Intelligent vehicle variable impedance driving assistance control method based on generative trajectory prediction

Abstract

This invention discloses a variable impedance driving assistance control method for intelligent vehicles based on generative trajectory prediction, comprising the following steps: (1) synchronously collecting multi-source data to construct a multimodal spatiotemporal context tensor; (2) using a generative trajectory prediction model to predict the driver's ideal operating trajectory; (3) based on the intention deviation between the actual operation and the ideal trajectory, integrating the potential energy of the dynamic driving risk field for comprehensive evaluation, and generating a tactile interaction tuple through a nonlinear strategy mapping mechanism; (4) establishing a variable impedance control model based on the tuple, calculating the auxiliary interaction torque, and driving the steer-by-wire motor to execute. Compared with the traditional torque superposition mode, this method establishes an implicit tactile interaction channel by adjusting the mechanical impedance of the steering wheel, realizing the adaptive smooth intervention of the auxiliary torque and the flexible alignment with the driving intention, effectively alleviating control conflict, and improving the handling stability of intelligent vehicles and the driver's interactive experience.

Description

Technical Field

[0001] This invention relates to the field of automotive intelligent interaction technology and motor vehicle automatic control driving safety technology, specifically to an intelligent vehicle variable impedance driving assistance control method based on generative trajectory prediction. Background Technology

[0002] With the rapid development of intelligent vehicle technology, intelligent driving assistance systems have gradually become an important component in improving driving safety and comfort. In terms of lateral assist control, traditional assist control systems mainly employ rule-based or simple prediction-based torque superposition intervention methods. That is, when the system detects lane departure or collision risk, it adjusts the vehicle's trajectory by applying an additional corrective torque to the steering system. While this method improves driving safety to some extent, its response mode has inherent lag and limitations, often failing to precisely align with the driver's real-time operational intentions, easily causing driver discomfort. Especially in emergency or complex driving situations, when system intervention is inconsistent with the driver's operational intentions, it can easily lead to control conflicts and oscillations between the human and machine, increasing the driver's workload and bringing higher safety risks. Therefore, designing an intelligent interactive system that can proactively understand the driver's intentions, smoothly integrate into the driver's operational loop, and dynamically adjust the assistance strategy according to real-time environmental risks plays a crucial role in improving the human-machine collaborative driving experience and realizing high-level autonomous driving technology.

[0003] Chinese invention patent application number CN201910333992.2 discloses a "vehicle steering assist control method, device, equipment, and storage medium." This method improves the adaptability of the power steering system to different steering scenarios by establishing a mapping relationship between steering operation signals and preset driver intentions to identify driving intentions and then querying a preset assist characteristic curve to adjust the output torque of the steering motor. However, its control channel and driver operation channel are in parallel, and the generated assist torque is prone to conflict with the driver's immediate operating intentions in terms of direction and amplitude. This makes it difficult to achieve the integration and coordination of human-machine operating torque, resulting in abrupt intervention, ambiguous control perception, and an inability to dynamically and continuously adjust the assist intensity according to environmental risks. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a variable impedance driving assistance control method for intelligent vehicles based on generative trajectory prediction. This method can achieve adaptive and smooth intervention of the assistance torque and flexible alignment with the driving intention based on accurate identification of the driver's intention and driving environment risks. It can effectively alleviate the conflict between human and machine control and significantly improve the handling stability of intelligent vehicles and the driver's interactive experience. The variable impedance control method makes the intervention of the assistance system more natural and human-like, and can better meet the driver's core needs for smooth steering feel and controllability.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] A method for intelligent vehicle variable impedance driving assistance control based on generative trajectory prediction includes the following steps;

[0007] (1) Simultaneously collect the driver's physiological characteristics data, vehicle operation behavior data and external environment perception data, perform time-series alignment processing, and use deep neural networks to extract spatial features and temporal features respectively. Construct a multimodal spatiotemporal context tensor representing the spatiotemporal dependence of people-vehicle-road through feature splicing.

[0008] (2) The driver intention recognition model is determined to be a generative trajectory prediction model. The multimodal spatiotemporal scenario tensor is input into the model, and the long-term causal dependency features in the data are mined by the self-attention mechanism to generate the ideal steering wheel angle sequence of the driver in the future preset time domain.

[0009] (3) Calculate the intentional deviation between the actual operation and the ideal trajectory, and make a comprehensive evaluation by combining the potential energy of the dynamic driving risk field. Through the nonlinear strategy mapping mechanism, generate tactile interaction tuples including compliance strategy, damping stabilization strategy and rigid intervention strategy, and dynamically output tactile interaction tuples containing virtual stiffness coefficient and virtual damping coefficient.

[0010] (4) Based on the tactile interaction tuple, a variable impedance control model is established, the auxiliary interaction torque used to adjust the steering feel is calculated, and the auxiliary interaction torque is executed by the steer-by-wire motor to adjust the equivalent rotational inertia and damping of the steering wheel, so as to realize tactile feedback to the driver.

[0011] As a further improvement to this technical solution:

[0012] Step (1) involves constructing a multimodal spatiotemporal context tensor representing the spatiotemporal dependency relationship between people, vehicles, and roads by concatenating the extracted spatial and temporal features. Specifically, this includes:

[0013] (101) Data acquisition: The driver's gaze point coordinates and pupil diameter sequence are collected using in-vehicle vision sensors as physiological characteristic data; the steering wheel angle, yaw rate and longitudinal speed are collected using vehicle data bus as vehicle handling behavior data; and lane line deviation and relative distance to obstacles are obtained using on-board sensors as external environment perception data.

