Driving assistance system based on artificial intelligence

Abstract

The invention relates to the technical field of automatic driving, and discloses a driving assistance system based on artificial intelligence, which comprises a mental inference engine used for constructing a mental model for quantifying the inherent driving decision preference of a target traffic participant for the target traffic participant, reasoning the probability distribution of the high-order intention of the target traffic participant based on the model, and further calculating the uncertainty metric value of the intention; and the decision planning module is used for adaptively switching between two modes based on a comparison result of the intention uncertainty measurement value and a preset threshold value. If the uncertainty exceeds a threshold value, entering an active cognitive detection mode, and generating a detection behavior aiming at reducing the uncertainty; and if not, entering a socialized game planning mode, and generating a socialized driving strategy giving consideration to the interests of the driver and other parties. According to the method, the accuracy of human behavior prediction and the personification level of interaction are improved, and the decision robustness and safety in an uncertain scene are enhanced.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to a driving assistance system based on artificial intelligence. Background Technology

[0002] With the continuous development of artificial intelligence and sensor technology, autonomous driving systems have significantly improved their capabilities in environmental perception and path planning. However, existing technologies still have inherent limitations when dealing with complex interaction scenarios with human drivers. Current decision-making and planning methods typically simplify other traffic participants into dynamic obstacles that follow predetermined physical motion models (such as constant speed and constant acceleration models). This approach ignores the inherent decision-making intentions of human drivers as intelligent agents, their individual driving styles, and the game-theoretic psychology involved in the interaction process. Therefore, in traffic scenarios requiring negotiation or cooperation (e.g., ramp merging, unsignalized roundabout passage), the behavior of autonomous vehicles often appears too conservative or mechanical, making it difficult to achieve smooth and efficient interaction and coordination with human drivers, thus affecting the efficiency of overall traffic flow and the naturalness of driving.

[0003] Furthermore, when faced with ambiguous human driver behavior and unclear intentions, the probability distribution of intentions output by traditional prediction models exhibits high uncertainty. Under this high level of cognitive uncertainty, existing technologies lack effective mechanisms to proactively eliminate or reduce this uncertainty. Systems typically can only passively adopt avoidance strategies that maximize safety margins, such as significant deceleration or prolonged stopping. This singular approach, in some cases, not only reduces traffic efficiency but may also cause unnecessary disturbances to the surrounding traffic environment due to atypical driving behavior. Therefore, enabling autonomous driving systems to infer the intentions of others and make proactive, human-like interactive decisions based on this is a pressing technical problem that needs to be solved in this field. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based driving assistance system that solves the problems of existing autonomous driving systems being conservative in behavior and having low interaction efficiency due to their lack of effective inference of the human driver's internal decision-making intentions and the lack of proactive interaction clarification mechanisms when faced with uncertainty of intentions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an artificial intelligence-based driving assistance system, comprising:

[0006] The environmental perception module is configured to acquire the physical state sequence of at least one target traffic participant in the scene in real time;

[0007] The mental inference engine, whose input is connected to the output of the environment perception module, is configured to receive the physical state sequence and perform the following operations:

[0008] (a) Construct a quantitative mental model for the target traffic participant, which is used to characterize the participant's intrinsic driving decision preferences;

[0009] (b) Combining the mental model with the current physical state, infer a probability distribution of a set of higher-order intentions of the participant, and calculate an intention uncertainty metric based on the probability distribution;

[0010] The decision planning module, whose input is connected to the output of the mental inference engine, is configured to receive the mental model, the probability distribution of the higher-order intention, and the intention uncertainty metric, and adaptively switch between two operating modes based on a comparison of the intention uncertainty metric with a preset threshold.

[0011] (i) Proactive cognitive detection mode: activated when the metric exceeds the threshold, used to generate and execute a detection behavior with the primary goal of reducing the uncertainty of the intent;

[0012] (ii) Socialized game planning mode: activated when the metric does not exceed the threshold, used to generate and execute a socialized driving strategy that takes into account the interests of itself and others.

[0013] Preferably, when constructing the mental model, the mental inference engine employs a reverse reinforcement learning algorithm to model the intrinsic reward function of the target traffic participant as a weighted linear combination of a set of preset driving behavior features, and solves the weight vector of the combination by analyzing the received physical state sequence, and uses the weight vector as the quantized mental model.

