Control methods, electronic devices and storage media for autonomous vehicles
By acquiring user voice information, extracting acoustic and semantic features, generating emotion scores, and adjusting autonomous driving strategies, the problem of insufficient representation of the emotional intensity of user voice commands in existing technologies is solved, thereby improving the coordination between longitudinal and lateral strategies for autonomous vehicle control.
Patent Information
- Application Number
- CN202611425809.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-14
- Publication Date
- 2026-11-17
AI Technical Summary
Existing autonomous vehicle control systems struggle to accurately represent the emotional intensity differences in user voice commands, resulting in insufficient coordination between longitudinal and lateral driving strategies and control behaviors that do not conform to user intentions.
By acquiring user voice information, extracting acoustic and semantic features, generating user emotion scores using emotion scoring models, and combining user historical driving data and professional driver data to adjust autonomous driving strategies, the coordination of longitudinal and lateral control is achieved.
It improves the coordination and coherence of longitudinal and lateral driving strategies for autonomous vehicle control, making the vehicle response more closely match the user's real-time needs and driving intentions.
Smart Images

Figure CN123104665A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of intelligent vehicle control, and more particularly to a control method, electronic device and storage medium for an autonomous vehicle. Background Technology
[0002] With the development of intelligent vehicles and voice interaction technology, existing vehicle control typically uses voice commands to perform operations such as speed adjustment, lane change control, or driving mode switching.
[0003] However, existing solutions mainly rely on discrete command mapping or preset mode switching, which makes it difficult to reflect the differences in emotional intensity in voice expression and is insufficient in the coordinated adjustment of longitudinal and lateral driving strategies.
[0004] Therefore, how to more accurately represent the intensity of user intent based on voice commands has become a technical problem that urgently needs to be solved. Summary of the Invention
[0005] The control method, electronic device, and storage medium for autonomous vehicles provided in this application aim to solve the aforementioned technical problems. This method, upon receiving voice information for controlling the autonomous driving of a vehicle, combines the user's emotional state reflected in the voice information to characterize the user's level of urgency, and adjusts the vehicle's autonomous driving strategy accordingly. This allows vehicle control to move beyond fixed operations corresponding to discrete voice commands, and instead better align with the user's real-time needs and driving intentions.
[0006] In a first aspect, embodiments of this application provide a control method for an autonomous vehicle, including:
[0007] Acquire user-input voice information, which is used to control the vehicle's autonomous driving.
[0008] Based on voice information, the user's emotion score is determined, which is used to indicate the degree of agitation in the user's emotions.
[0009] Adjust the vehicle's autonomous driving strategy based on emotion scores.
[0010] In one possible embodiment, acoustic features and semantic features are extracted from speech information;
[0011] The acoustic and semantic features are input into a pre-defined emotion scoring model to obtain the user's emotion score.
[0012] In one possible embodiment, the acoustic features include at least one of speech rate features, volume features, and pitch features.
[0013] In one possible embodiment, extracting semantic features from speech information includes:
[0014] Extracting text information from speech information;
[0015] Text information is input into a preset semantic extraction model to obtain semantic features.
[0016] In one possible embodiment, the sentiment scoring model is shown in the following formula:
[0017]
[0018] in, Rate your mood. For activation function, As the first weighting coefficient, Acoustic characteristics, This is the second weighting coefficient. For semantic features, This is a bias term.
[0019] In one possible embodiment, the vehicle's autonomous driving strategy is adjusted based on emotion scores, including:
[0020] Determine the first and second strategies;
[0021] The first and second strategies, along with the emotion score, are input into a pre-defined strategy fusion model to obtain and output the third strategy.
[0022] In one possible embodiment, the first strategy is the output of a first strategy model, which is trained based on the user's historical driving data.
[0023] In one possible embodiment, the first policy model includes a first loss function, which is used to optimize the first policy model for safety and comfort.
[0024] In one possible embodiment, the first loss function is shown in the following formula:
[0025]
[0026] in, The first loss value, For the fitting term, For the policy vector, For data labels, For comfort constraints, For security constraints, , These are the weighting coefficients.
[0027] In one possible embodiment, the second strategy is the output of a second strategy model, which is trained based on driving data from professional drivers.
[0028] In one possible embodiment, the second policy model includes a second loss function used to optimize the efficiency of the second policy model.
[0029] In one possible embodiment, the second loss function is shown in the following formula:
[0030]
[0031] in, This is the second loss value. For the fitting term, For the policy vector, For data labels, For speed, Speed limits for roads, For safety constraints, Minimum collision time, , These are the weighting coefficients.
[0032] In one possible embodiment, the policy fusion model is shown in the following formula:
[0033]
[0034] in, As the third strategy, Rate your mood. As the first strategy, This is the second strategy.
[0035] In one possible embodiment, it also includes:
[0036] If the security check of the third strategy fails, the third strategy will be modified.
[0037] In one possible embodiment, it also includes:
[0038] Based on voice information, a first prompt message is generated, which is used to remind the user to pay attention to safety.
[0039] Play the first prompt message.
[0040] In one possible embodiment, after obtaining the user's voice input, the method further includes:
[0041] Delete the voice message if it is determined that the user does not have control permissions.
[0042] Generate a second prompt message to inform the user that they do not have control permissions;
[0043] Play the second prompt message.
[0044] Secondly, embodiments of this application provide an electronic device, including: a memory and a processor;
[0045] The memory stores the instructions that the computer executes;
[0046] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.
[0047] Thirdly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.
[0048] By acquiring voice information used to control the autonomous driving of vehicles and determining the user's emotion score based on the voice information, the intensity of emotion in the voice expression can be incorporated into the vehicle control decision, thereby enabling the autonomous driving strategy to be adjusted to better meet actual needs based on the user's emotional state. By adjusting the vehicle's autonomous driving strategy based on the emotion score, the coordination and coherence of longitudinal and lateral driving control can be improved. Attached Figure Description
[0049] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0050] Figure 1 This is a schematic diagram illustrating an application scenario of the control method for autonomous vehicles provided in the embodiments of this application;
[0051] Figure 2 A flowchart illustrating the control method for an autonomous vehicle provided in this application embodiment. Figure 1 ;
[0052] Figure 3 A flowchart illustrating the control method for an autonomous vehicle provided in this application embodiment. Figure 2 ;
[0053] Figure 4 This is a schematic diagram of the structure of the control device for an autonomous vehicle provided in an embodiment of this application;
[0054] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.
