Trajectory fractal feature and emotional evolution fused driving behavior prediction method

Abstract

The invention discloses a driving behavior prediction method fusing track fractal features and emotional evolution, and relates to the technical field of driving behavior analysis, and the method comprises the steps: S1, obtaining vehicle motion parameters, traffic situation indexes, a driving track coordinate sequence, a driver physiological signal sequence, and multi-modal driving data of a synchronously labeled emotional state label sequence, carrying out the preprocessing of the multi-modal driving data, and generating test sample data; s2, based on the driving track coordinate sequence, fractal dimension features of the driving track are extracted through a fractal geometric analysis method; and S3, based on the vehicle motion parameters, the traffic situation indexes and the emotional state label state, constructing and training a hidden Markov model, and taking the emotional state as a hidden state of the hidden Markov model. According to the invention, the fractal dimension of the driving track is used as a quantitative index, key discrimination information except traditional motion parameters is provided for the model, and the prediction precision is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of driving behavior analysis technology, specifically to a method for predicting driving behavior by fractal features of trajectories and emotional evolution. Background Technology

[0002] Driving behavior prediction is a key technology in the fields of intelligent transportation and autonomous driving. Its core lies in accurately inferring the driver's future operational intentions through multi-source data such as vehicle trajectory and environmental information. Currently, related research mainly follows a data-driven paradigm, utilizing machine learning and deep learning models to fit and predict historical trajectory sequences.

[0003] The invention patent application with application number CN117786526A discloses a vehicle trajectory prediction method that considers the characteristics of the driver's emotional state. This application aims to solve the problem that existing vehicle trajectory prediction technologies are mostly based on the physical motion characteristics of vehicles and do not fully consider the nonlinear impact of the driver's emotional fluctuations on the decision-making process, resulting in limited prediction accuracy in complex traffic scenarios; it also fails to deeply reveal how emotional state structurally affects the geometric shape of the driving trajectory in space, and lacks quantitative modeling of the dynamic evolution process of emotions and its correlation with behavioral patterns.

[0004] In addition, although some studies have introduced fractal theory into the field of transportation, such as the patent disclosure with application number CN101290713A, which aims to explore its periodicity or trend, it has no direct relationship with the geometric complexity of the microscopic single vehicle driving trajectory in two-dimensional space, and does not involve the psychological state factors of the driver; another type of technology, such as application number CN110379193B, focuses entirely on autonomous driving behavior planning strategies in specific scenarios, which belong to the traditional control methods driven by rules or optimization, and neither analyzes the spatial structural characteristics of the trajectory nor considers the driver's emotions.

[0005] Existing technical solutions have the following limitations: In terms of emotion modeling, most are single emotion state recognition or simple feature fusion, lacking in-depth modeling of the dynamic transfer of multiple types of emotions and their coupling mechanism with behavioral decision-making; in terms of trajectory feature characterization, they generally rely on traditional kinematic parameters or coordinate point sequences, failing to quantify the spatial structural complexity and cross-scale self-similarity characteristics of the trajectory caused by emotional perturbation from a geometric perspective, making it difficult for the model to capture the deep mechanism of behavior patterns driven by emotions; therefore, we propose a driving behavior prediction method that integrates trajectory fractal features and emotion evolution. Summary of the Invention

[0006] To address the aforementioned technical issues, this technical solution provides a driving behavior prediction method that integrates trajectory fractal features and emotion evolution. This solution resolves the problem of lacking in-depth modeling of the dynamic transfer of multiple types of emotions and their coupling mechanism with behavioral decision-making.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] Methods for predicting driving behavior that integrate trajectory fractal features and emotional evolution include:

[0009] S1. Acquire multimodal driving data including vehicle motion parameters, traffic situation indicators, driving trajectory coordinate sequence, driver physiological signal sequence, and synchronously labeled emotional state label sequence, and preprocess the multimodal driving data to generate test sample data;

[0010] S2. Based on the driving trajectory coordinate sequence, extract the fractal dimension features of the driving trajectory using fractal geometry analysis methods;

[0011] S3. Based on the vehicle motion parameters, traffic situation indicators and emotional state label states, construct and train a hidden Markov model, take the emotional state as the hidden state of the hidden Markov model, take the driving behavior state as the observable state, and output the emotional state vector at the current moment.