[0014] Preferably, the combination of preset driving behavior characteristics includes at least: efficiency characteristics for measuring travel speed and time, safety characteristics for measuring headway and lateral distance, and comfort characteristics for measuring longitudinal and lateral acceleration.

[0015] Preferably, when reasoning about the higher-order intention, the mental inference engine uses a Bayesian belief network to calculate the posterior probability distribution of the higher-order intention, taking the mental model, the current physical state, and the current traffic environment context as input evidence.

[0016] Preferably, when calculating the intent uncertainty metric, the mental inference engine uses the Shannon entropy formula to calculate the posterior probability distribution.

[0017] Preferably, in the active cognitive detection mode, the decision planning module is further configured as follows:

[0018] Generate a set of candidate detection behaviors that conform to social norms;

[0019] Each candidate detection behavior is internally simulated to assess its expected information gain, and finally the detection behavior with the greatest expected information gain is selected and executed.

[0020] Preferably, the detection behavior is a driving action with minimal physical impact but perceptible to a human driver, selected from the group consisting of: slight lateral deviation within the current lane, a brief and gentle speed pulse, and a momentary single or multiple flashes of the turn signal.

[0021] Preferably, in the social game planning model, the decision planning module is further configured as follows:

[0022] The current driving scenario is modeled as a multi-agent game model involving the vehicle itself and the target traffic participants;

[0023] The game model is solved based on a total social utility function, which is a weighted sum of the vehicle's own utility and the utility of the target traffic participant predicted using the mental model.

[0024] Preferably, in the total social utility function, the weight of the target traffic participant's utility is determined by an online adjustable social factor parameter, which is used to dynamically balance egoism and altruism in driving decisions.

[0025] Preferably, the system further includes a human-computer interaction interface, which is connected to the mental inference engine and the decision planning module, and is configured as follows:

[0026] Present the current high-order intention inference results regarding the target traffic participant to the driver in a visual manner;

[0027] When the system is in the active cognitive detection mode, it informs the driver that the system is performing a detection action and its purpose.

[0028] This invention provides a driver assistance system based on artificial intelligence. It has the following beneficial effects:

[0029] 1. This invention employs a reverse reinforcement learning algorithm to construct quantified mental models of other traffic participants, no longer treating them as simple physical moving points, but as intelligent agents capable of quantifying their inherent driving preferences. This understanding at the "mental" level enables the system to go beyond pure trajectory extrapolation when predicting the behavior of other vehicles, making predictions that are more in line with human driving logic. As a result, in complex interactive scenarios such as merging and yielding, it generates more natural, smoother, and more easily understood anthropomorphic driving strategies for human drivers.

[0030] 2. This invention utilizes a quantification mechanism for intent uncertainty and innovatively designs an active cognitive detection mode. When the system's judgment of a key target's intent is highly uncertain, it will not make decisions based on luck or excessive conservatism as in traditional methods. Instead, it will proactively execute subtle probing actions aimed at acquiring information, such as slight changes in vehicle posture. This proactive probing effectively elicits a clear response from the target, thereby quickly and proactively eliminating cognitive ambiguity and providing a reliable basis for subsequent safety decisions. This significantly improves the system's safety and adaptability in complex and ambiguous environments.

[0031] 3. This invention employs a social game theory planning model in its decision-making process. Its core is based on a total social utility function that simultaneously considers the utility of the system and that of others. By introducing adjustable social factors, the system can make cooperative and altruistic decisions while ensuring its own safety and efficiency, while also appropriately considering the interests of other traffic participants. This "empathetic" driving approach avoids traffic conflicts and efficiency bottlenecks that may result from purely self-serving decisions, contributing to a more harmonious and orderly traffic environment and thus possessing the potential to improve local traffic flow. Attached Figure Description

[0032] Figure 1 This is a system flowchart of the present invention;

[0033] Figure 2 This is a schematic diagram of the method of the present invention. Specific implementation methods

[0035] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Please see the appendix Figure 1 -Appendix Figure 2 This invention provides an artificial intelligence-based driving assistance system, comprising:

[0037] The environmental perception module is used to accurately and in real time perceive the dynamic traffic environment around the vehicle and convert it into structured data that can be processed by subsequent modules.