[0055] The accompanying drawings have illustrated specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation
[0056] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0057] Figure 1 This is a schematic diagram illustrating an application scenario of the control method for autonomous vehicles provided in the embodiments of this application, such as... Figure 1 As shown, the driver issues commands to the vehicle via voice, which are usually in natural language (such as...). Figure 1 The "hurry up" command in the system refers to the vehicle receiving driver commands via an in-vehicle voice interaction system. The central control unit then combines this information with perception data to execute actions such as acceleration, deceleration, lane changing, or driving mode adjustment. This type of system is typically deployed within a control architecture comprised of an in-vehicle microphone, voice recognition module, driving strategy decision-making module, and vehicle actuators.
[0058] Existing solutions often employ a direct mapping from voice content to control actions, or map voice commands to preset driving modes, with mode parameters then driving the vehicle's execution. This approach typically focuses only on the literal meaning of the command, lacking differentiation in the urgency, intensity, and emotional nuances of the voice expression. This results in the same command producing nearly identical control outcomes under different tones. Furthermore, control strategies often adjust only a single driving dimension, lacking coordination between longitudinal speed and acceleration and lateral lane changes and following decisions, easily leading to inconsistencies between action responses and driving intentions. After a task is completed, the system requires manual restoration to its original state, as control behavior rarely naturally reverts. These issues make the vehicle's response in complex road environments feel abrupt, impacting the driving experience and undermining user trust in autonomous driving interaction.
[0059] Based on the aforementioned technical challenges, voice information is used as the carrier of user intent input, and emotion scoring is introduced into the vehicle's autonomous driving control process as an intermediate variable characterizing the degree of irritability. Voice information is first used to identify the user's current control intent, and then combined with the emotional differences in the voice expression to generate a corresponding emotion score. The vehicle's autonomous driving strategy is adjusted around this score, causing control actions to change according to the intensity of the user's intent, and maintaining consistency between longitudinal and lateral strategies during execution.
[0060] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.
[0061] Figure 2 A flowchart illustrating the control method for an autonomous vehicle provided in this application embodiment. Figure 1 The methods described above in the embodiments of this application can be applied to any electronic device. For example... Figure 2 As shown, the method includes:
[0062] S201. Obtain voice information input by the user, which is used to control the autonomous driving of the vehicle.
[0063] For example, voice information can be acquired by a voice acquisition device installed in the vehicle. When the vehicle is in autonomous driving, assisted driving, or driving mode switching mode, when the user issues a voice command related to vehicle control, the controller acquires the voice command and forms voice information, which is then transmitted to the subsequent processing module. The voice information can be a single short voice command or a continuous input segment, as long as the voice content is related to the vehicle's autonomous driving control. By acquiring this voice information, the system can directly incorporate user control requests into the autonomous driving control process without changing the user's original operating methods, providing a basic input for determining subsequent emotion scores. Based on the above analysis, this step completes the voice input acquisition in the context of autonomous driving control, and its output is voice data that can be used for voice analysis and strategy linkage processing.
[0064] S202. Based on voice information, determine the user's emotion score, which is used to represent the user's level of agitation.
[0065] For example, a score representing the user's emotional state can be determined based on voice information. The emotion score can be a discrete level or a continuously varying numerical value. The numerical range represents the strength of the user's current level of agitation in their voice; a higher score indicates greater agitation, and a lower score indicates less agitation. This emotion score serves as an intermediate quantity, converting the user's emotional state in their voice into a numerical parameter that the control system can directly access. Subsequent autonomous driving strategies are then adjusted around this parameter. Based on the above analysis, this step completes the conversion from voice input to a quantitative emotion parameter, outputting an agitation score that can be used for control decisions.
[0066] S203. Adjust the vehicle's autonomous driving strategy based on emotion scores.
[0067] For example, after receiving an emotion score, the controller maps the score to a set of autonomous driving strategy parameters and updates the corresponding strategy based on the magnitude of the emotion score. Specifically, strategy adjustment can be achieved by updating at least one of the following: target speed, control acceleration, distance parameters, lane change response parameters, or trajectory control parameters. The updated parameters are then sent to the vehicle's actuators or autonomous driving decision-making module, thereby altering the vehicle's autonomous driving behavior in the current driving scenario. In this way, the vehicle's autonomous driving strategy is no longer fixedly mapped based solely on the literal content of the voice, but rather adjusts the driving control method in conjunction with the level of agitation in the user's voice, ensuring that the autonomous driving behavior aligns with the intensity of the user's intent. Based on the above analysis, this step completes the mapping and updating of emotion scores to vehicle control parameters, outputting the adjusted autonomous driving strategy and its corresponding control commands.
[0068] This application provides a control method for autonomous vehicles, including acquiring user-input voice information, determining the user's emotion score based on the voice information, and adjusting the vehicle's autonomous driving strategy based on the emotion score. In this application, after the voice information is collected as a control input, it is first analyzed to obtain an emotion score representing the degree of agitation. This score is then written into the autonomous driving control link as a basis for strategy adjustment, enabling relevant control parameters to change synchronously with the strength of the user's tone. This processing method establishes a correspondence between user emotions and the vehicle's autonomous driving decisions, allowing control actions to be updated in conjunction with the degree of agitation in the voice, forming an executable and retrievable control flow. It should be understood that the above example is merely illustrative and not limiting. Without departing from the technical concept of this application, the voice acquisition method, emotion score generation method, and the specific form of strategy adjustment parameters can all be adapted to the vehicle platform, sensor configuration, and control architecture.
[0069] Figure 3 A flowchart illustrating the control method for an autonomous vehicle provided in this application embodiment. Figure 2 .like Figure 3 As shown, the method includes:
[0070] S301. Obtain voice information input by the user, which is used to control the autonomous driving of the vehicle.
[0071] S302. Extract acoustic and semantic features from speech information.
[0072] S303. Input the acoustic features and semantic features into the preset emotion scoring model to obtain the user's emotion score.
