label data for subsequent hierarchical longitudinal driving intention recognition. Some of the obtained longitudinal driving intention labels are shown below. Figure 4 As shown.
[0087] In this embodiment, during the longitudinal driving intention sub-classification process, clustering features are selected based on the correspondence between driver operation behavior and vehicle motion state under different driving states. For acceleration state, the acceleration state feature vector is composed of longitudinal acceleration, accelerator pedal opening, and accelerator pedal opening change rate. Here, longitudinal acceleration reflects the actual acceleration degree of the vehicle, accelerator pedal opening reflects the driver's current acceleration operation degree, and accelerator pedal opening change rate reflects the change process of the driver's acceleration operation. K-means clustering is performed on the acceleration state using these three features to distinguish between slow acceleration, normal acceleration, and rapid acceleration.
[0088] For deceleration, the deceleration feature vector is composed of vehicle speed, longitudinal acceleration, and brake fluid pressure. Vehicle speed characterizes the vehicle's current motion, longitudinal acceleration reflects changes in vehicle speed, and brake fluid pressure reflects the degree of driver braking. For cruising, since it primarily involves maintaining a relatively stable speed, vehicle speed is used as the clustering feature for cruising, and based on this, cruising is divided into low-speed cruising, medium-speed cruising, and high-speed cruising.
[0089] Furthermore, road slope alters the vehicle's longitudinal force state, affecting its actual acceleration and braking requirements. Therefore, before generating longitudinal driving intention features, slope compensation is applied to at least one of the vehicle's speed, longitudinal acceleration, and brake fluid pressure based on the road slope. Specifically, the longitudinal motion component caused by the slope can be determined based on the slope information of the road the vehicle is currently traveling on, and corrected from the corresponding vehicle motion data. This reduces the impact of changes in vehicle state caused by uphill or downhill road conditions on the driver's judgment of longitudinal driving intentions; details will not be elaborated further here. The slope-compensated feature data is then used for subsequent K-means clustering and longitudinal driving intention label generation.
[0090] In this embodiment, the lateral driving intention recognition model and the three longitudinal driving intention recognition models for acceleration, deceleration, and cruising are all constructed using a CNN-BiLSTM network. Figure 5 As shown, the CNN-BiLSTM network includes a feature input layer, a convolutional feature extraction layer, a BiLSTM temporal feature extraction layer, and a classification output layer. The feature input layer receives multi-dimensional driving data arranged according to a preset time window; the convolutional feature extraction layer extracts local variation features between adjacent time sampling points and different input features; the BiLSTM temporal feature extraction layer extracts temporal features from the input sequence from both forward and reverse directions to obtain temporal correlation features of driving behavior within the time window; the classification output layer outputs the corresponding driving intention category based on the extracted features.
[0091] like Figure 6As shown, in the lateral driving intention recognition model, lateral displacement, longitudinal displacement, and steering wheel angle are used as input features. Lateral and longitudinal displacements describe changes in vehicle trajectory, while the steering wheel angle reflects the driver's steering input. The obtained lateral driving intention recognition results are as follows: Figure 7 As shown. For the acceleration driving intention recognition model, vehicle speed, longitudinal acceleration, accelerator pedal opening, gear position, and lateral driving intention recognition results are input into the model; for the deceleration driving intention recognition model, vehicle speed, longitudinal acceleration, brake fluid pressure, gear position, and lateral driving intention recognition results are input into the model; for the cruise driving intention recognition model, vehicle speed, gear, engine speed, and lateral driving intention recognition results are input into the model. Thus, the three longitudinal recognition models employ feature combinations corresponding to their respective longitudinal states, while simultaneously utilizing lateral driving intention recognition results to supplement the current driving scenario information.
[0092] During model training, the aforementioned natural driving data with driving intention labels is divided into multiple training and validation samples according to a preset time window, and the network parameters are updated using the backpropagation algorithm. The network parameters can be iteratively updated using the Adam optimization algorithm, and hyperparameters such as the number of convolutional kernels, the number of BiLSTM hidden units, the number of training epochs, and the batch size can be determined using Bayesian optimization methods. The time window length can be set to 2–6 seconds, preferably 3 seconds, and the samples are divided into training and validation data according to a preset ratio.
[0093] In this embodiment, after obtaining the lateral driving intention recognition result, the longitudinal driving intention is recognized hierarchically using the longitudinal driving intention recognition model, rather than directly using a single model to recognize all longitudinal subcategories.