[0012] S4. Construct and train a GRU model that integrates fractal features and emotion evolution. The input layer receives standardized vehicle motion parameters, traffic situation indicators, driver physiological signal sequences, fractal dimension features and emotion state vectors. After processing by multiple GRU units, the predicted driving behavior category is used as the output.

[0013] S5. Input the preprocessed test sample data into the trained GRU model that integrates fractal features and emotion evolution, and output the prediction results of six types of driving behaviors: acceleration following, constant speed following, deceleration following, acceleration lane change, constant speed lane change, and deceleration lane change.

[0014] Preferably, step S1 specifically includes:

[0015] The multimodal driving data was collected through a high-fidelity driving simulation test platform. The driving simulation test platform is based on UC-win / Road software to construct simulation scenarios including urban main roads, suburban expressways and commercial street roads. It is equipped with a force feedback steering wheel, real vehicle pedals and a multi-channel data synchronous acquisition system to synchronously collect the vehicle motion parameters, driving trajectory coordinate sequence and driver physiological signal sequence at a sampling frequency of not less than 50Hz.

[0016] Preferably, step S1 further includes a data preprocessing step:

[0017] The driving trajectory coordinate sequence is uniformly transformed into a vehicle body coordinate system centered on the vehicle through coordinate transformation;

[0018] The continuous time series data is segmented according to a preset fixed time window to generate sample segments for model input.

[0019] The vehicle motion parameters are normalized, and the driver's physiological signals are standardized.

[0020] The emotional state label sequence is encoded into a corresponding category value.

[0021] Preferably, the fractal geometry analysis method is the box counting method, which specifically includes the following steps:

[0022] A series of square grids with different side lengths ε are set to cover the two-dimensional trajectory curve formed by mapping the driving trajectory coordinate sequence;

[0023] For each given side length ε, calculate the minimum number of square boxes N(ε) required to completely cover the two-dimensional trajectory curve;

[0024] The fractal dimension D of the driving trajectory is calculated based on the number of boxes N(ε) and the corresponding side length ε.

[0025] Preferably, in step S3, the specific steps for constructing the Hidden Markov Model are as follows:

[0026] Define the set of hidden states These correspond to six basic emotional states: anger, happiness, sadness, fear, surprise, and disgust; define the set of observed states. These correspond to six types of driving behaviors: accelerating and following, maintaining a constant speed, decelerating and following, accelerating and changing lanes, maintaining a constant speed and changing lanes, and decelerating and changing lanes.

[0027] Based on the vehicle motion parameters and traffic situation indicators, and according to preset driving behavior discrimination rules, each driving operation at any given moment is classified as a specific driving behavior state, thereby generating a driving behavior state observation sequence synchronized with the emotional state label sequence. ,in I is the defined set of driving behavior states;

[0028] Determine the parameter set of the Hidden Markov Model The parameter set λ consists of the initial state probability distribution π, the state transition probability matrix A, and the observation probability matrix B.

[0029] Preferably, in step S3, the specific steps for training the Hidden Markov Model are as follows:

[0030] Using the driving behavior state observation sequence O as input, the Baum-Welch algorithm is applied to the parameter set of the Hidden Markov Model. Perform iterative optimization until the model converges to the likelihood function of the observed sequence;

[0031] For a given observation sequence O, using the trained model parameters The probability of the driver being in various emotional states at each moment is calculated using a forward-backward algorithm.

[0032] The probability distribution of all predefined emotional states calculated at each moment is used as the emotional state vector representing the driver's emotional state at that moment.

[0033] Preferably, in step S4, constructing a GRU model that integrates fractal features and emotion evolution specifically includes:

[0034] The preprocessed vehicle motion parameters, traffic situation indicators, driver physiological signal sequence, fractal dimension features extracted in step S2, and emotional state vector output in step S3 are concatenated and aligned to form a unified temporal feature vector.

[0035] A neural network with a gated recurrent unit as the core temporal processing layer is constructed. The input dimension of the network matches the dimension of the temporal feature vector, and the output dimension matches the number of driving behavior categories to be predicted.

[0036] The final output layer of the network is configured as a Softmax layer, which is used to map the final hidden state of the network to the predicted probabilities of various driving behaviors.

[0037] Preferably, in step S4, training the GRU model that integrates fractal features and emotion evolution specifically includes:

[0038] The temporal feature vector is input into the network, and then sequentially passes through the gated recurrent unit layer for temporal feature extraction and state update. Finally, it is processed by the Softmax output layer to obtain the predicted probability distribution of the driving behavior category at each time step.