[0073] The voice information refers to the voice information input by the driver during in-vehicle voice interaction. It includes both acoustic information that characterizes the vocal state and semantic information that characterizes the expressed meaning. Acoustic features are features extracted from the voice information that reflect the physical properties of the voice, used to characterize the vocal state in dimensions such as speech rate, volume, pitch, and pause rhythm. Semantic features are features extracted from the voice information that reflect the semantic content and expressive tendencies, used to reflect the user's level of agitation in dimensions such as word choice, sentence structure, and command intensity. The emotion scoring model is a pre-defined multimodal mapping model, which can employ a neural network model, a support vector regression model, or other models capable of feature fusion and emotion regression calculation. In practical applications, other models can also be selected, and this application does not impose any limitations on this.
[0074] In emotion scoring, the acquired speech information is first preprocessed, including noise reduction, endpoint detection, and framing, to obtain stable speech information. Then, acoustic features such as Mel-frequency cepstral coefficients, fundamental frequency, energy envelope, speech rate variation, and pitch fluctuation are extracted based on the speech information. Speech recognition is then performed to generate text. Semantic features are further extracted by combining word vectors, syntactic features, sentiment word distribution, and the proportion of imperative sentences. After dimensionality alignment, the acoustic and semantic features are input into the emotion scoring model. The model jointly encodes and fuses these two types of features, outputting an emotion score corresponding to the user's current level of agitation. The score can be a continuous value or a discrete level value.
[0075] By feeding acoustic features reflecting vocalization state and semantic features reflecting text meaning into the emotion scoring model, the model can simultaneously use speech expression and semantic content to comprehensively represent user emotions, thereby outputting an emotion score that matches the actual tone of voice and providing input for subsequent adjustments to autonomous driving strategies based on emotion scores.
[0076] S304. Adjust the vehicle's autonomous driving strategy based on emotion scores.
[0077] Based on the foregoing embodiments, the acoustic features further include at least one of speech rate features, volume features, and pitch features.
[0078] Among them, speech rate features characterize the speed of speech, reflecting whether the user speaks quickly; volume features characterize changes in loudness, reflecting whether the user's volume increases; and pitch features characterize changes in high and low pitch, reflecting changes in the user's tone. These acoustic features are all acoustic dimensional information extracted from speech information and are used as input features in the emotion scoring model for calculation.
[0079] In this solution, after acquiring the user's input voice information, the in-vehicle voice interaction system first preprocesses the voice information to perform noise reduction, endpoint detection, and framing. Then, it extracts acoustic features from the valid voice information. For speech rate features, the system can calculate based on the number of syllables per unit time, phoneme duration, or the relationship between the duration of the voice information and the length of the content. For volume features, the system can characterize them based on the energy amplitude, sound pressure level, or short-time average amplitude of the voice signal. For pitch features, the system can characterize them based on fundamental frequency trajectory, formant variations, or pitch contours. The extracted acoustic features can be input into the emotion scoring model independently or together with semantic features to output the corresponding emotion score.
[0080] When only speech rate features are extracted, the system characterizes the user's emotional agitation by the speed of speech output; when only volume features are extracted, the system represents emotional changes by the volume of the voice; and when only pitch features are extracted, the system reflects the tone of voice by changes in pitch. If two or three of the speech rate, volume, and pitch features are used simultaneously, the system normalizes each feature and combines them into a feature vector, which is then input into a preset model to complete the joint scoring. The combination method of acoustic features can be set according to the vehicle's computing power, recognition accuracy requirements, and model training configuration. In practical applications, other forms of feature combination can also be selected, which this application does not limit.
[0081] By using at least one of speech rate, volume, and pitch as acoustic input, the emotion scoring model can obtain a representational basis for the urgency, intensity of urging, and changes in tone of voice, and generate an emotion score that better reflects the intensity of the current driving intention, thus providing input for adjusting autonomous driving strategies. With this acoustic feature construction method, the system's ability to distinguish different tone states is improved, and the consistency between the emotion score and the driver's actual control intention is correspondingly enhanced.
[0082] Based on the aforementioned embodiments, further, semantic features are extracted from the speech information, including: extracting text information from the speech information; and inputting the text information into a preset semantic extraction model to obtain semantic features.
[0083] The text information refers to the written content formed after speech recognition of the speech information, which carries the driver's instructions, modifiers, and emotional expressions. The semantic extraction model is a pre-trained or pre-configured text analysis model. Its input is text information, and its output is semantic features that can represent the user's intention or emotional expression. The semantic features can be semantic representations in vector form, or feature values used to represent the emotional tendency of the text, the intensity of urging, or the intention to control.
[0084] In this embodiment, after receiving voice information, the vehicle's in-vehicle voice interaction system first converts the voice information into text information using a speech recognition module. The speech recognition module can employ an end-to-end recognition network, a joint recognition structure based on acoustic and language models, or other model structures capable of speech-to-text conversion. After obtaining the text information, it is input into a semantic extraction model. This model can employ an encoding network based on a pre-trained language model, a recurrent neural network, a Transformer network, or a combination thereof, and outputs a feature representation corresponding to the text's semantics. This feature representation reflects the type of control request, emotional language, tone intensity, and semantic tendency related to autonomous driving in the statement, and further serves as one of the inputs to a subsequent emotion scoring model.
[0085] In practical applications, the semantic extraction model can be deployed within the vehicle's central control unit or on an edge computing node connected to the vehicle. The speech recognition module and the semantic extraction model exchange text information via message queues or interface data frames. The model's parameters can be trained and updated based on driving scenario corpora to enhance its semantic discrimination ability for expressions such as "hurry up" and "change lanes immediately." This application does not impose any limitations on this.
[0086] In this way, voice information is first transcribed into text information, and then semantic features are generated by the semantic extraction model. This enables the system to directly obtain text semantic representations related to driving intentions and emotional expressions, and provides an input basis for subsequent emotion scoring. This allows the autonomous driving strategy to be adjusted in combination with semantic-level intention differences.
[0087] In one possible implementation, the sentiment rating model is shown in the following formula:
[0088] in, Rate your mood. For activation function, As the first weighting coefficient, Acoustic characteristics, This is the second weighting coefficient. For semantic features, This is a bias term.