[0038] The environmental perception module is based on a multimodal sensor array, which is rigidly mounted on the vehicle body. This sensor array specifically includes:

[0039] A forward-looking high dynamic range camera is used to capture visual image data of the road ahead of the vehicle;

[0040] Forward-facing solid-state LiDAR is used to acquire high-density three-dimensional spatial point cloud data within a fan-shaped field of view in front of the vehicle;

[0041] In addition, corner radars and a forward-facing long-range millimeter-wave radar are deployed at the four corners of the vehicle to detect and measure the radial distance, relative speed and azimuth of the target.

[0042] During system startup or offline calibration, the environmental perception module executes a sensor extrinsic parameter calibration process. This process accurately determines the pose parameters of all sensors in a unified vehicle coordinate system (e.g., with the rear axle center as the origin), namely the rotation matrix and translation vector. Simultaneously, the module's internal clock synchronization mechanism (e.g., based on the Precision Time Protocol, PTP) ensures that data frames acquired from each sensor carry a unified, high-precision timestamp, providing a foundation for subsequent alignment of multi-sensor data in the time dimension.

[0043] The environmental perception module receives raw data streams from various sensors and first performs independent single-modal perception processing. Specifically: for visual image data acquired by the camera, a convolutional neural network (CNN) model is used to perform object detection, outputting the 2D bounding boxes of objects in the image, object categories (e.g., vehicles, pedestrians), and confidence scores. For 3D point cloud data acquired by LiDAR, through a series of processes such as ground point segmentation, point cloud clustering (e.g., the density-based spatial clustering algorithm DBSCAN), and 3D bounding box fitting, preliminary estimates of the 3D position, size, and heading angle of each independent object in the environment are output. For the target list acquired by millimeter-wave radar, each target point directly provides accurate range and radial velocity information.

[0044] After completing single-modal perception processing, the environmental perception module executes a multi-sensor fusion process within a shared Cartesian coordinate system (i.e., a unified vehicle coordinate system). At the heart of this process is a multi-target tracker that maintains a dynamic list of targets, each representing a continuously tracked traffic participant. Within each computation cycle, the outputs from the aforementioned single-modal perception processing (2D detection boxes, 3D detection boxes, and radar target points) are fed into a data association unit. This unit uses an allocation algorithm (e.g., the Hungarian algorithm or Joint Probabilistic Data Association (JPDA)) to calculate the association between the current sensor detection results and the target tracks maintained by the tracker.

[0045] For each successfully associated existing track, an Extended Kalman Filter (EKF) is used to update its state. The filter first performs a prediction step, using a motion model (e.g., a constant turn rate and velocity CTRV model) to predict the target's physical state from the previous time step to the current time step. Then, the filter performs an update step, using measurements from one or more sensors associated with the target at the current time step to refine the predicted state, thereby obtaining a posterior-optimal estimate of the target's current physical state. If a detection cannot be associated with any existing track, that detection is used to initialize a new track.

[0046] Through the multi-target tracking process described above, the final output of the environment perception module is to provide a physical state vector s describing the kinematic information of each traffic participant (denoted as agent i) that is being stably tracked in the scene at any discrete time t. i (t). The mathematical expression for this state vector is:

[0047] s i (t)=[p x,i (t),p y,i (t),v x,i (t),v y,i (t),a x,i (t),a y,i (t),h i (t)];

[0048] In the aforementioned state vector s i In the expression of (t), the symbols are defined as follows:

[0049] p x,i (t) and p y,i (t) represents the estimated lateral and longitudinal positions of agent i at time t in the vehicle's unified coordinate system, respectively.

[0050] v x,i (t) and vy,i (t) represents the estimated values ​​of the lateral and longitudinal velocities of agent t in the coordinate system at time t, respectively.

[0051] a x,i (t) and a y,i (t) represents the estimated values ​​of the lateral acceleration and longitudinal acceleration of agent i at time t in this coordinate system. These values ​​can be obtained by differentiating the velocity components or by directly estimating them in the filter state.

[0052] h i (t) represents the heading angle of agent i, which is the estimated value of the angle between its heading and the longitudinal axis of the vehicle coordinate system.