[0014] (102) Data preprocessing and time-series alignment: First, the data collected in step (101) is converted to different units, and an improved t-text test is used to remove outlier values ​​from the heterogeneous multi-source data. Then, timestamp synchronization is performed, and all heterogeneous data are resampled to a uniform system sampling frequency using a linear interpolation algorithm. Z-score standardization is then performed to eliminate dimensional differences, forming a time-aligned multi-source input sequence. The multi-source input sequence specifically includes physiological feature sequences. Vehicle handling behavior sequence and external environment perception sequence ;

[0015] (103) Spatial encoding of physiological features, The images are converted into eye-tracking heatmaps and head posture Euler angle images, and then physiological feature space feature vectors are extracted using a convolutional neural network. ; It consists of feature maps output by convolutional layers, and the coordinates on these feature maps are... pixel value at The calculation formula is:

[0016] (1)

[0017] in, For activation functions; and For eye-tracking heatmaps and head pose images in position Pixel value at; and The convolution kernel corresponding to the input channel is located at position Weight parameters at the location; For bias terms; and These are the height and width of the convolution kernel, respectively; and The row and column indices for sliding the convolution kernel;

[0018] (104) Vehicle behavior timing coding, Input a Long Short-Term Memory (LSTM) network to extract temporal feature vectors of vehicle behavior It utilizes a gating mechanism to capture long-term operational habits, and its internal state update formula is:

[0019] (2)

[0020] (3)

[0021] (4)

[0022] (5)

[0023] (6)

[0024] (7)

[0025] in, Input the vehicle behavior at the current moment; This is the hidden state from the previous moment; These are the forget gate, input gate, and output gate, respectively. In cellular state, Candidate cell state; This is the corresponding weight matrix; This is the corresponding bias vector; It is the sigmoid function; This indicates element-wise multiplication; the final output is... for ;

[0026] (105) For It is mapped to an environmental risk feature vector through a fully connected encoding layer. ;

[0027] (106) Multimodal tensor construction and positional encoding, which converts feature vectors The sequence is concatenated along the feature channel dimension; sinusoidal positional encoding is then superimposed to preserve the temporal information of the sequence, ultimately constructing a multimodal spatiotemporal context tensor. :

[0028] (8)

[0029] in, For the position encoding matrix, This indicates a vector concatenation operation.

[0030] As a further improvement to this technical solution:

[0031] The generative trajectory prediction model in step (2) adopts a transformer Encoder-Decoder architecture, and specifically includes the following steps:

[0032] (201) Context encoding, which encodes the multimodal spatiotemporal context tensor. The input is fed into the encoder; the encoder consists of multi-layered stacked multi-head self-attention modules and a feedforward neural network, used to extract and fuse long-term temporal dependency features from the input tensor, and output a contextual semantic feature matrix. ;

[0033] (202) Autoregressive trajectory decoding, which decodes the context semantic feature matrix The input is fed into the trajectory generation decoder; the trajectory generation decoder includes a masked self-attention module, a cross-attention module, and a feedforward neural network, and operates in an autoregressive manner; at each time step, based on the generated steering wheel angle sequence prefix, it is fused through the cross-attention mechanism. Predict the conditional probability distribution of the steering wheel angle at the next moment;

[0034] (203) Sequence sampling output: Sample the probability distribution of the final output of step (202) to obtain the initial steering wheel angle sequence for the next N time moments. ;

[0035] (204) Trajectory smoothing and motion verification: The smoothing filter is called to smooth the initial steering wheel angle sequence output in step (203). Filtering is performed to obtain a smoothed steering angle sequence. Then, the kinematic feasibility of the smoothed steering angle sequence is determined to verify whether it meets the vehicle's kinematic constraints and actual performance requirements. Finally, an ideal steering wheel angle sequence that meets the conditions is output. .

[0036] As a further improvement to this technical solution:

[0037] The generative trajectory prediction model employs a multi-head self-attention mechanism that fuses relative positional biases. The calculation formula for this mechanism is as follows:

[0038] (9)

[0039] (10)

[0040] in, These are the query matrix, key matrix, and value matrix obtained by linear projection of the input, respectively; This is the scaling factor; Given a learnable relative position bias matrix, its elements are... The value depends only on the position. and relative distance It is used to explicitly model the relative positional dependencies between temporal features; For the number of attention heads, each attention head Independently compute attention; This is the linear projection matrix used for the output after splicing.

[0041] As a further improvement to this technical solution:

[0042] Step (3) Dynamic driving risk field construction and potential energy calculation, the specific steps include:

[0043] (301) Construct a static road constraint potential field, and generate a lateral gravitational potential field with the center line of the front lane as the gravitational target. Using the lane boundary as the repulsive force source, a static repulsive potential field is generated that increases exponentially with the deviation distance. The basic driving trajectory of the vehicle is constrained by these constraints, and the calculation formula is as follows:

[0044] (11)

[0045] (12)

[0046] in, This is the gravitational gain coefficient. The vehicle's lateral position. The lateral coordinates of the lane centerline; The repulsive force intensity at the road boundary. and These are the distances from the vehicle to the left and right boundaries, respectively. This is a parameter representing the range of influence of the repulsive force field.