[0089] Here, α represents the quantification of the user's level of agitation; the larger the value, the stronger the sense of urgency in the driver's voice. σ is a mapping function used to convert the linear combination result into an output value suitable for representing the degree of emotion. Used to adjust the contribution ratio of acoustic features to emotion scores. b is used to adjust the contribution ratio of semantic features in the sentiment score, and b is used to perform baseline correction on the linear combination results.
[0090] In practical implementation, the system extracts acoustic features from speech information. and semantic features Then, the two types of features are input into their respective weighting units. and The training-obtained floating-point parameter matrix or scalar parameters can be multiplied with corresponding features to form a linear combination term, which is then superimposed with a bias term b to form an intermediate result. The intermediate result is input into an activation function σ to output an emotion score α. The activation function can be a sigmoid function, a hyperbolic tangent function, or an equivalent nonlinear mapping to constrain the output within a preset range. The emotion scoring model can be deployed in an in-vehicle central control unit or voice interaction processing module and can be accessed through a parameter file stored in non-volatile memory. In practical applications, other training configurations for the model parameters can also be selected; this application does not impose any limitations on this.
[0091] When a driver issues control requests such as "faster" or "accelerate," the system first determines the control intent based on the speech content. Then, it uses the model to fuse the acoustic fluctuations and semantic urgency in the speech to obtain a corresponding emotion score, which is used as the basis for subsequent adjustments to the autonomous driving strategy. This approach allows acoustic and semantic information to participate in emotion determination, enabling the output of continuous and controllable quantitative results. This ensures that the intensity of vehicle control aligns with the intensity of user expression and provides a stable input basis for subsequent strategy adjustments.
[0092] In one possible implementation, the autonomous driving strategy of the vehicle is adjusted based on the emotion score, including: determining a first strategy and a second strategy; inputting the first strategy, the second strategy, and the emotion score into a preset strategy fusion model to obtain and output a third strategy.
[0093] The first strategy is one of the strategies used in the strategy fusion process together with the second strategy, and is output by the first strategy model; the second strategy is one of the strategies used in the strategy fusion process together with the first strategy, and is output by the second strategy model; the strategy fusion model is a preset model used to fuse the first strategy, the second strategy, and the emotion score to generate the third strategy; the third strategy is the resulting strategy output by the strategy fusion model, which is used to actually adjust the autonomous driving behavior of the vehicle.
[0094] In implementation, after receiving a strategy adjustment instruction based on emotion score, the controller first retrieves a first strategy and a second strategy corresponding to the current driving scenario from a preset strategy library. The strategy library can store longitudinal and lateral control parameters corresponding to different emotion score ranges. The first strategy can represent the vehicle's acceleration, deceleration, or following decisions, while the second strategy can represent the vehicle's lane changing, lane keeping, or path tracking decisions. Both are generated by independent strategy networks or can be directly output from a rule parameter table. After obtaining the emotion score, the controller inputs the first strategy, the second strategy, and the emotion score into a strategy fusion model. The strategy fusion model performs weighted calculations, gating selection, or feature concatenation operations on the two strategies based on the emotion score to output a third strategy that matches the user's current emotion intensity. The strategy fusion model can be deployed as a neural network model, a weighted fusion model, or a parameterized model with lookup table functionality. In practical applications, other model forms can also be selected, and this application does not limit this.
[0095] The resulting third strategy can be continuously adjusted according to changes in emotion scores, ensuring consistency between longitudinal and lateral control. When the vehicle executes this third strategy, it can obtain autonomous driving control results adapted to the current emotional state, thus making the control response more aligned with the user's intentions.
[0096] Building upon the foregoing embodiments, the first strategy is further defined as being output by a first strategy model, which is trained based on the user's historical driving data. The first strategy model outputs a basic strategy that matches the user's driving habits, and can serve as the initial control basis in the vehicle's autonomous driving strategy.
[0097] The first strategy model is trained based on the user's historical driving data, which characterizes the user's handling preferences and control tendencies during past driving experiences. Historical driving data can include records of acceleration, deceleration, following, lane changing, and speed maintenance generated by the user in urban roads, highways, congested areas, and ramp lane-changing scenarios. The system organizes these records into time-series samples and inputs them into the first strategy model for training, enabling the model to learn the user's typical driving style under different road conditions. The first strategy model can be constructed using a neural network-based strategy network. Its input receives vehicle speed, longitudinal acceleration, lateral acceleration, steering input, changes in distance to other vehicles, and corresponding scenario identifiers from the historical driving data. Its output generates first strategy parameters reflecting the user's basic driving preferences. After training, this model can generate relatively stable, conservative, or user-compliant autonomous driving strategies based on the user's historical driving characteristics. In practical applications, other model structures can also be selected, which this application does not limit.
[0098] During operation, the vehicle control system first loads the corresponding historical driving data based on the user's identity and then calls the pre-trained first strategy model to generate a first strategy. This first strategy is then used as the base strategy input into subsequent strategy fusion processes, enabling the vehicle to adjust its strategy within the constraints of the user's existing driving habits after receiving the current voice intent and emotion score. Because the first strategy is derived from the user's long-term driving records, the vehicle's speed changes, following distance, and lane change response in autonomous driving control can remain consistent with the user's past preferences, thus making the final output control strategy both personalized and continuous.
[0099] After adopting the above method, the first strategy no longer relies on fixed preset parameters, but is trained by the user's historical driving data. It can reflect the user's basic driving style, keep the autonomous driving control results consistent with the user's habits, and provide a stable basic strategy input for subsequent strategy fusion based on emotion scoring.
[0100] In one possible implementation, the first policy model includes a first loss function, which is used to optimize the first policy model for both safety and comfort.
[0101] The first strategy model is used to output a first strategy that matches the user's historical driving data with the current emotion score. The first loss function is embedded in the parameter update process of the first strategy model as a training objective, and forms an error constraint between the strategy output corresponding to the training samples and the preset objective. The first loss function is used to optimize the safety and comfort of the first strategy model. Specifically, it incorporates safety and comfort constraints into the loss calculation and balances different constraints through a weighted method, so that the model can simultaneously suppress unsafe control outputs and overly abrupt control changes when learning driving strategies.