[0053] In each computation cycle, the environmental perception module generates and updates the corresponding physical state vectors for all identified and tracked target traffic participants in the scene. These physical state sequences, consisting of a series of historical state vectors, are transmitted as a structured data stream from the output of the module to the input of the mental inference engine, providing the necessary quantified observation data for subsequent mental modeling and intent reasoning steps.

[0054] The input of the mental inference engine is connected to the output of the aforementioned environmental perception module, and is used to receive the sequence of physical states of the target traffic participants output by that module. The output of this engine is then connected to the input of the subsequent decision-making and planning module. Its function is to establish a quantitative internal decision preference model for the target traffic participant, and based on this model, to perform probabilistic inference on the participant's current intention, ultimately outputting its mental model, the probability distribution of higher-order intentions, and a quantitative measure of intention uncertainty.

[0055] The internal functionality of the mental inference engine can be implemented by two sequentially operating units: a mental modeling unit and an intention reasoning unit. First, the mental modeling unit processes the input sequence of historical physical states to generate a mental model, and then the intention reasoning unit uses this mental model and the latest physical state to perform inference.

[0056] The function of the mental modeling unit is to construct a mental model that quantifies the intrinsic driving style of the target traffic participant i. This unit models the participant's driving decision-making process as a process of seeking to maximize its intrinsic reward function R. i The mathematical expression for a weighted linear combination of a predefined driving behavior feature vector, parameterized as such, is:

[0057]

[0058] In the aforementioned expression for the reward function, the symbols are defined as follows:

[0059] s represents the system state at any given time, which includes the physical states of the vehicle and the target traffic participant i.

[0060] 'a' represents the driving behavior of the target traffic participant i in state 's', such as a specific combination of longitudinal acceleration and lateral angular velocity.

[0061] φ(s,a) is a feature function vector used to map the state-behavior pair (s,a) to a multidimensional feature space.

[0062] The specific composition of this vector may include:

[0063] Efficiency feature φ used to characterize traffic efficiency eff Its value is positively correlated with the speed along the centerline of the road;

[0064] Safety feature φ used to characterize safety margin safe Its value is positively correlated with the headway of the vehicle or the lateral distance to the lane line;

[0065] and comfort features φ used to characterize ride smoothness comf Its value is positively correlated with the reciprocal of the absolute value of acceleration or jerk.

[0066] w i It is a weight vector, where each element corresponds to a weight of a feature in the feature vector φ(s,a). This weight vector w i This is defined as a quantitative mental model of the target traffic participant i.

[0067] To solve for the unknown weight vector w i The mental modeling unit employs an inverse reinforcement learning algorithm. Specifically, this unit uses the Maximum Entropy Inverse Reinforcement Learning (Max-EntIRL) framework. The goal of this framework is to find the reward function with the largest entropy, i.e., the least biased reward function, among all reward functions that can explain the observed driving trajectory. Upon receiving a driving trajectory τ containing multiple time-state-behavior pairs from the environmental perception module... i Then, the unit determines the weight vector w by solving an optimization problem. i This ensures that, under the reward function defined by the weight, the observed trajectory τ i The probability of occurrence is maximized. The mathematical expression for this probability is:

[0068]

[0069] Among them, Z(w i) is the partition function used for normalization. This optimization problem can be solved using gradient-based numerical optimization methods. The resulting weight vector w i That is, it is output as the mental model.

[0070] The input of the intent reasoning unit is connected to the output of the mental modeling unit and the output of the environment perception module. Its function is to apply this information to the already constructed mental model. i Based on this, and combined with the current physical state and traffic environment context, the higher-order intentions of the target traffic participant i are inferred. This unit constructs this reasoning process as a Bayesian belief network (BBN).

[0071] The Bayesian belief network is calculated over a series of evidence E (including current driving behavior a) i (t), traffic scenario C, and mental model w i Given the conditions, each predefined higher-order intention I i,j (For example, j=1 represents "keep in lane", j=2 represents "change lanes to the left") The posterior probability P(I) is the probability of keeping in lane. i,j The calculation of |E) follows Bayes' theorem.

[0072] P(I i,j |a i (t),C,w i )=η·P(a i (t)|I i,j ,C,w i )·P(I i,j |C,w i );

[0073] Here, η is a normalization constant used to ensure that the sum of the probabilities of all possible intentions is 1.