[0039] The error loss between the predicted probability distribution and the corresponding real driving behavior category label is calculated using the cross-entropy loss function.

[0040] The gradients of each parameter in the network are calculated based on the error loss using the backpropagation algorithm; the Adam optimizer is then used to iteratively update the network parameters according to the gradients.

[0041] Preferably, in step S4, training the GRU model that integrates fractal features and emotion evolution further includes:

[0042] The observation probability matrix and behavior transition probability matrix of the Hidden Markov Model trained in step S3 are introduced as logical prior knowledge.

[0043] In the loss calculation step, a regularization term is added to constrain the output probability distribution of the prediction model to be consistent with the logical prior knowledge, thereby achieving collaborative modeling of the emotion evolution law.

[0044] Preferably, step S5 specifically includes:

[0045] The preprocessed test sample data is used to generate input samples containing the time-series feature vector in the same way as in the training phase.

[0046] The input samples are fed into the trained GRU model that integrates fractal features and emotion evolution. Through forward computation of the model, the predicted probabilities of each driving behavior category at each time step are obtained.

[0047] Based on the predicted probability, the final driving behavior prediction result is output, which is one of six types of behavior: accelerating to follow, constant speed following, decelerating to follow, accelerating to change lanes, constant speed to change lanes, and decelerating to change lanes.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] This invention proposes the fractal dimension of driving trajectory as a quantitative indicator, characterizing the nonlinear perturbation features of driving behavior from the perspective of spatial geometric complexity. This provides the model with key discriminative information beyond traditional motion parameters, significantly improving prediction accuracy. This invention constructs an interpretable modeling framework of "emotional state evolution → driving trajectory spatial features → driving behavior decision-making." Hidden Markov models are used to explicitly model the transition of emotional states and their probabilistic correlation with driving behavior output, enabling the model to not only output prediction results but also simultaneously provide inferences about the driver's potential emotional state. Attached Figure Description

[0050] Figure 1 This is a flowchart of the method of the present invention;

[0051] Figure 2 This is a schematic diagram illustrating the calculation of fractal dimension using the box counting method of the present invention;

[0052] Figure 3 This invention provides the fractal dimension distribution of driving trajectories under different emotional states.

[0053] Figure 4 This invention relates to the multi-scale variation of the fractal dimension of driving trajectories under different emotional states. Detailed Implementation

[0054] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0055] Reference Figure 1 As shown, the driving behavior prediction method that integrates trajectory fractal features and emotion evolution includes:

[0056] S1. Acquire multimodal driving data including vehicle motion parameters, traffic situation indicators, driving trajectory coordinate sequence, driver physiological signal sequence, and synchronously labeled emotional state label sequence, and preprocess the multimodal driving data to generate test sample data;

[0057] S2. Based on the driving trajectory coordinate sequence, extract the fractal dimension features of the driving trajectory using fractal geometry analysis methods;

[0058] S3. Based on the vehicle motion parameters, traffic situation indicators and emotional state label states, construct and train a hidden Markov model, take the emotional state as the hidden state of the hidden Markov model, take the driving behavior state as the observable state, and output the emotional state vector at the current moment.

[0059] S4. Construct and train a GRU model that integrates fractal features and emotion evolution. The input layer receives standardized vehicle motion parameters, traffic situation indicators, driver physiological signal sequences, fractal dimension features and emotion state vectors. After processing by multiple GRU units, the predicted driving behavior category is used as the output.

[0060] S5. Input the preprocessed test sample data into the trained GRU model that integrates fractal features and emotion evolution, and output the prediction results of six types of driving behaviors: acceleration following, constant speed following, deceleration following, acceleration lane change, constant speed lane change, and deceleration lane change.

[0061] Step S1 specifically includes:

[0062] The multimodal driving data is collected through a high-fidelity driving simulation test platform. The driving simulation test platform is based on UC-win / Road software to construct simulation scenarios including urban main roads, suburban expressways and commercial street roads. It is equipped with a force feedback steering wheel, real vehicle pedals and a multi-channel data synchronous acquisition system to synchronously collect the vehicle motion parameters, driving trajectory coordinate sequence and driver physiological signal sequence at a sampling frequency of not less than 50Hz.

[0063] The main urban roads have high traffic volume and frequent disruptions; the suburban expressways have high vehicle speeds and risks such as construction and road occupation; and the commercial districts have mixed pedestrian and vehicle traffic and dense traffic lights.