[0102] In practical implementation, the safety constraint can correspond to penalties for indicators such as vehicle longitudinal speed, acceleration, distance between vehicles, or lateral lane change risk. When the strategy output by the first strategy model exceeds the preset safety boundary, the loss value increases accordingly. The comfort constraint can correspond to penalties for indicators such as acceleration change rate, steering angle change rate, or braking impact. When the strategy causes a sharp change in control, the loss value increases accordingly. The first loss function can be expressed as a weighted sum of the basic strategy error term, the first safety term, and the first comfort term. The weights of each term are preset according to the vehicle type, road scenario, and driving mode, and can be iteratively optimized using sample playback during the training phase. This application is not limited to this form; in practical applications, this loss function can also be extended by combining other constraint terms, which this application does not limit.
[0103] With the above settings, the first strategy model can learn safety boundaries and smooth control rules simultaneously during training, so that the output first strategy can maintain stable and smooth control characteristics while satisfying driving intentions, and maintain more consistent strategy performance in complex road environments. This makes the subsequent integration with the second strategy smoother, and the vehicle's actions are more in line with the needs of autonomous driving control.
[0104] In one possible implementation, the first loss function is shown in the following formula:
[0105] in, The first loss value, For the fitting term, For the policy vector, For data labels, For comfort constraints, For security constraints, , These are the weighting coefficients.
[0106] in, Used to characterize the overall optimization objective of the first policy model during the training process. This is used to measure the deviation between the policy vector output by the model and the data labels corresponding to the training samples. `pred` represents the control policy output by the model under the current sample conditions, and `label` represents the target policy that matches the current scene. Used to constrain the smoothness of changes in vehicle control variables. This is used to introduce a safety penalty when the remaining time after a collision is less than a set threshold. , These are used to adjust the influence of comfort constraints and safety constraints on the total loss, respectively.
[0107] In its implementation, the first strategy model receives scene features and target control labels corresponding to the user's historical driving data, outputs a strategy vector corresponding to the current scene, and substitutes this strategy vector and label into MSE(pred,label) to calculate the fitting error. Simultaneously, the model calculates Jerk based on the acceleration change rate corresponding to the strategy vector and uses it as a comfort constraint term in the loss calculation. For the safety constraint term, it calculates max(0, 1.5 -TTC) based on the time remaining between the vehicle and the target ahead (TTC). This term is zero when TTC is greater than or equal to 1.5, and generates a positive penalty value when TTC is less than 1.5, thus enabling the model to impose stronger constraints on near-field risk states during training. Weight coefficients , Scalar parameter configuration is adopted, and the parameters are preset before training or adjusted according to the sample distribution during training to achieve a balance between fitting accuracy, comfort, and safety in the loss function. In practical applications, the weighting coefficients can also be selected with other parameter values based on different vehicle models, road environments, or driving styles; this application does not impose any limitations on this.
[0108] During the training process, after the training samples are constructed from historical driving data, the model first updates its parameters based on the error between the policy vector and the label. Then, it corrects the update direction by incorporating the constraints corresponding to Jerk and TTC, ensuring that the output policy maintains consistency with the target control while limiting excessively rapid control changes and high-risk situations. Through this loss structure, the first policy model can output a smoother and safer control policy, providing a foundation for subsequent adjustments to the autonomous driving policy fused with emotion scoring. By adopting this loss function, the model training can simultaneously consider fitting effect, comfort, and safety, making the policy output more in line with actual driving control needs, thereby improving the stability and consistency of policy adjustments.
[0109] In one possible implementation, the second strategy is the output of a second strategy model, which is trained based on driving data from professional drivers.
[0110] In this application, the second strategy model is used to learn the driving operations of professional drivers under different road conditions, speed ranges, and traffic densities, and maps the learned high-efficiency driving patterns into a second strategy. The professional driver's driving data can include time-series data related to vehicle control, such as steering wheel angle, throttle opening, braking intensity, speed changes, lateral displacement, and following distance. This data, combined with corresponding scene annotations, can be input into the model for training, enabling the model to establish a correspondence between scene features and driving decisions. The second strategy model can employ temporal neural networks, recurrent neural networks, attention networks, or other model structures capable of handling continuous driving trajectories. During training, supervised learning is used to output a strategy vector that conforms to the driving style of a professional driver. The output second strategy represents a tendency towards faster response, more decisive decision-making, and more tightly coupled actions in autonomous driving control. In practical applications, other network structures can also be selected for this model, and this application does not impose any limitations on this.
[0111] When the vehicle's autonomous driving strategy needs to be adjusted based on the user's current intent, the second strategy output by the second strategy model serves as one of the inputs for strategy fusion, participating in subsequent fusion calculations along with the first strategy and emotion score. Since the second strategy is derived from training data on professional drivers, its output provides a relatively stable and efficient driving reference, ensuring that the fused control result satisfies changes in user intent intensity while also taking into account the rapid response characteristics of professional driving styles. This approach allows the second strategy to be learned directly from training data without relying on manually set rules, thus enabling it to output strategy representations consistent with professional driving habits in different driving scenarios and providing a computable strategy basis for the vehicle's final control commands.
[0112] By employing a second strategy model trained on data from professional drivers, the system can output a second strategy with professional driving characteristics. This second strategy serves as an effective input during the autonomous driving strategy adjustment process, enabling the vehicle to obtain a strategy supplement that more closely resembles high-level driving behavior in voice-interaction-triggered control scenarios. Since this second strategy reflects the driving style of professional drivers in complex scenarios, its combination with the first strategy ensures that vehicle control results remain consistent in terms of response speed, action continuity, and scenario adaptability, thereby improving the naturalness and executability of autonomous driving strategy adjustments.
[0113] In one possible implementation, the second policy model includes a second loss function, which is used to optimize the efficiency of the second policy model.
[0114] The second strategy model is a policy network trained on driving data from professional drivers, used to output a second strategy that matches an efficient driving style. The second loss function is an optimization function used to constrain the model's training objective; its role is to guide the second strategy model towards higher driving efficiency during training and to ensure the output strategy remains consistent with the rapid decision-making characteristics of professional drivers. Efficiency refers to the vehicle completing driving control with shorter response times, more reasonable speed changes, and more coherent action transitions while meeting safety boundary conditions.