[0047] (302) Construct a time-varying repulsive field for dynamic obstacles, targeting the sensed first... A dynamic obstacle is used to construct an anisotropic Gaussian repulsive field modulated by relative velocity, which is located globally. Dynamic repulsive potential energy generated at the location for:

[0048] (13)

[0049] in, The inherent danger potential peak value set according to the type of obstacle; This is the current position of the obstacle; These are the variance parameters that characterize the extent of risk diffusion in the longitudinal and lateral directions, respectively. This represents the relative velocity components between the vehicle and the obstacle in the longitudinal and lateral directions. The speed influence factor is used to directly embed the relative speed components of the vehicle and the obstacle in the longitudinal and lateral directions into the denominator of the Gaussian distribution, thereby more accurately assessing complex driving risks.

[0050] (303) Composite risk field synthesis: The static potential field mentioned above is linearly weighted and superimposed with the repulsive potential fields of all dynamic obstacles to obtain the global dynamic risk potential energy characterizing the comprehensive danger level of the current location. :

[0051] (14)

[0052] in, These are the static field weighting coefficients. For the first The repulsive field weight coefficient of a dynamic obstacle.

[0053] As a further improvement to this technical solution:

[0054] The nonlinear strategy mapping mechanism in step (3) includes the following steps:

[0055] First, define the intention bias normalization factor. and environmental risk normalization factor :

[0056] (15)

[0057] (16)

[0058] Among them, the sensitivity coefficient Deviation tolerance and high-risk threshold These are configuration parameters preset based on driver characteristics or vehicle dynamics. This represents the actual steering wheel angle. The next moment's steering angle in the ideal steering wheel angle sequence;

[0059] Secondly, based on the intention bias normalization factor and environmental risk normalization factor Calculate the virtual stiffness coefficient and virtual damping coefficient , forming tactile interaction tuples :

[0060] (17)

[0061] (18)

[0062] in, The maximum virtual stiffness allowed by the system. Basic damping, This is the damping gain coefficient. As a regulating factor;

[0063] The pattern classification of interaction strategies is based on decision factors. and The system exhibits smooth transition characteristics across three main categories of interaction strategies due to dynamic changes; when When the value approaches 0, the system enters a compliant strategy, the steering wheel feels light, and the driver has complete control; when... When the system is approximately in the middle range, it enters a damped stabilization strategy; at this time, if At lower levels, the system is in auxiliary stability control mode, with lower stiffness and moderately increased damping to provide a stable feel and eliminate operational jitter; if At higher speeds, the system enters a softer alert mode, where stiffness takes effect and, in conjunction with increased damping, the driver experiences a gentle resistance, creating a subtle tactile cue to guide corrective action; when... When the value approaches 1, the system enters a rigid intervention strategy; at this point, if The system is in a rigid warning mode with low stiffness and damping, assisting the driver in precise hazard avoidance; if When the stiffness is also high, the system is in a strong guidance mode, with the stiffness approaching its maximum value, outputting an extremely strong self-correcting torque, forcibly taking over lateral control to avoid collision risks; through the adaptive switching of the above modes, the system achieves a continuous and smooth transition from complete compliance to forced intervention.

[0064] As a further improvement to this technical solution:

[0065] The variable impedance control model in step (4) includes the following:

[0066] First, based on haptic interaction tuples Calculate the auxiliary interaction torque that ultimately acts on the steering column. Its control law is:

[0067] (19)

[0068] Among them, stiffness coefficient and virtual damping coefficient , This represents the actual steering wheel angle. The next moment's steering angle in the ideal steering wheel angle sequence. The current steering wheel angular velocity, The angular velocity at the next moment in the ideal steering wheel angle sequence;

[0069] Secondly, the steer-by-wire motor executes auxiliary interactive torque to adjust the equivalent moment of inertia and damping of the steering wheel, thereby providing tactile feedback to the driver; when the system is in a compliant strategy... The system provides only minimal damping to maintain the driver's original control feel to the greatest extent possible; when in damped stability control strategy... Increased torque output becomes the dominant factor in torque output, used to suppress unstable or excessively rapid steering maneuvers; when in a rigid intervention strategy, Approaching The system generates a strong guiding torque to correct high-risk or severely deviated driving maneuvers.

[0070] This invention fully considers the accurate understanding of the driver's operating intentions, the real-time assessment of driving environment risks, and the high-order requirements for the dynamic compliance of human-machine interaction. By combining generative trajectory prediction with variable impedance control, it significantly improves the human-like level of the assistance system intervention and the driver's acceptance while ensuring the lateral stability of the vehicle.

[0071] Compared with the prior art, the beneficial effects of the present invention are:

[0072] This method first constructs a multimodal spatiotemporal context tensor and employs a generative trajectory prediction model to achieve high-precision and forward-looking prediction of the driver's future operating trajectory. Then, it designs a strategy that integrates a dynamic driving risk field and a nonlinear strategy mapping mechanism, quantifying prediction bias and environmental risk into continuously adjustable virtual stiffness and damping coefficients, forming a tactile interaction tuple. Finally, through a variable impedance control model, it transforms auxiliary commands into real-time adjustments to the equivalent mechanical impedance of the steering wheel, thereby establishing an implicit and smooth tactile interaction channel. Compared to the abrupt intervention and control conflict issues of the torque superposition control method described in the background, this method establishes an implicit tactile interaction channel between the driver and the assistance system by adjusting the mechanical impedance characteristics of the steering wheel. This achieves adaptive and smooth intervention of the auxiliary torque and flexible alignment with driving intentions, effectively alleviating control conflict problems and enabling the system to achieve adaptive adjustment within a continuous range, balancing driving safety and driving experience.