[0094] First, the vehicle's current state is determined to be one of the following basic states: parking, acceleration, deceleration, or cruising. For the parking state, since the vehicle speed is 0, the parking intention is directly output without further sub-classification. For the acceleration state, the vehicle speed, longitudinal acceleration, accelerator pedal opening, gear position, and the obtained lateral driving intention recognition results within the current time window are input into the acceleration driving intention recognition model, which then outputs one of three options: gentle acceleration, normal acceleration, or rapid acceleration. For the deceleration state, the vehicle speed, longitudinal acceleration, brake fluid pressure, gear position, and the lateral driving intention recognition results are input into the deceleration driving intention recognition model, which outputs one of three options: gentle deceleration, normal deceleration, or rapid deceleration. For the cruising state, the vehicle speed, gear, engine speed, and the lateral driving intention recognition results are input into the cruise driving intention recognition model, which outputs one of three options: low-speed cruise, medium-speed cruise, or high-speed cruise. The obtained longitudinal driving intention recognition results are as follows: Figure 8 As shown.
[0095] In this model, the lateral driving intention recognition result is not used as auxiliary information after the longitudinal driving intention recognition is completed, but rather directly participates in temporal feature extraction during the input stage of the longitudinal driving intention recognition model. Therefore, the longitudinal driving intention recognition model can simultaneously utilize vehicle longitudinal motion information, driver longitudinal operation information, vehicle dynamic state, and the current lateral driving state for identification. For example, when a vehicle is turning or changing lanes, its longitudinal speed change may differ from that of straight-line driving. By introducing the lateral driving intention recognition result, the longitudinal driving intention recognition model can distinguish longitudinal behavior based on the current lateral motion state.
[0096] In this embodiment, after completing the driving intention recognition, the result is incorporated into driving style modeling. The driving style feature data is not built for a single moment, but rather accumulated over a continuous time span based on natural driving data. Specifically, over a preset, relatively long time span, vehicle speed, longitudinal acceleration, accelerator pedal opening, accelerator pedal opening rate of change, brake fluid pressure, brake fluid pressure rate of change, and driving intention are recorded in time series format, thereby forming seven-dimensional driving style feature data for each driver.
[0097] Among them, driving intention, as a representation of short-term driving behavior, reflects the driver's recurring operating patterns in different traffic scenarios as its changes over a continuous time range; the remaining vehicle state and driver operating characteristics are used to describe the long-term behavioral distribution formed by the driver during actual driving. Using these seven types of features together as input for driving style modeling can avoid describing driver behavior solely through single statistical measures such as mean and variance.
[0098] A seven-dimensional Gaussian Model (GMM) is established for each driver. The GMM consists of multiple Gaussian components combined according to their corresponding mixture probabilities, describing the joint probability distribution of seven-dimensional driving style feature data. The GMM model parameters include the mean vector, covariance matrix, and mixing coefficients of each Gaussian component. The EM algorithm is used to iteratively estimate these model parameters: in the expectation step, the posterior probability of each driving data sample belonging to different Gaussian components is calculated; in the maximization step, the mean vector, covariance matrix, and mixing coefficients of each Gaussian component are updated based on the posterior probability; and the expectation and maximization steps are repeated until the change in model parameters or the log-likelihood function satisfies a preset convergence condition. Finally, a seven-dimensional driving style probability model corresponding to each driver is obtained.
[0099] In this embodiment, as Figure 9-13As shown, after obtaining the seven-dimensional GMM corresponding to each driver, the GMM-KL method is used to classify the driver's driving style. First, the number of representative driving style models is determined according to the pre-set number of driving style categories, and a corresponding number of representative GMMs are selected or initialized from the GMMs corresponding to each driver. Then, for each driver to be classified, the KL divergence between its GMM and each representative GMM is calculated, and the category corresponding to the representative GMM with the smallest KL divergence is determined as the driver's driving style category.
[0100] Let p and q be the multidimensional Gaussian probability distributions of the two drivers, then the KL divergence between them is defined as follows: .
[0101] Since the KL divergence between two GMMs typically involves integrating a probability density function, direct calculation can be computationally expensive or lack an analytical expression. Therefore, probabilistic sampling can be used to obtain samples for numerical computation. After completing the first round of driver classification, the driving style feature data corresponding to all drivers in the same driving style category are merged, and a representative GMM for that category is rebuilt based on the merged data. The rebuilt representative GMM is used for the next round of driver classification.