[0115] In its implementation, the second loss function jointly models speed-related constraints, safety-related constraints, and a fitting term related to strategy deviation, and then sums them by weight to form the total loss value. Speed-related constraints characterize target vehicle speed, acceleration changes, or speed tracking errors; safety-related constraints characterize limitations related to distance to the vehicle ahead, collision risk, or minimum safe distance; and the fitting term ensures the model output closely approximates the target strategy vector in the professional driver's data. During training, the professional driver's trajectory data, speed data, and control commands are input into the second strategy model. The network parameters are then updated via backpropagation using the second loss function, ensuring the model balances efficiency and safety boundaries when outputting acceleration, deceleration, or lane-changing decisions.
[0116] The second loss function imposes efficiency constraints on the training output, enabling the second policy model to learn a more compact control decision pattern and output a policy that is more in line with efficient driving scenarios under the same safety requirements. This structure allows the model to provide a second policy with faster response and more continuous execution when subsequently fused with the first policy, thereby achieving higher operational efficiency and more stable policy consistency in the autonomous driving control process.
[0117] In one possible implementation, the second loss function is shown in the following formula:
[0118] in, This is the second loss value. For the fitting term, For the policy vector, For data labels, For speed, Speed limits for roads, For safety constraints, Minimum collision time, , These are the weighting coefficients.
[0119] in, Used to characterize the overall loss during the training process of the second strategy model, the larger the value, the more significantly the current output deviates from the target in terms of fitting accuracy, speed efficiency, and safety constraints. This is used to measure the mean squared error between the policy vector and the data labels. `pred` corresponds to the sequence of control parameters output by the model, and `label` corresponds to the target policy obtained from professional driver driving data. `v` represents the vehicle speed corresponding to the output of the second policy model. The speed limit is the allowable speed limit under current road conditions, and together they constitute a speed efficiency constraint. This is used to represent the minimum collision time among all potential collision targets under the current strategy. Used in A penalty is introduced when the value falls below the threshold; 0.8 is the preset safety threshold. , These are used to adjust the influence of the speed efficiency constraint and the safety constraint, respectively, so that the loss function can make a unified trade-off between different constraints during training.
[0120] In its implementation, the second strategy model uses the road speed limit, target distance, relative speed, and trajectory prediction results output by the vehicle perception module as input for constraint calculation. It then aligns the model-generated strategy vector with the labeled data to calculate the Mean Squared Error (MSE). The speed efficiency term is constructed by comparing the ratio of the current predicted speed to the road speed limit. When the predicted speed is lower than the speed limit, this term still imposes constraints on deviations from higher traffic efficiency. The safety constraint term calculates the minimum collision time by combining the trajectories of the preceding vehicle, side vehicle, and potential conflict targets, and generates a non-zero penalty when the time is less than 0.8. The weight coefficient can be set as a fixed constant, or different values can be used in the early training and convergence phases to adapt to control preferences under different road scenarios. In practical applications, other values can also be selected for this coefficient, which this application does not limit.
[0121] The loss function simultaneously constrains the policy fitting effect, driving efficiency, and collision risk during training, so that the control results output by the second policy model can not only closely resemble professional driving samples, but also maintain a speed performance close to the road's permissible limit, and promptly increase the penalty signal when danger approaches, thereby guiding the model to learn policy parameters that are more in line with the characteristics of efficient driving and maintain safety boundaries.
[0122] After adopting this loss function, the second strategy model can output a strategy vector that is highly consistent with the training samples and takes into account traffic efficiency when facing voice intent-driven driving control tasks. It also forms a clear constraint on the case of insufficient collision time, so that the fused autonomous driving strategy can maintain a more stable balance between responsiveness and safety.
[0123] In one possible implementation, the policy fusion model is shown in the following formula:
[0124]
[0125] in, As the third strategy, Rate your mood. As the first strategy, This is the second strategy.
[0126] The policy fusion model linearly fuses the first and second policies based on emotion scores to form a third policy with a continuous transition. Pfinal represents the fused autonomous driving control output, Pbase represents the basic driving policy, Pagg represents the aggressive driving policy, and α represents the degree of agitation corresponding to the voice emotion, and is used as a fusion weight in the calculation.
[0127] In this implementation, the first strategy is output by a first strategy model trained based on the user's historical driving data, and the second strategy is output by a second strategy model trained based on professional driver driving data. The first strategy model corresponds to a more stable control tendency, suitable for normal driving needs when the emotion score is low; the second strategy model corresponds to a more aggressive control tendency, suitable for stronger response needs when the emotion score is high. The strategy fusion model uses a linear weighting method to interpolate the two types of strategies. The larger the value of α, the more the output third strategy is biased towards the second strategy; the smaller the value of α, the more the output third strategy is biased towards the first strategy, thereby enabling longitudinal speed, acceleration, and lateral lane change decisions to be adjusted synchronously with the emotion score.
[0128] In practical implementation, the emotion score α can be limited to between 0 and 1. After obtaining this score, the control unit writes it into the fusion module and performs numerical calculations with Pbase and Pagg to obtain Pfinal, which is then output to the decision interface corresponding to the vehicle actuator. This fusion relationship does not rely on fixed threshold switching but uses continuous interpolation calculation. Therefore, when the driver's tone gradually changes from calm to urgent, the vehicle control strategy will not change abruptly but will gradually transition to a strategy with higher response intensity. If the first or second strategy model specifically adopts a neural network, regression model, or other strategy prediction model, other models of this component can also be selected in practical applications, and this application does not limit this.
[0129] By employing the aforementioned fusion relationship, the third strategy achieves a smooth transition between basic driving comfort and aggressive responsiveness, maintaining a consistent mathematical mapping between control outputs under different emotional intensities. This results in a more natural strategy switching effect for the vehicle in voice control scenarios. Since the third strategy is continuously generated from the same fusion formula, it reduces jumps in control commands between different driving styles, ensuring that the autonomous driving response matches the intensity of the user's current intent.