[0064] Using a professional multi-channel physiological recorder (such as the Biopac MP160 system), the driver's electrocardiogram (ECG), electromyography (EMG), skin conductance response (GSR), and respiratory rate (RSP) are collected synchronously at a sampling frequency of 50Hz.

[0065] Emotional labels are obtained through a paradigm of "emotional arousal - driving simulation - video review". Before the driving experiment, audiovisual materials highly related to traffic scenarios, such as accident videos, traffic jam scenes, and music of a specific style, are used to induce target emotions. After the driving experiment, the subjects are invited to watch the video recording of their driving process and conduct structured interviews based on the "driver's cognitive model under the influence of emotion" to review their emotional state and cognitive process at the moment of operation. This allows for simultaneous manual labeling of six types of emotions, namely anger, happiness, sadness, fear, surprise, and disgust, on the entire driving data.

[0066] Step S1 further includes a data preprocessing step:

[0067] Remove obvious outliers (such as those exceeding 3 times the standard deviation), and use linear interpolation to complete less than 5% of missing data; otherwise, delete the missing data. Ensure that vehicle data, physiological signals, and emotion labels are strictly synchronized on the timeline.

[0068] The driving trajectory coordinate sequence is transformed into a vehicle body coordinate system centered on the vehicle through coordinate transformation; the vehicle body coordinate system is a vehicle body coordinate system with the vehicle's forward direction as the x-axis and the vertical direction as the y-axis, eliminating the influence of scene coordinate system differences on trajectory analysis;

[0069] The continuous time series data is segmented according to a preset fixed time window of 5 seconds to generate sample segments for model input; all samples are randomly divided into training set and test set in a 7:3 ratio.

[0070] The vehicle motion parameters are normalized using Min-Max, and the driver's physiological signals are normalized using Z-score.

[0071] The emotional state label sequence is encoded into a corresponding category value.

[0072] The fractal geometry analysis method is the box counting method, and the specific calculation process is as follows: Figure 2 The following are included:

[0073] A series of square grids with different side lengths ε are defined to cover the two-dimensional trajectory curve formed by mapping the driving trajectory coordinate sequence; where the side length (ε) is a square grid with successively decreasing side lengths, for example, ε starts from 1 / 2 of the side length of the outer envelope rectangle of the trajectory and is halved each time until a minimum resolution is reached;

[0074] For each given side length ε, calculate the minimum number of square boxes N(ε) required to completely cover the two-dimensional trajectory curve;

[0075] Based on the number of boxes N(ε) and the corresponding side length ε, the fractal dimension D of the driving trajectory is calculated. According to fractal theory, N(ε) and ε satisfy a power law relationship. Linear fitting is performed on multiple sets of (ε, N(ε)) data in a double logarithmic coordinate system (logN(ε) ~ log(1 / ε)), and the slope of the resulting straight line is the desired fractal dimension D. The calculation formula can be expressed as:

[0076]

[0077] For each emotion category, 40 data points of 5 seconds each were selected, and their driving trajectory indicators were quantitatively analyzed. The calculated trajectory fractal dimension is as follows: Figure 3 As shown, the overall trend indicates that the fractal dimension of driving trajectories influenced by emotions such as anger, disgust, and surprise is relatively high, with a mean ranging from 1.28 to 1.34, reflecting strong nonlinear perturbations in driving behavior driven by these emotions. Conversely, the fractal dimension of trajectories influenced by inhibitory emotions such as fear and sadness is relatively low, with a mean concentrated between 1.11 and 1.22, reflecting a conservative and defensive driving strategy. The fractal dimension under the influence of positive emotions is generally in the middle range, reflecting a moderately flexible driving mode.

[0078] To further verify the stability and discriminative power of fractal features of driving trajectories under emotional states, and to explore self-similarity features in fractal geometry, the fractal dimension of driving trajectories was calculated across multiple spatial scales. Twenty data sets were selected for each emotion, with each trajectory exceeding 500 meters. The trajectories were segmented according to different scales, with a minimum scale of 50 meters and a maximum scale of 500 meters, increasing progressively from 50 meters. The distribution of fractal dimensions of driving trajectories across multiple scales under emotional states is shown below. Figure 4 As shown;

[0079] The results show that, under the same emotional state, the mean of the fractal dimension exhibits extremely high consistency across different scale intervals, with small fluctuations, demonstrating significant scale invariance. Under the emotion of "Anger," the mean of the fractal dimension is 1.3328±0.0184, the maximum is 1.3655, the minimum is 1.2960, and the standard deviation is less than 0.02. This phenomenon also exists under other emotional states, such as "Surprise" and "Happiness," where the fractal dimensions are 1.2723±0.0258 and 1.2241±0.0228, respectively. The standard deviations for all six emotion categories are less than 0.03, indicating that the fractal dimension, as a measure of trajectory complexity, possesses good self-similarity across different scales, demonstrating the iterative characteristics of trajectories across scales. This also provides a theoretical basis for extracting spatial distribution patterns and regularities from small-scale trajectories and applying them to large-scale behavioral feature prediction.