[0102] By repeatedly performing the process of calculating the difference between the driver probability model and the representative probability model, reclassifying drivers, and re-estimating the category probability model, the representative GMM is gradually updated based on the actual data of drivers within its category. When the category of each driver remains unchanged in two consecutive classification processes, or when the number of category changes is less than a preset threshold, the driving style classification process is considered to have reached a convergence state, and the driving style category corresponding to each driver is output.
[0103] In this embodiment, an improved Metropolis-Hastings method is used to sample the GMM probability distribution to obtain probability samples for numerical calculation of the KL divergence. The sampling of the GMM probability distribution using the improved Metropolis-Hastings method specifically includes:
[0104] 1) Given a random state transition matrix Q T The stationary distribution is the Gaussian mixture model p. (x) Collect samples n m There are 1 sample, the sampling threshold is n1, and the initial sample is x0;
[0105] 2) Repeat the following sampling steps n1+n m Next: a) Obtain sample x based on conditional probability distribution Q(x|xt). *b) Obtain the probability value u by sampling from the uniform distribution uniform(0, 1). c) If u < (x t , x * ) = min { p (j) Q (j,i) / p (i) Q (i,j) If x ,1} then accept the transfer x t → x * , that is, x t+1 = x * Otherwise, the transfer will not be accepted, i.e., x t+1 =x t .
[0106] 3) Let the sample set collected after the n1th transition be (x n1 , x n1+1 , …, x n1+nm-1 ), which is the required sample to be collected.
[0107] The collected probability samples are input into the corresponding Gaussian mixture model to obtain probability values. Based on the probability values of the probability samples corresponding to different drivers, the KL divergence is calculated to determine the degree of difference between the probability models of different drivers' driving styles. Specifically, this includes:
[0108] For each driver's corresponding GMM, the obtained sampling points are input into the corresponding GMM probability density function to obtain the probability value of each sampling point. Based on the sampling results of the two driver probability models, the difference between the probability distributions is numerically calculated to obtain the corresponding KL divergence. Furthermore, the calculated KL divergence is used for driving style category classification and representative GMM update process, thereby completing the measurement of the difference between the driving style probability models of different drivers.
[0109] This invention uses driving intention recognition results as input features for driving style modeling, achieving correlation modeling and joint recognition of driving behavior at different time scales. Compared to existing methods that separately identify driving intention and driving style, this invention can more fully reveal the intrinsic connection between short-term operational behavior and long-term driving habits, improving the accuracy of driving behavior recognition results.
[0110] This invention employs a layered driving intention recognition structure. First, lateral driving intentions are identified. Then, the lateral driving intention recognition results, road slope, and vehicle trajectory information are introduced as additional features into the longitudinal driving intention recognition process, fully considering the coupling relationship between the driving environment and lateral and longitudinal driving behaviors. This method can identify the driver's lateral and longitudinal driving intentions and reduce misjudgments of longitudinal driving intentions. The final longitudinal driving intention recognition accuracy reaches 99.3%. Furthermore, this invention mainly utilizes existing vehicle signals such as speed, acceleration, steering wheel angle, and vehicle trajectory, eliminating the need for high-investment additional equipment such as eye trackers and EEG devices, thus demonstrating good engineering deployment feasibility.
[0111] This invention combines continuous natural driving data and driving intention recognition results for driving style modeling, and utilizes Geometric Models (GMMs) to represent each driver's long-term driving behavior from a probability distribution perspective. Compared to methods that only use single statistical features such as mean, variance, and event frequency, this invention preserves the overall distribution, multimodal characteristics, and joint relationships between different variables of natural driving data, improving the representation ability of complex random driving behavior. Furthermore, this invention uses KL divergence to calculate the probability distribution differences between GMM models of different drivers, measuring the similarity between different drivers from the overall data distribution perspective, thereby improving the accuracy and stability of driving style classification. The classification accuracy of long-term driving behavior, i.e., driving style, reaches 98.6%.