[0130] Based on the foregoing embodiments, it further includes:
[0131] If the security check of the third strategy fails, the third strategy will be modified.
[0132] The third strategy is the target control strategy output by the strategy fusion model. Safety verification is a check process to determine whether the third strategy meets safety requirements. Correction is the process of adjusting the parameters of the third strategy that fails the safety verification, modifying it into an executable strategy that meets safety constraints. In this embodiment, the third strategy safety verification can be performed based on preset safety rules. These preset safety rules include at least one or more of the following: current vehicle speed, road speed limit, target vehicle distance, minimum collision time, and lateral lane change conditions. When the acceleration, deceleration, following distance, lane change timing, or steering angle corresponding to the third strategy exceeds the safety boundary, the third strategy is determined to have failed the safety verification, and correction logic is triggered.
[0133] In its implementation, the control unit jointly verifies the longitudinal and lateral control variables in the third strategy and inputs the current vehicle state, environmental perception results, and road constraints into the safety check module. The safety check module can output the verification results using a rule engine, threshold judgment, or constraint optimization. When the verification results indicate that the third strategy poses a risk, the control unit corrects it. Corrections include reducing the target speed, increasing the following distance, delaying lane change requests, smoothing steering angular velocity constraints, or reassigning other control parameters related to autonomous driving. The corrected third strategy is then sent to the vehicle actuators only after meeting safety rules. This correction process can be completed in a closed loop within the onboard controller. In practical applications, other controller models can also be used, and this application does not limit this.
[0134] By making immediate corrections to third strategies that fail safety checks, the fused control intent can retain user emotion-driven characteristics while meeting vehicle operation safety boundaries, and the final output control parameters can be executable and consistent, thus making the autonomous driving control process more stable.
[0135] Based on the foregoing embodiments, the method further includes: generating a first prompt message based on voice information, the first prompt message being used to remind the user to pay attention to safety; and playing the first prompt message.
[0136] In this application, the first prompt information is a safety reminder generated based on the acquired voice information. This information is used to remind the user to pay attention to the surrounding traffic environment, vehicle operating status, and takeover risks during autonomous driving control. The first prompt information can be automatically generated by the in-vehicle voice interaction system in conjunction with the control request corresponding to the voice information. Furthermore, the content can be matched according to the control intent, tone intensity, and vehicle operating status contained in the current voice information, ensuring the reminder content is suitable for the current interaction scenario. In a specific implementation, the first prompt information can be a voice broadcast or a synthesized voice containing preset safety prompt statements. After generation, the system outputs the information through the in-vehicle speaker, allowing the user to directly hear the safety reminder.
[0137] In implementation, after recognizing voice information, the vehicle control unit first parses the corresponding control request and determines whether a safety reminder needs to be output based on the current autonomous driving status. When the voice information is detected to involve autonomous driving control or is related to third-party strategy correction, the control unit calls the prompt generation module to generate a first prompt message and then sends the message to the audio playback module for playback. This audio playback module can be integrated with the vehicle's multimedia system, using a speech synthesis engine to output prompt audio, or directly calling a pre-stored prompt voice file for playback. In practical applications, other models of the playback unit can also be selected, and this application does not limit this choice.
[0138] By simultaneously generating and playing the first prompt message upon receiving voice information, users can be promptly informed of the current autonomous driving control status and safety precautions, and receive clear audio reminders during policy adjustments or corrections, thus making the interaction process more seamless. Because the prompt message is combined with the corresponding control request process for the voice message, users can receive safety notifications without additional user interface operations, thereby improving the effectiveness of safety prompts in in-vehicle interaction and the consistency of system response.
[0139] In one possible implementation, after obtaining the user's voice input, the method further includes: deleting the voice information when it is determined that the user does not have control permissions; generating a second prompt message to inform the user that they do not have control permissions; and playing the second prompt message.
[0140] Control permissions are used to determine whether a user is qualified to control the vehicle's autonomous driving system, and this status serves as the basis for deciding whether to retain voice information subsequently. Voice information consists of the user's input control request, conveying their driving control intentions in voice form. The second prompt message is the content output when permissions are insufficient, clearly informing the user that they currently lack control privileges; playing the second prompt message outputs this message to the user via voice.
[0141] In this embodiment, after receiving voice information collected by the vehicle microphone and obtained by the voice recognition module, the system first calls the permission verification module to verify the current user's identity. Verification can be based on account login status, vehicle key binding relationship, seat occupancy information, or preset permission rules. When the aforementioned verification result indicates that the user does not have control permissions, the control unit immediately deletes the corresponding voice information from the pending processing cache, message queue, or temporary storage area, and stops the voice information from further entering the control intent parsing, emotion scoring, or strategy fusion processing link, so that the voice information does not participate in subsequent vehicle control. Subsequently, the system generates a second prompt message according to a preset prompt template. The second prompt message may include broadcast content such as "The current user does not have control permissions and cannot execute the vehicle control request," and is output by the vehicle speaker. To adapt to the voice interaction architecture of different vehicle models, the prompt message generation module can be configured to communicate with the voice synthesis engine, synthesize the audio into an audio stream, and then play it. In practical applications, this component can also be selected from other models; this application does not limit this.
[0142] By simultaneously deleting voice information and broadcasting a second prompt when permission verification fails, the system can prevent unauthorized input from entering the autonomous driving control link. At the same time, it directly provides feedback on the permission status to the user, ensuring that the permission control logic is consistent with the voice interaction output, thereby improving the standardization and clarity of vehicle control management.
[0143] Figure 4 This is a schematic diagram of the structure of the control device for an autonomous vehicle provided in an embodiment of this application, as shown below. Figure 4 As shown, the control device 40 for the autonomous vehicle provided in this embodiment includes:
[0144] The acquisition module 401 is used to acquire voice information input by the user, which is used to control the autonomous driving of the vehicle.
[0145] The determination module 402 is used to determine the user's emotion score based on voice information. The emotion score is used to represent the degree of agitation of the user's emotions.
[0146] Control module 403 is used to adjust the vehicle's autonomous driving strategy based on emotion scores.
[0147] In one possible embodiment, acoustic features and semantic features are extracted from speech information;
[0148] The acoustic and semantic features are input into a pre-defined emotion scoring model to obtain the user's emotion score.