[0080] Comparing the fractal dimensions of trajectories across different emotional states reveals significant differences in their means, demonstrating strong discriminative power. For example, comparing "Sadness" and "Anger," the former's mean fractal dimension is only 1.1003, significantly lower than the latter's 1.3328. Similarly, the mean for "Fear" is 1.1341, also significantly lower than "Disgust's" 1.2915. This trend indicates a clear difference in the spatial characteristics of driving trajectories under the influence of extroverted emotions (anger, surprise, disgust) and repressed emotions (sadness, fear). This discriminative property based on fractal dimension not only demonstrates the profound impact of emotional evolution on the trajectory perturbation structure but also provides a crucial quantitative indicator for constructing a driving behavior recognition model during emotional evolution. Especially when the differences in trajectory dynamics are small, the structural complexity reflected by the fractal dimension can serve as a supplementary feature to enhance the model's ability to discriminate emotion-driven behaviors.

[0081] In step S3, the specific steps for constructing the hidden Markov model are as follows:

[0082] Define the set of hidden states These correspond to six basic emotional states: anger, happiness, sadness, fear, surprise, and disgust; define the set of observed states. These correspond to six types of driving behaviors: accelerating and following, maintaining a constant speed, decelerating and following, accelerating and changing lanes, maintaining a constant speed and changing lanes, and decelerating and changing lanes.

[0083] Based on the vehicle motion parameters and traffic situation indicators, and according to preset driving behavior discrimination rules, each driving operation at any given moment is classified as a specific driving behavior state, thereby generating a driving behavior state observation sequence synchronized with the emotional state label sequence. ,in I is the defined set of driving behavior states;

[0084] Determine the parameter set of the Hidden Markov Model The parameter set λ consists of the initial state probability distribution π, the state transition probability matrix A, and the observation probability matrix B.

[0085]

[0086] In the formula, This is the initial state probability distribution, used to represent the various emotional states at the initial time. The probability of; Let be the state transition probability matrix, where Used to represent the emotional state from time t-1 Emotional state at time t The probability of; Let be the observation probability matrix, where the elements are... Used to indicate the emotional state at time t. Under these conditions, driving behavior was observed. The probability of.

[0087] In step S3, the specific steps for training the Hidden Markov Model are as follows:

[0088] Using the driving behavior state observation sequence O as input, the Baum-Welch algorithm is applied to the parameter set of the Hidden Markov Model. Perform iterative optimization until the model converges to the likelihood function of the observed sequence;

[0089] For a given observation sequence O, using the trained model parameters The probability of the driver being in various emotional states at each moment is calculated using a forward-backward algorithm.

[0090] The probability distribution of all predefined emotional states calculated at each moment is used as the emotional state vector representing the driver's emotional state at that moment.

[0091] Based on the above, and through training and analysis using experimental data, the initial probability of driving emotion is determined. Transition probability matrix between emotional states :

[0092]

[0093] Probability matrix of driving behavior under different emotional states for:

[0094]

[0095] Under anger, the probability of "accelerating lane change" was highest (0.351), indicating that risky behavior is more easily induced in a state of high arousal. Under surprise, the probability of "decelerating lane change" was highest (0.365), reflecting the driving effect of sudden emotions causing changes in operational behavior and spatial needs. Under sadness and fear, the probability of "decelerating and following" was relatively high (0.313 and 0.302), indicating that repressed emotions tend to lead to conservative and cautious driving strategies. Under disgust, the probability of "accelerating and following" was relatively high (0.301), reflecting that drivers exhibit more direct speed control behavior under strong negative emotions. The overall trend shows that the HMM model can effectively capture the coupling relationship between emotion transfer and behavior change, and the probability transfer of driving behavior under different emotions.