[0073] This invention establishes an integrated architecture encompassing perception, prediction, decision-making, and execution. The generative trajectory prediction model provides the system with human-like intent understanding capabilities, a prerequisite for intelligent decision-making; the variable impedance control model executes assistance strategies in an ergonomic manner, key to experience optimization; the two work synergistically to achieve precise matching of the assistance torque with the driver's expectations in terms of timing, direction, and intensity. This collaborative control paradigm provides a safe, compliant, and reliable solution for advanced human-machine co-driving systems. Attached Figure Description

[0074] Figure 1 This is a flowchart of the multimodal spatiotemporal scenario tensor construction process of the present invention;

[0075] Figure 2This is a structural diagram of the generative intent prediction model of the present invention;

[0076] Figure 3 This is a schematic diagram of the dynamic driving risk field construction of the present invention;

[0077] Figure 4 This is a schematic diagram of the nonlinear strategy mapping mechanism of the present invention;

[0078] Figure 5 This is a schematic diagram of the physical principle of the variable impedance control of the present invention;

[0079] Figure 6 This is a flowchart of the overall process framework of the present invention. Detailed Implementation

[0080] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be noted that in the description of the present invention, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0081] A method for intelligent vehicle variable impedance driving assistance control based on generative trajectory prediction, such as Figure 6 It includes the following steps;

[0082] Step 1) Data acquisition: The driver's gaze point coordinates and pupil diameter sequence are collected using in-vehicle vision sensors as physiological characteristic data; the steering wheel angle, yaw rate and longitudinal speed are collected using the vehicle data bus as vehicle handling behavior data; and lane line deviation and relative distance to obstacles are obtained using on-board sensors as external environment perception data.

[0083] Step 2) Constructing the multimodal spatiotemporal scenario tensor

[0084] See Figure 1 As can be seen, since the default units of some of the collected data are not easily understood and observed intuitively, the units were first converted after data collection. Vehicle status data was converted to Standard International Units (SI), and environmental perception data and driver physiological data were converted to the vehicle coordinate system. Secondly, each data point in each data group was segmented, and an improved t-text method was used to remove outliers from each segment. For each data segment, a specific data quantity was selected as the observation sample set, and a statistic was defined. :

[0085] (1)

[0086] in, The observed value to be detected; and Remove from the sample set The arithmetic mean and standard deviation are calculated after these parameters.

[0087] Set judgment criteria: If Then determine Outliers are identified and either removed or replaced by interpolation. The critical value is... for:

[0088] (2)

[0089] Where n is the sample size; It has degrees of freedom. of Distribution Quantiles. The cumulative probability corresponding to the quantile is given by the test criterion. Decision. This embodiment uses a two-sided test, taking:

[0090] (3)

[0091] Where k is the test standard;

[0092] After unit conversion and outlier removal, the multimodal data was resampled to a uniform frequency using a linear interpolation algorithm. Subsequently, Z-score normalization was employed to eliminate dimensional differences between different physical quantities (such as vehicle speed in km / h and turning angle in degrees).

[0093] (4)

[0094] in, for The original value at time; and They are respectively The sample mean and standard deviation are included within the time window prior to the specified time. For a very small positive number, the denominator is guaranteed to be non-zero;

[0095] Physiological feature sequences are formed after standardization. Vehicle handling behavior sequence and external environment perception sequence To address the data characteristics of different modalities, the following parallel neural network encoding channels were designed:

[0096] Physiological feature spatial coding, The images are converted into eye-tracking heatmaps and head posture Euler angle images, and then physiological feature space feature vectors are extracted using a convolutional neural network. ;vector It consists of feature maps output by convolutional layers, and the coordinates on these feature maps are... pixel value at The calculation formula is:

[0097] (5)

[0098] in, The activation function (such as the ReLU function); and For eye-tracking heatmaps and head pose images in position Pixel value at; and The convolution kernel corresponding to the input channel is located at position Weight parameters at the location; For bias terms; and These are the height and width of the convolution kernel, respectively; and The row and column indices for sliding the convolution kernel;

[0099] Vehicle behavior timing coding will Input a Long Short-Term Memory (LSTM) network to extract temporal feature vectors of vehicle behavior The gating mechanism captures long-term operational habits, and its internal state update formula is as follows:

[0100] (6)

[0101] (7)

[0102] (8)

[0103] (9)

[0104] (10)

[0105] (11)

[0106] in, Input the vehicle behavior at the current moment; This is the hidden state from the previous moment; These are the forget gate, input gate, and output gate, respectively. In cellular state, Candidate cell state; This is the corresponding weight matrix; This is the corresponding bias vector; It is the sigmoid function; This indicates element-wise multiplication; the final output is... for ;

[0107] Environmental perception feature mapping, for It is mapped to an environmental risk feature vector through a fully connected encoding layer. ;

[0108] Multimodal tensor construction and positional encoding, transforming feature vectors The sequence is concatenated along the feature channel dimension; sinusoidal positional encoding is then superimposed to preserve the temporal information of the sequence, ultimately constructing a multimodal spatiotemporal context tensor. :

[0109] (12)

[0110] (13)

[0111] (14)

[0112] in, For the position encoding matrix, This represents a vector concatenation operation; Indicates the absolute position index of the time step; Indexed by feature dimension; This represents the total dimension of the feature vectors.