[0149] In one possible embodiment, the acoustic features include at least one of speech rate features, volume features, and pitch features.
[0150] In one possible embodiment, extracting semantic features from speech information includes:
[0151] Extracting text information from speech information;
[0152] Text information is input into a preset semantic extraction model to obtain semantic features.
[0153] In one possible embodiment, the sentiment scoring model is shown in the following formula:
[0154]
[0155] in, Rate your mood. For activation function, As the first weighting coefficient, Acoustic characteristics, This is the second weighting coefficient. For semantic features, This is a bias term.
[0156] In one possible embodiment, the vehicle's autonomous driving strategy is adjusted based on emotion scores, including:
[0157] Determine the first and second strategies;
[0158] The first and second strategies, along with the emotion score, are input into a pre-defined strategy fusion model to obtain and output the third strategy.
[0159] In one possible embodiment, the first strategy is the output of a first strategy model, which is trained based on the user's historical driving data.
[0160] In one possible embodiment, the first policy model includes a first loss function, which is used to optimize the first policy model for safety and comfort.
[0161] In one possible embodiment, the first loss function is shown in the following formula:
[0162]
[0163] in, The first loss value, For the fitting term, For the policy vector, For data labels, For comfort constraints, For security constraints, , These are the weighting coefficients.
[0164] In one possible embodiment, the second strategy is the output of a second strategy model, which is trained based on driving data from professional drivers.
[0165] In one possible embodiment, the second policy model includes a second loss function used to optimize the efficiency of the second policy model.
[0166] In one possible embodiment, the second loss function is shown in the following formula:
[0167]
[0168] in, This is the second loss value. For the fitting term, For the policy vector, For data labels, For speed, Speed limits for roads, For safety constraints, Minimum collision time, , These are the weighting coefficients.
[0169] In one possible embodiment, the policy fusion model is shown in the following formula:
[0170]
[0171] in, As the third strategy, Rate your mood. As the first strategy, This is the second strategy.
[0172] In one possible embodiment, it also includes:
[0173] If the security check of the third strategy fails, the third strategy will be modified.
[0174] In one possible embodiment, it also includes:
[0175] Based on voice information, a first prompt message is generated, which is used to remind the user to pay attention to safety.
[0176] Play the first prompt message.
[0177] In one possible embodiment, after obtaining the user's voice input, the method further includes:
[0178] Delete the voice message if it is determined that the user does not have control permissions.
[0179] Generate a second prompt message to inform the user that they do not have control permissions;
[0180] Play the second prompt message.
[0181] The control device for the autonomous vehicle provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.
[0182] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 5 As shown, the electronic device 50 provided in this embodiment includes at least one processor 501 and a memory 502. Optionally, the device 50 further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus.
[0183] In a specific implementation, at least one processor 501 executes computer execution instructions stored in memory 502, causing at least one processor 501 to perform the above-described method.
[0184] The specific implementation process of processor 501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.
[0185] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.
[0186] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.
[0187] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.
[0188] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.
[0189] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.
[0190] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.
[0191] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.
[0192] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.
[0193] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0194] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0195] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0196] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.
[0197] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.
Claims
1. A control method for an autonomous vehicle, characterized in that, include: Acquire user-inputted voice information, which is used to control the autonomous driving of the vehicle; Based on the voice information, the user's emotion score is determined, and the emotion score is used to represent the degree of the user's impatience. The autonomous driving strategy of the vehicle is adjusted based on the emotion score.
2. The method according to claim 1, characterized in that, Determining the user's emotion score based on the voice information includes: Extract acoustic and semantic features from the speech information; The acoustic features and semantic features are input into a preset emotion scoring model to obtain the user's emotion score.
3. The method according to claim 2, characterized in that, The acoustic features include at least one of speech rate features, volume features, and pitch features.
4. The method according to claim 2, characterized in that, Extracting semantic features from the speech information includes: Extract text information from the speech information; The text information is input into a preset semantic extraction model to obtain semantic features.
5. The method according to claim 2, characterized in that, The emotion scoring model is shown in the following formula: in, Rate your mood. For activation function, As the first weighting coefficient, Acoustic characteristics, This is the second weighting coefficient. For semantic features, This is a bias term.
6. The method according to claim 1, characterized in that, The step of adjusting the vehicle's autonomous driving strategy based on the emotion score includes: Determine the first and second strategies; The first strategy, the second strategy, and the emotion score are input into a preset strategy fusion model to obtain and output a third strategy.
7. The method according to claim 6, characterized in that, The first strategy is output by the first strategy model, which is trained based on the user's historical driving data.
8. The method according to claim 7, characterized in that, The first strategy model includes a first loss function, which is used to optimize the first strategy model for both safety and comfort.
9. The method according to claim 8, characterized in that, The first loss function is shown in the following formula: in, The first loss value, For the fitting term, For the policy vector, For data labels, For comfort constraints, For security constraints, , These are the weighting coefficients.
10. The method according to claim 6, characterized in that, The second strategy is the output of the second strategy model, which is trained based on driving data from professional drivers.
11. The method according to claim 10, characterized in that, The second strategy model includes a second loss function, which is used to optimize the efficiency of the second strategy model.
12. The method according to claim 11, characterized in that, The second loss function is shown in the following formula: in, This is the second loss value. For the fitting term, For the policy vector, For data labels, For speed, Speed limits for roads, For safety constraints, Minimum collision time, , These are the weighting coefficients.
13. The method according to claim 6, characterized in that, The strategy fusion model is shown in the following formula: in, As the third strategy, Rate your mood. As the first strategy, This is the second strategy.
14. The method according to claim 6, characterized in that, Also includes: If the security verification of the third strategy fails, the third strategy shall be modified.
15. The method according to claim 14, characterized in that, Also includes: Based on the voice information, a first prompt message is generated, which is used to remind the user to pay attention to safety. Play the first prompt message.
16. The method according to claim 14, characterized in that, After obtaining the user's voice input, the method further includes: When it is determined that the user does not have control permissions, the voice information is deleted; A second prompt message is generated, which is used to prompt the user that they do not have control permissions; Play the second prompt message.
17. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-16.
18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-16.