[0096] Training data containing fractal features yields a transition probability matrix of driving behavior during the evolution of driver emotions, part of which is shown below:

[0097]

[0098]

[0099]

[0100]

[0101] The aforementioned behavioral transition probability matrices reveal the dynamic changes in driving behavior under different emotional evolution paths. When emotions are relatively stable, such as transitioning from one positive emotion to another, the behavioral patterns show strong consistency, with a relatively balanced distribution of transition probabilities between behaviors, and an overall fluctuation range of only 0.102–0.236, reflecting the stability of driving behavior. However, when emotions are agitated, such as the self-reinforcing process from anger to anger, the differences in behavioral transitions significantly increase, with the probability range rising to 0.075–0.397. Among these, the self-cyclical transition probability of lane-changing acceleration to lane-changing acceleration is the highest, at 0.397, indicating that under strong negative emotions, drivers are more likely to maintain aggressive driving techniques. By modeling the behavioral transition probabilities under emotional evolution, we can deepen our understanding of the influence mechanism of emotions on driving behavior and provide data support for the construction of subsequent behavioral prediction models.

[0102] In step S4, constructing the GRU model that integrates fractal features and emotion evolution (FHEA-GRU) specifically includes:

[0103] The preprocessed vehicle motion parameters, traffic situation indicators, driver physiological signal sequence, fractal dimension features extracted in step S2, and emotional state vector output in step S3 are concatenated and aligned to form a unified temporal feature vector.

[0104] A neural network with a gated recurrent unit as the core temporal processing layer is constructed. The input dimension of the network matches the dimension of the temporal feature vector, and the output dimension matches the number of driving behavior categories to be predicted.

[0105] The final output layer of the network is configured as a Softmax layer, which is used to map the final hidden state of the network to the predicted probabilities of various driving behaviors.

[0106] In step S4, training the GRU model that integrates fractal features and emotion evolution specifically includes:

[0107] The temporal feature vector is input into the network, and then sequentially passes through the gated recurrent unit layer for temporal feature extraction and state update. Finally, it is processed by the Softmax output layer to obtain the predicted probability distribution of the driving behavior category at each time step.

[0108] The error loss between the predicted probability distribution and the corresponding real driving behavior category label is calculated using the cross-entropy loss function.

[0109] The gradients of each parameter in the network are calculated based on the error loss using the backpropagation algorithm; the Adam optimizer is then used to iteratively update the network parameters according to the gradients.

[0110] In step S4, training the GRU model that integrates fractal features and emotion evolution specifically includes:

[0111] The observation probability matrix and behavior transition probability matrix of the Hidden Markov Model trained in step S3 are introduced as logical prior knowledge.

[0112] In the loss calculation step, a regularization term is added to constrain the output probability distribution of the prediction model to be consistent with the logical prior knowledge, thereby achieving collaborative modeling of the emotion evolution law. Specifically, the observation probability matrix B obtained from HMM training and the behavior transition probability matrix P under emotion transition are used as prior knowledge. This regularization term encourages the behavior prediction probability distribution output by the GRU model to be as close as possible to the behavior probability distribution (from B and P) given by the HMM in the corresponding hidden state (which can be approximated by the emotion state vector). This essentially uses the "emotion-behavior" statistical law mined by HMM as a logical constraint to guide the learning of the deep network, realizing the synergy between the mechanism model and the data-driven model.

[0113] Step S5 specifically includes:

[0114] The preprocessed test sample data is used to generate input samples containing the time-series feature vector in the same way as in the training phase.

[0115] The input samples are fed into the trained GRU model that integrates fractal features and emotion evolution. Through forward computation of the model, the predicted probabilities of each driving behavior category at each time step are obtained.

[0116] Based on the predicted probability, the final driving behavior prediction result is output, which is one of six types of behavior: accelerating to follow, constant speed following, decelerating to follow, accelerating to change lanes, constant speed to change lanes, and decelerating to change lanes.