[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for recognizing long- and short-term driving behaviors using deep learning and generative probabilistic models, characterized in that: include: Acquire natural driving data from multiple drivers during driving under different road and traffic conditions, and perform time synchronization and noise reduction on the natural driving data; The vehicle's longitudinal and lateral positions in a preset coordinate system are calculated based on natural driving data to obtain the vehicle's driving trajectory. Determine the lateral driving intent label based on the vehicle's driving trajectory; Determine longitudinal driving intent labels based on vehicle speed; A lateral driving intention recognition model is constructed based on lateral driving intention labels and natural driving data. The natural driving data to be recognized is then input into the lateral driving intention recognition model to obtain the lateral driving intention recognition result of the current driver. The longitudinal driving intention recognition model is used to perform hierarchical recognition of longitudinal driving intentions to obtain the current driver's longitudinal driving intention recognition result; Natural driving data within a continuous time frame is combined with lateral driving intent recognition results and longitudinal driving intent recognition results to form driving style feature data; Gaussian mixture models are established for the driving style feature data of different drivers, and the Gaussian mixture models are used to represent the joint probability distribution of the corresponding drivers in multiple driving feature dimensions. The driving style differences between drivers are determined based on the KL divergence between Gaussian mixture models corresponding to different drivers, and the driving style of different drivers is classified according to the driving style differences.
2. The method for long-term and short-term driving behavior recognition based on deep learning and generative probabilistic models according to claim 1, characterized in that: The natural driving data includes at least one of the following: vehicle speed, longitudinal acceleration, lateral acceleration, steering wheel angle, yaw rate, accelerator pedal opening, brake fluid pressure, gear, engine speed, accelerator pedal opening change rate, and brake fluid pressure change rate. The time synchronization includes aligning the natural driving data collected by different vehicle sensors according to a unified time reference, and resampling the natural driving data with different sampling frequencies using cubic spline interpolation. The denoising process includes using wavelet denoising to suppress noise in the time-synchronized natural driving data.
3. The method for long-term and short-term driving behavior recognition based on deep learning and generative probabilistic models according to claim 1, characterized in that: The lateral driving intentions include going straight, turning left, turning right, changing lanes to the left, and changing lanes to the right. The lateral driving intention labels are obtained by clustering vehicle trajectories using the K-means clustering algorithm. During the trajectory clustering process, the Fréchet distance is used to calculate the distance between the trajectory to be clustered and the cluster center trajectory. The Fréchet distance is used to simultaneously characterize the spatial position difference and trajectory point order difference between the trajectory to be clustered and the cluster center trajectory, and the clustering results are respectively assigned to going straight, turning left, turning right, changing lanes to the left, and changing lanes to the right. The vehicle's trajectory is determined according to the following formula: ; ; This represents the longitudinal position of the vehicle. The lateral position of the vehicle. For the turning angle of the wheel, The corrected longitudinal velocity, For lateral velocity; as well as The calculation method is as follows: ; ; ; The yaw rate is angular velocity. The measured vehicle speed value, For longitudinal acceleration, For lateral acceleration, For steering wheel angle, This is the steering gear ratio.
4. The method for long-term and short-term driving behavior recognition based on deep learning and generative probabilistic models according to claim 1, characterized in that: The longitudinal driving intentions include stopping, accelerating, decelerating, and cruising; among them, acceleration, deceleration, and cruising are each divided into three driving intention levels; The longitudinal driving intent label is determined using a combination of coarse and fine classification. In the coarse classification, a vehicle is defined as being in a stopped state when its speed is 0; a vehicle is defined as being in a stopped state when the accelerator pedal opening is greater than 0 and the longitudinal acceleration is greater than... The acceleration state is defined as follows: when the accelerator pedal opening is greater than 0 and the absolute value of the longitudinal acceleration is not greater than... The state is determined to be cruise mode; data other than the aforementioned parking mode, acceleration mode, and cruise mode are determined to be deceleration mode. In the detailed classification, K-means clustering is performed on acceleration, deceleration and cruise states respectively. Acceleration state is divided into slow acceleration, normal acceleration and rapid acceleration, deceleration state is divided into slow deceleration, normal deceleration and rapid deceleration, and cruise state is divided into low-speed cruise, medium-speed cruise and high-speed cruise.
5. The method for long-term and short-term driving behavior recognition based on deep learning and generative probabilistic models according to claim 4, characterized in that: When further classifying the acceleration states, longitudinal acceleration, accelerator pedal opening, and accelerator pedal opening change rate are used as clustering features; When classifying the deceleration state, vehicle speed, longitudinal acceleration, and brake fluid pressure are used as clustering features; When classifying the cruise states in detail, vehicle speed is used as the clustering feature; It also performs slope compensation based on at least one of the vehicle speed, longitudinal acceleration, and brake oil pressure according to the road gradient.