[0113] Step 3) Construct a generative trajectory prediction model based on the transformer Encoder-Decoder architecture.

[0114] See Figure 2 It can be seen that the multimodal spatiotemporal context tensor constructed in step 2) The input is fed into a generative trajectory prediction model, which outputs an ideal steering wheel angle sequence within a preset time domain. The specific steps are as follows:

[0115] (201) Context encoding, which encodes the multimodal spatiotemporal context tensor. The input is transformed into a highly abstract contextual semantic feature matrix by the encoder. The encoder consists of N identical coding layers stacked together, with each layer containing two sub-layers.

[0116] The first sub-layer integrates a multi-head self-attention mechanism with relative positional bias, specifically including:

[0117] First, Through three different linear projection matrices Mapped to a query matrix, a key matrix, and a value matrix:

[0118] (15)

[0119] Then, a scaled dot product attention with relative position bias B is introduced:

[0120] (16)

[0121] in, This is the scaling factor; Given a learnable relative position bias matrix, its elements are... The value depends only on the position. and relative distance ;

[0122] Finally, the outputs of multiple attention heads are concatenated and subjected to a linear transformation. Output:

[0123] (17)

[0124] in, For the number of attention heads, each attention head Independently compute attention; This is the linear projection matrix used for the output after splicing.

[0125] The second sub-layer is a location-based feedforward neural network:

[0126] (18)

[0127] in, For the input of the current sub-layer, These are the learnable parameters for this sublayer;

[0128] (202) Autoregressive trajectory decoding, which decodes the context semantic feature matrix The input is fed to the trajectory generation decoder; the decoder consists of a masked self-attention module, a cross-attention module, and a feedforward neural network, and operates in an autoregressive manner. At each time step, based on the generated steering wheel angle sequence prefix, it is fused through a cross-attention mechanism. Predict the conditional probability distribution of the steering wheel angle at the next moment;

[0129] (203) Sequence sampling output: Sample the probability distribution of the final output of step (202) to obtain the initial steering wheel angle sequence for the next N time moments. ;

[0130] (204) Trajectory smoothing and kinematic verification: The initial steering wheel angle sequence output in step (203) is filtered by a smoothing filter. While preserving the overall trend and characteristics of the waveform, high-frequency noise is filtered out to obtain a smoothed angle sequence. Subsequently, the kinematic feasibility of the smoothed angle sequence is determined, and the maximum angle amplitude and maximum angle velocity constraints are set to ensure that the final output is an ideal steering wheel angle sequence. Control commands should match the vehicle's actual capabilities to avoid sudden changes.

[0131] Step 4) Dynamic driving risk field design

[0132] See Figure 3 It can be seen that by establishing a static field based on physical constraints and a dynamic field based on velocity vector correction, the complex traffic environment can be quantified into a computable potential energy distribution. The specific steps are as follows:

[0133] Construct a static road constraint potential field, and generate a lateral gravitational potential field with the centerline of the preceding lane as the gravitational target. Using the lane boundary as the repulsive force source, a static repulsive potential field is generated that increases exponentially with the deviation distance. The basic driving trajectory of the vehicle is constrained by these constraints, and the calculation formula is as follows:

[0134] (19)

[0135] (20)

[0136] in, This is the gravitational gain coefficient. The vehicle's lateral position. The lateral coordinates of the lane centerline; The repulsive force intensity at the road boundary. and These are the distances from the vehicle to the left and right boundaries, respectively. This refers to the parameter representing the range of influence of the repulsive force field.

[0137] Construct a time-varying repulsive field for dynamic obstacles, targeting the sensed first... A dynamic obstacle is used to construct an anisotropic Gaussian repulsive field modulated by relative velocity, which is located globally. Dynamic repulsive potential energy generated at the location for:

[0138] (twenty one)

[0139] in, The inherent danger potential peak value set according to the type of obstacle; This is the current position of the obstacle; These are the variance parameters that characterize the extent of risk diffusion in the longitudinal and lateral directions, respectively. This represents the relative velocity components between the vehicle and the obstacle in the longitudinal and lateral directions. As a speed-affecting factor;

[0140] Composite risk field synthesis involves linearly weighting and superimposing the aforementioned static potential field with the repulsive potential fields of all dynamic obstacles to obtain a global dynamic risk potential energy characterizing the overall danger level of the current location. :

[0141] (twenty two)

[0142] in, These are the weighting coefficients for the static gravitational field and the dynamic repulsive field, respectively. For the first The repulsive field weight coefficients of dynamic obstacles. By adjusting these three types of weights, driving styles can be flexibly defined.

[0143] Step 5) Design of nonlinear strategy mapping mechanism.