[0117] After constructing and training the FHEA-GRU model, which integrates fractal features and emotion evolution, it can be deployed for practical driving behavior prediction tasks, and the model's performance and core feature contributions can be systematically verified. The specific application process and verification results of this embodiment are as follows:

[0118] First, the model application follows a standardized inference process. The raw driving data to be predicted (which can come from a simulation platform or real-vehicle road tests) is preprocessed and its features extracted strictly according to the same procedures as the training phase. This includes: data cleaning and synchronization, trajectory coordinate transformation to the vehicle body coordinate system, segmentation into 5-second time windows, Min-Max normalization of vehicle motion parameters, Z-score standardization of physiological signals, and simultaneous encoding of emotional state labels. Then, the pre-trained Hidden Markov Model (HMM) is used to decode the driving behavior sequences in the preprocessed data, generating corresponding emotional state probability vectors. Simultaneously, box counting is used to calculate the fractal dimension of the driving trajectory within each time window. Finally, the standardized vehicle motion parameters, traffic situation indicators, physiological signals, fractal dimension, and emotional state vector are concatenated into a unified temporal feature vector, which is then input into the pre-trained FHEA-GRU model. The model outputs the predicted probability distributions for six types of driving behaviors—accelerating follow-up, constant speed follow-up, decelerating follow-up, accelerating lane change, constant speed lane change, and decelerating lane change—through forward calculation. The one with the highest probability is taken as the final driving behavior prediction result at that moment.

[0119] Secondly, to quantitatively evaluate the effectiveness and innovation of this invention, we conducted systematic testing on three typical scenario test sets covering urban main roads, suburban expressways, and commercial street roads. Experimental results show that the complete model proposed in this invention (FHEA-GRU) exhibits superior comprehensive prediction performance. More importantly, through carefully designed ablation experiments, we quantitatively evaluated the key contribution of trajectory fractal features: when fractal dimension features were removed from the model input, the prediction accuracy of the model decreased significantly in all test scenarios, with the most pronounced performance degradation in the most complex environment and frequently disturbed trajectory commercial street scenario. This comparative result directly confirms that the fractal dimension features extracted from driving trajectories can effectively capture the nonlinear, multi-scale perturbation patterns exhibited in the spatial geometry of driving operations driven by emotions. This is an indispensable core technical element for improving the accuracy of behavior recognition in complex scenarios and enhancing the model's generalization ability.

[0120] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A driving behavior prediction method that integrates trajectory fractal features and emotional evolution, characterized in that, include: S1. Acquire multimodal driving data including vehicle motion parameters, traffic situation indicators, driving trajectory coordinate sequence, driver physiological signal sequence, and synchronously labeled emotional state label sequence, and preprocess the multimodal driving data to generate test sample data; S2. Based on the driving trajectory coordinate sequence, extract the fractal dimension features of the driving trajectory using fractal geometric analysis methods; S3. Based on the vehicle motion parameters, traffic situation indicators and emotional state label states, construct and train a hidden Markov model, take the emotional state as the hidden state of the hidden Markov model, take the driving behavior state as the observable state, and output the emotional state vector at the current moment. S4. Construct and train a GRU model that integrates fractal features and emotion evolution. The input layer receives standardized vehicle motion parameters, traffic situation indicators, driver physiological signal sequences, fractal dimension features and emotion state vectors. After processing by multiple GRU units, the predicted driving behavior category is used as the output. S5. Input the preprocessed test sample data into the trained GRU model that integrates fractal features and emotion evolution, and output the prediction results of six types of driving behaviors: acceleration following, constant speed following, deceleration following, acceleration lane change, constant speed lane change, and deceleration lane change.

2. The driving behavior prediction method based on the fusion of trajectory fractal features and emotion evolution as described in claim 1, characterized in that, Step S1 specifically includes: The multimodal driving data was collected through a high-fidelity driving simulation test platform. The driving simulation test platform was built based on UC-win / Road software to construct simulation scenarios including urban main roads, suburban expressways and commercial street roads. It was equipped with a force feedback steering wheel, real vehicle pedals and a multi-channel data synchronous acquisition system, which synchronously collected the vehicle motion parameters, driving trajectory coordinate sequence and driver physiological signal sequence at a sampling frequency of not less than 50Hz.

3. The driving behavior prediction method based on the fusion of trajectory fractal features and emotional evolution as described in claim 1, characterized in that, Step S1 further includes a data preprocessing step: The driving trajectory coordinate sequence is transformed into a vehicle body coordinate system centered on the vehicle through coordinate transformation; The continuous time series data is segmented according to a preset fixed time window to generate sample segments for model input. The vehicle motion parameters are normalized, and the driver's physiological signals are standardized. The emotional state label sequence is encoded into a corresponding category value.