6. The method for long-term and short-term driving behavior recognition based on deep learning and generative probabilistic models according to claim 1, characterized in that: The lateral driving intention recognition model and each longitudinal driving intention recognition model are constructed using a CNN-BiLSTM network. The CNN-BiLSTM network comprises, in sequence, a feature input layer, a convolutional feature extraction layer, a bidirectional long short-term memory network temporal feature extraction layer, and a classification output layer; When training the lateral driving intention recognition model, lateral displacement, longitudinal displacement, and steering wheel angle are used as input features. When training the acceleration driving intention recognition model, vehicle speed, longitudinal acceleration, accelerator pedal opening, gear position, and lateral driving intention recognition results are used as input features. When training the deceleration driving intention recognition model, vehicle speed, longitudinal acceleration, brake fluid pressure, gear position, and lateral driving intention recognition results are used as input features. When training the cruise driving intention recognition model, vehicle speed, gear position, engine speed, and lateral driving intention recognition results are used as input features.
7. The method for long-term and short-term driving behavior recognition based on deep learning and generative probabilistic models according to claim 6, characterized in that: The longitudinal driving intention recognition model is used to perform hierarchical recognition of longitudinal driving intentions, specifically including: Select the corresponding longitudinal driving intent recognition model based on the vehicle's current state of parking, acceleration, deceleration, or cruising; When the vehicle is parked, the parking intention is output directly. When the vehicle is accelerating, the acceleration driving intention recognition model is called to identify slow acceleration, normal acceleration or rapid acceleration. When the vehicle is decelerating, the deceleration driving intention recognition model is called to identify slow deceleration, normal deceleration or rapid deceleration. When the vehicle is cruising, the cruise driving intention recognition model is called to identify low-speed cruise, medium-speed cruise or high-speed cruise. Each longitudinal driving intention recognition model uses the lateral driving intention recognition result as input feature, so that the current lateral driving behavior state participates in the temporal feature extraction and classification of longitudinal driving intention.
8. The method for long-term and short-term driving behavior recognition based on deep learning and generative probabilistic models according to claim 1, characterized in that: The driving style feature data consists of natural driving data within a continuous time range and the corresponding driving intention recognition results. The driving style feature data includes vehicle speed, longitudinal acceleration, accelerator pedal opening, accelerator pedal opening change rate, brake fluid pressure, brake fluid pressure change rate, and driving intention. A seven-dimensional Gaussian mixture model is established for each driver's driving style feature data. The model parameters of the seven-dimensional Gaussian mixture model are iteratively estimated using the expectation-maximization (EM) algorithm to obtain the driving style probability model of the corresponding driver. The driving style probability model is used to characterize the joint probability distribution of the corresponding driver on the seven-dimensional driving style features.
9. The method for long-term and short-term driving behavior recognition based on deep learning and generative probabilistic models according to claim 8, characterized in that: The seven-dimensional Gaussian mixture model corresponding to different drivers is used as the probability model to be classified, and the number of driving style categories is determined in advance. The representative Gaussian mixture model corresponding to the number of driving style categories is initialized. Calculate the KL divergence between each probability model to be classified and each representative Gaussian mixture model, and determine the driving style category corresponding to the representative Gaussian mixture model with the smallest KL divergence as the driving style category of the driver corresponding to the probability model to be classified. After classifying the driving style categories of all drivers, the driving style feature data of drivers belonging to the same driving style category are merged, and a representative Gaussian mixture model corresponding to that driving style category is rebuilt. Repeat the process of classifying driver driving style categories and updating the representative Gaussian mixture model until the driving style category for each driver remains unchanged.
10. The method for long-term and short-term driving behavior recognition based on deep learning and generative probabilistic models according to claim 9, characterized in that: An improved Metropolis-Hastings method is used to sample the probability distribution of the Gaussian mixture model to obtain probability samples for KL divergence calculation, specifically including: Initialize the state transition probability distribution, initial samples, and preset number of samplings, and generate candidate samples from the current samples based on the state transition probability distribution; The acceptance probability judgment value is generated from the uniform distribution, and the acceptance of the candidate sample is determined based on the probability value of the Gaussian mixture model corresponding to the candidate sample and the current sample, as well as the state transition probability. After completing a preset number of pre-sampling operations, a preset number of probability samples are continuously collected.