[0144] See Figure 4 As can be seen, the specific steps for transforming intention bias and environmental risk into continuously changing tactile control parameters based on digital models are as follows;

[0145] First, define the intention bias normalization factor. and environmental risk normalization factor :

[0146] (twenty three)

[0147] (twenty four)

[0148] Among them, the sensitivity coefficient Deviation tolerance and high-risk threshold These are configuration parameters preset based on driver characteristics or vehicle dynamics. This represents the actual steering wheel angle. The next moment's steering angle in the ideal steering wheel angle sequence;

[0149] Secondly, based on the intention bias normalization factor and environmental risk normalization factor Calculate the virtual stiffness coefficient and virtual damping coefficient , forming tactile interaction tuples :

[0150] (25)

[0151] (26)

[0152] in, The maximum virtual stiffness allowed by the system. Basic damping, This is the damping gain coefficient. As a regulating factor;

[0153] Finally, the interaction strategy is categorized into three modes. Based on the dynamic changes of decision factors α and β, the system exhibits smooth transition characteristics across three main interaction strategies. When β approaches 0, the system enters a compliant strategy, resulting in a light steering feel and giving the driver complete operational freedom. When β is approximately in the middle range, the system enters a damping stability strategy. If α is low, the system is in an auxiliary stability mode, with lower stiffness and moderately increased damping to provide a stable feel and eliminate operational vibrations. If α is high, the system is in a soft prompting mode, where stiffness begins to take effect, combined with increased damping, allowing the driver to experience a soft resistance, forming a weak tactile prompt to guide corrective actions. When β approaches 1, the system enters a rigid intervention strategy. If α is low, the system is in a rigid warning mode, with both stiffness and damping at high levels to assist the driver in precise hazard avoidance. If α is also high, the system is in a strong guidance mode, with stiffness approaching its maximum value, outputting a very strong self-centering torque to forcibly take over lateral control and avoid collision risks. Through adaptive switching of the above modes, the system achieves a continuous and smooth transition from complete compliance to mandatory intervention.

[0154] Step 6) Design of the physical principle of variable impedance control

[0155] See Figure 5 As can be seen, unlike the direct torque superposition mode used in traditional driving systems, this embodiment constructs a second-order mechanical system that includes virtual stiffness and virtual damping, incorporating tactile interaction elements. Transformed into the torque of the steer-by-wire system Its control law equation is as follows:

[0156] (27)

[0157] in, This represents the actual steering wheel angle. The next moment's steering angle in the ideal steering wheel angle sequence. The current steering wheel angular velocity, The angular velocity at the next moment in the ideal steering wheel angle sequence;

[0158] First item As a location-guided item, when the intention deviation is large and the risk is high, due to The increase in torque significantly guides the driver to follow the ideal trajectory; the second item This is a dynamically stabilizing factor. When environmental risks increase, due to... Increasing this torque results in increased resistance to steering wheel rotation, thus suppressing sudden steering maneuvers or involuntary vibrations by the driver in dangerous situations and improving handling stability. Therefore, real-time adjustment... and The system can dynamically reconstruct the mechanical impedance characteristics of the steering mechanism, thereby achieving implicit tactile guidance and interaction without severing the driver's mechanical connection.

[0159] Finally, the steer-by-wire motor executes an auxiliary interactive torque to adjust the equivalent moment of inertia and damping of the steering wheel, thereby providing tactile feedback to the driver.

[0160] See Figure 6 As can be seen, the intelligent vehicle variable impedance driving assistance control method based on generative trajectory prediction in this embodiment firstly collects the driver's physiological, operational, and environmental data simultaneously. Through feature encoding and time alignment, a multimodal spatiotemporal context tensor is constructed. A generative trajectory prediction model with a self-attention mechanism is used to calculate the driver's ideal steering wheel angle sequence in a future preset time domain. Secondly, the intentional deviation between the actual operation and the ideal trajectory is calculated and comprehensively evaluated in combination with the potential energy of the dynamic driving risk field. Three types of interaction strategies—compliance, damped stability control, or rigid intervention—are generated through a nonlinear strategy mapping mechanism, and a tactile interaction tuple containing virtual stiffness and virtual damping is dynamically output accordingly. Finally, a variable impedance control model is established based on the tuple, the auxiliary interaction torque is calculated, and the steer-by-wire motor is driven to execute. Thus, without directly superimposing the counteracting torque, the method provides the driver with continuous and smooth tactile feedback from gentle prompts to forced guidance, achieving intelligent assistance for the vehicle's lateral movement.

[0161] The advantages of this embodiment are:

[0162] This method employs a generative trajectory prediction model capable of understanding driver intentions with high accuracy and foresight. By integrating a dynamic driving risk field with a nonlinear strategy mapping mechanism, complex driving scenarios are quantified into continuously adjustable mechanical impedance parameters. Finally, an implicit and smooth tactile interaction channel is constructed through variable impedance control. Compared to traditional torque superposition methods, this approach enables the assistance system to adaptively match the timing, direction, and intensity of intervention with driver intentions and environmental risks, thereby ensuring safety while effectively mitigating human-machine control conflicts and improving operational smoothness and interactive experience.

[0163] The above description is merely a preferred embodiment of the present invention and does not limit the scope of the present invention. All equivalent structural changes made based on the description and drawings of the present invention are included within the scope of the present invention.