4. The driving behavior prediction method that integrates trajectory fractal features and emotion evolution according to claim 1, characterized in that, The fractal geometric analysis method is the box counting method, which specifically includes the following steps: A series of square grids with different side lengths ε are set to cover the two-dimensional trajectory curve formed by mapping the driving trajectory coordinate sequence; For each given side length ε, calculate the minimum number of square boxes N(ε) required to completely cover the two-dimensional trajectory curve; The fractal dimension D of the driving trajectory is calculated based on the number of boxes N(ε) and the corresponding side length ε.

5. The driving behavior prediction method that integrates trajectory fractal features and emotion evolution according to claim 1, characterized in that, In step S3, the specific steps for constructing the hidden Markov model are as follows: Define the set of hidden states These correspond to the six basic emotional states: anger, happiness, sadness, fear, surprise, and disgust. Define the set of observed states These correspond to six types of driving behaviors: accelerating and following, maintaining a constant speed, decelerating and following, accelerating and changing lanes, maintaining a constant speed and changing lanes, and decelerating and changing lanes. Based on the vehicle motion parameters and traffic situation indicators, and according to preset driving behavior discrimination rules, each driving operation at any given moment is classified as a specific driving behavior state, thereby generating a driving behavior state observation sequence synchronized with the emotional state label sequence. ,in I is the defined set of driving behavior states; Determine the parameter set of the Hidden Markov Model The parameter set λ consists of the initial state probability distribution π, the state transition probability matrix A, and the observation probability matrix B.

6. The driving behavior prediction method that integrates trajectory fractal features and emotion evolution according to claim 5, characterized in that, In step S3, the specific steps for training the Hidden Markov Model are as follows: Using the driving behavior state observation sequence O as input, the Baum-Welch algorithm is applied to the parameter set of the Hidden Markov Model. Perform iterative optimization until the model converges to the likelihood function of the observed sequence; For a given observation sequence O, using the trained model parameters The probability of the driver being in various emotional states at each moment is calculated using a forward-backward algorithm. The probability distribution of all predefined emotional states calculated at each moment is used as the emotional state vector representing the driver's emotional state at that moment.

7. The driving behavior prediction method that integrates trajectory fractal features and emotion evolution according to claim 1, characterized in that, In step S4, constructing the GRU model that integrates fractal features and emotion evolution specifically includes: The preprocessed vehicle motion parameters, traffic situation indicators, driver physiological signal sequence, fractal dimension features extracted in step S2, and emotional state vector output in step S3 are concatenated and aligned to form a unified temporal feature vector. A neural network with a gated recurrent unit as the core temporal processing layer is constructed. The input dimension of the network matches the dimension of the temporal feature vector, and the output dimension matches the number of driving behavior categories to be predicted. The final output layer of the network is configured as a Softmax layer, which is used to map the final hidden state of the network to the predicted probabilities of various driving behaviors.

8. The driving behavior prediction method that integrates trajectory fractal features and emotion evolution according to claim 7, characterized in that, In step S4, training the GRU model that integrates fractal features and emotion evolution specifically includes: The temporal feature vector is input into the network, and then sequentially passes through the gated recurrent unit layer for temporal feature extraction and state update. Finally, it is processed by the Softmax output layer to obtain the predicted probability distribution of the driving behavior category at each time step. The error loss between the predicted probability distribution and the corresponding real driving behavior category label is calculated using the cross-entropy loss function. The gradients of each parameter in the network are calculated based on the error loss using the backpropagation algorithm; the Adam optimizer is then used to iteratively update the network parameters according to the gradients.

9. The driving behavior prediction method that integrates trajectory fractal features and emotion evolution according to claim 8, characterized in that, In step S4, training the GRU model that integrates fractal features and emotion evolution specifically includes: The observation probability matrix and behavior transition probability matrix of the Hidden Markov Model obtained in step S3 are introduced as logical prior knowledge. In the loss calculation step, a regularization term is added to constrain the output probability distribution of the prediction model to be consistent with the logical prior knowledge, thereby achieving collaborative modeling of the emotion evolution law.

10. The driving behavior prediction method that integrates trajectory fractal features and emotion evolution according to claim 9, characterized in that, Step S5 specifically includes: The preprocessed test sample data is used to generate input samples containing the time-series feature vector in the same way as in the training phase. The input samples are fed into the trained GRU model that integrates fractal features and emotion evolution. Through forward computation of the model, the predicted probabilities of each driving behavior category at each time step are obtained. Based on the predicted probability, the final driving behavior prediction result is output, which is one of six types of behavior: accelerating to follow, constant speed following, decelerating to follow, accelerating to change lanes, constant speed to change lanes, and decelerating to change lanes.

Images