Claims

1. A method for intelligent vehicle variable impedance driving assistance control based on generative trajectory prediction, characterized in that, Includes the following steps: (1) Simultaneously collect the driver's physiological characteristics data, vehicle operation behavior data and external environment perception data, perform time-series alignment processing, and use deep neural networks to extract spatial features and temporal features respectively. Construct a multimodal spatiotemporal context tensor representing the spatiotemporal dependence of people-vehicle-road through feature splicing. (2) The driver intention recognition model is determined to be a generative trajectory prediction model. The multimodal spatiotemporal scenario tensor is input into the model, and the long-term causal dependency features in the data are mined by the self-attention mechanism to generate the ideal steering wheel angle sequence of the driver in the future preset time domain. (3) Calculate the intentional deviation between the actual operation and the ideal trajectory, and make a comprehensive evaluation by combining the potential energy of the dynamic driving risk field. Through the nonlinear strategy mapping mechanism, generate tactile interaction tuples including compliance strategy, damping stabilization strategy and rigid intervention strategy, and dynamically output tactile interaction tuples containing virtual stiffness coefficient and virtual damping coefficient. (4) Based on the tactile interaction tuple, a variable impedance control model is established, the auxiliary interaction torque used to adjust the steering feel is calculated, and the auxiliary interaction torque is executed by the steer-by-wire motor to adjust the equivalent rotational inertia and damping of the steering wheel, so as to realize tactile feedback to the driver.

2. The intelligent vehicle variable impedance driving assistance control method based on generative trajectory prediction according to claim 1, characterized in that, The construction of the multimodal spatiotemporal scenario tensor in step (1) specifically includes the following steps: (101) Data acquisition: The driver's gaze point coordinates and pupil diameter sequence are collected using in-vehicle vision sensors as physiological characteristic data; the steering wheel angle, yaw rate and longitudinal speed are collected using vehicle data bus as vehicle handling behavior data; and lane line deviation and relative distance to obstacles are obtained using on-board sensors as external environment perception data. (102) Data preprocessing and time-series alignment: First, the data collected in step (101) is converted to different units, and an improved t-text test is used to remove outlier values ​​from the heterogeneous multi-source data. Then, timestamp synchronization is performed, and all heterogeneous data are resampled to a uniform system sampling frequency using a linear interpolation algorithm. Z-score standardization is then performed to eliminate dimensional differences, forming a time-aligned multi-source input sequence. The multi-source input sequence specifically includes physiological feature sequences. Vehicle handling behavior sequence and external environment perception sequence ; (103) Spatial encoding of physiological features, The images are converted into eye-tracking heatmaps and head posture Euler angle images, and then physiological feature space feature vectors are extracted using a convolutional neural network. ; It consists of feature maps output by convolutional layers, and the coordinates on these feature maps are... pixel value at The calculation formula is: （1） in, For activation functions; and For eye-tracking heatmaps and head pose images in position Pixel value at; and The convolution kernel corresponding to the input channel is located at position Weight parameters at the location; For bias terms; and These are the height and width of the convolution kernel, respectively; and The row and column indices for sliding the convolution kernel; (104) Vehicle behavior timing coding, Input a Long Short-Term Memory (LSTM) network to extract temporal feature vectors of vehicle behavior It utilizes a gating mechanism to capture long-term operational habits, and its internal state update formula is as follows: （2） （3） （4） （5） （6） （7） in, Input the vehicle behavior at the current moment; This is the hidden state from the previous moment; These are the forget gate, input gate, and output gate, respectively. In cellular state, Candidate cell state; This is the corresponding weight matrix; This is the corresponding bias vector; It is the sigmoid function; This indicates element-wise multiplication; the final output is... for ; (105) Environmental perception feature mapping, for It is mapped to an environmental risk feature vector through a fully connected encoding layer. ; (106) Multimodal tensor construction and positional encoding, which converts feature vectors The sequence is concatenated along the feature channel dimension; sinusoidal positional encoding is then superimposed to preserve the temporal information of the sequence, ultimately constructing a multimodal spatiotemporal context tensor. : （8） in, For the position encoding matrix, This indicates a vector concatenation operation.

3. The intelligent vehicle variable impedance driving assistance control method based on generative trajectory prediction according to claim 1, characterized in that, The generative trajectory prediction model in step (2) adopts a transformer Encoder-Decoder architecture, and specifically includes the following steps: (201) Context encoding, which encodes the multimodal spatiotemporal context tensor. The input is fed into the encoder; the encoder consists of multi-layered stacked multi-head self-attention modules and a feedforward neural network, used to extract and fuse long-term temporal dependency features from the input tensor, and output a contextual semantic feature matrix. ; (202) Autoregressive trajectory decoding, which decodes the context semantic feature matrix The input is fed into the trajectory generation decoder; the trajectory generation decoder includes a masked self-attention module, a cross-attention module, and a feedforward neural network, and operates in an autoregressive manner; at each time step, based on the generated steering wheel angle sequence prefix, it is fused through the cross-attention mechanism. Predict the conditional probability distribution of the steering wheel angle at the next moment; (203) Sequence sampling output: Sample the probability distribution of the final output of step (202) to obtain the initial steering wheel angle sequence for the next N time moments. ; (204) Trajectory smoothing and motion verification: The smoothing filter is called to smooth the initial steering wheel angle sequence output in step (203). Filtering is performed to obtain a smoothed steering angle sequence. Then, the kinematic feasibility of the smoothed steering angle sequence is determined to verify whether it meets the vehicle's kinematic constraints and actual performance requirements. Finally, an ideal steering wheel angle sequence that meets the conditions is output. .

4. The intelligent vehicle variable impedance driving assistance control method based on generative trajectory prediction according to claim 3, characterized in that, The generative trajectory prediction model employs a multi-head self-attention mechanism that fuses relative positional biases. The calculation formula for this mechanism is as follows: （9） （10） in, These are the query matrix, key matrix, and value matrix obtained by linear projection of the input, respectively; This is the scaling factor; Given a learnable relative position bias matrix, its elements are... The value depends only on the position. and relative distance It is used to explicitly model the relative positional dependencies between temporal features; For the number of attention heads, each attention head Independently compute attention; This is the linear projection matrix used for the output after splicing.

5. The intelligent vehicle variable impedance driving assistance control method based on generative trajectory prediction according to claim 1, characterized in that, The design method for the dynamic driving risk field in step (3) is as follows: (301) Construct a static road constraint potential field, and generate a lateral gravitational potential field with the center line of the front lane as the gravitational target. Using the lane boundary as the repulsive force source, a static repulsive potential field is generated that increases exponentially with the deviation distance. The basic driving trajectory of the vehicle is constrained by these constraints, and the calculation formula is as follows: （11） （12） in, This is the gravitational gain coefficient. The vehicle's lateral position. The lateral coordinates of the lane centerline; The repulsive force intensity at the road boundary. and These are the distances from the vehicle to the left and right boundaries, respectively. The parameter represents the range of influence of the repulsive force field. (302) Construct a time-varying repulsive potential field for dynamic obstacles, targeting the sensed first... A dynamic obstacle is used to construct an anisotropic Gaussian repulsive field modulated by relative velocity, which is located globally. Dynamic repulsive potential energy generated at the location for: （13） in, The inherent danger potential peak value set according to the type of obstacle; This is the current position of the obstacle; These are the variance parameters that characterize the extent of risk diffusion in the longitudinal and lateral directions, respectively. This represents the relative velocity components between the vehicle and the obstacle in the longitudinal and lateral directions. As a speed-affecting factor; (303) Composite risk field synthesis: The static road constraint potential field and the time-varying repulsive potential field of all dynamic obstacles are linearly weighted and superimposed to obtain the global dynamic risk potential energy characterizing the comprehensive danger level of the current location. : （14） in, These are the static field weighting coefficients. For the first The repulsive field weight coefficient of a dynamic obstacle.

6. The intelligent vehicle variable impedance driving assistance control method based on generative trajectory prediction according to claim 1, characterized in that, The nonlinear strategy mapping mechanism in step (3) includes the following steps: First, define the intention bias normalization factor. and environmental risk normalization factor : （15） （16） Among them, the sensitivity coefficient Deviation tolerance and high-risk threshold These are configuration parameters preset based on driver characteristics or vehicle dynamics. This represents the actual steering wheel angle. The next moment's steering angle in the ideal steering wheel angle sequence; Secondly, based on the intention bias normalization factor and environmental risk normalization factor Calculate the virtual stiffness coefficient and virtual damping coefficient , forming tactile interaction tuples : （17） （18） in, The maximum virtual stiffness allowed by the system. Basic damping, This is the damping gain coefficient. As a regulating factor; Finally, the interaction strategy is divided into patterns based on decision factors. and The system exhibits smooth transition characteristics across three main categories of interaction strategies due to dynamic changes; when When the value approaches 0, the system enters a compliant strategy, the steering wheel feels light, and the driver has complete control; when... When the system is approximately in the middle range, it enters a damped stabilization strategy; at this time, if At lower levels, the system is in auxiliary stability control mode, with lower stiffness and moderately increased damping to provide a stable feel and eliminate operational jitter; if At higher speeds, the system enters a softer alert mode, where stiffness takes effect and, in conjunction with increased damping, the driver experiences a gentle resistance, creating a subtle tactile cue to guide corrective action; when... When the value approaches 1, the system enters a rigid intervention strategy; at this point, if The system is in a rigid warning mode with low stiffness and damping, assisting the driver in precise hazard avoidance; if When the stiffness is also high, the system is in a strong guidance mode, with the stiffness approaching its maximum value, outputting an extremely strong self-correcting torque, forcibly taking over lateral control to avoid collision risks; through the adaptive switching of the above modes, the system achieves a continuous and smooth transition from complete compliance to forced intervention.

7. The intelligent vehicle variable impedance driving assistance control method based on generative trajectory prediction according to claim 1, characterized in that, The variable impedance control model in step (4) includes the following: First, based on haptic interaction tuples Calculate the auxiliary interaction torque that ultimately acts on the steering column. Its control law is: （19） Among them, stiffness coefficient and virtual damping coefficient , This represents the actual steering wheel angle. The next moment's steering angle in the ideal steering wheel angle sequence. The current steering wheel angular velocity, The angular velocity at the next moment in the ideal steering wheel angle sequence; Secondly, the steer-by-wire motor executes auxiliary interactive torque to adjust the equivalent moment of inertia and damping of the steering wheel, thereby providing tactile feedback to the driver; when the system is in a compliant strategy... The system provides only minimal damping to maintain the driver's original control feel to the greatest extent possible; when in damped stability control strategy... Increased torque output becomes the dominant factor in torque output, used to suppress unstable or excessively rapid steering maneuvers; when in a rigid intervention strategy, Approaching The system generates a strong guiding torque to correct high-risk or severely deviated driving maneuvers.

Images