[0074] P(a i (t)|…) is the likelihood, P(I…) is the likelihood, and P(I…) is the likelihood. i,j |…) represents the prior probability. The output of this unit is a complete belief distribution Bel(I) containing all possible intentions and their corresponding probabilities. i ).

[0075] In the belief distribution of obtaining intention, Bel(I) i Following this, the intent reasoning unit is further configured to calculate the uncertainty of the belief distribution. This unit uses the Shannon entropy formula to quantify this uncertainty, obtaining a scalar value H(Bel(I)). i This value is defined as a measure of intent uncertainty. Its mathematical expression is:

[0076] H(Bel(I i ))=-∑ jP(I i,j )·log2P(I i,j );

[0077] Wherein, P(I i,j ) is intention I i,j The posterior probability. The higher the entropy value, the more dispersed the intention distribution and the higher the uncertainty.

[0078] Finally, the mental inference engine will generate a mental model w i Bel(I) distribution of higher-order intentions i ) and the measure of intent uncertainty H(Bel(I) i The data is transmitted from its output end to the input end of the decision planning module, providing a basis for subsequent decision mode selection and behavior planning.

[0079] In one specific embodiment, the decision planning module is the core actuator for implementing the adaptive driving behavior described in this invention. The input of this module is connected to the output of the aforementioned mental inference engine and is configured to receive the mental model w. i The belief distribution of the higher-order intentions, Bel(I) i ) and the intent uncertainty metric H(Bel(I) i The output of this module is connected to a vehicle control actuator to transmit the final generated driving decision commands.

[0080] The function of the decision-making and planning module is to switch between two mutually exclusive operating modes based on the received intent uncertainty metric: an active cognitive detection mode and a social game planning mode. This switching is deterministic, based on the intent uncertainty metric H(Bel(I)). i )) with an internally stored, system-calibrated preset uncertainty threshold ∈ H The comparison results between them.

[0081] Specifically, at the beginning of each decision cycle, the decision planning module performs a comparison operation. If the comparison result is H(Bel(I)... i ))>∈ H Then the module enters active cognitive detection mode. If the comparison result is H(Bel(I) i ))≤∈ H Then the module enters the social game planning mode. These two modes employ different optimization methods.

[0082] Logic for generating goals and behaviors.

[0083] When entering active cognitive detection mode, the module aims to generate and execute a driving action that minimizes intent uncertainty with maximum efficiency. To this end, the module performs the following steps: First, it generates a discrete action set A consisting of candidate detection behaviors that comply with traffic regulations and social norms and have minimal physical disturbances. p .

[0084] Behaviors in this set include, for example, a small lateral drift within the current lane at a specific angular velocity, a longitudinal velocity pulse with a specific acceleration and duration, or a momentary flashing of a turn signal.

[0085] Next, for set A p Each candidate detection behavior a in p This module evaluates the expected information gain (EIG) of the behavior through internal forward simulation. The expected information gain is defined as the result of performing action a. p Then, the expected reduction in uncertainty regarding intentions. Its calculation formula is:

[0086] EIG(a p )=H(Bel(I i ))-∑ o∈O P(o|a p )H(Bel′(I i )|o,a p );

[0087] In the aforementioned expression for expected information gain, the symbols are defined as follows:

[0088] H(Bel(I i )) is a measure of the uncertainty of the current intention before the action is performed.

[0089] O is the action to be performed (a) p Then, the set of all possible reactions of the target traffic participants.

[0090] P(o|a p ) is the action to be performed. p Subsequently, the probability of the target traffic participant reacting o was observed, and this probability was obtained using the mental model w. i Obtained through forward prediction.

[0091] H(Bel′(I i )|o,a p ) is in the execution of action a p And after observing response o, the new belief distribution Bel′(I) for higher-order intentions was observed. iThe posterior entropy is calculated using EIG(a). After calculating the expected information gain of all candidate detection behaviors, this module selects the one that makes EIG(a) the most suitable. p The largest behavior This serves as the final decision for the detection behavior and is then output.

[0092] When entering the social game planning mode, the goal of this module is to generate a driving strategy that can effectively coordinate with surrounding traffic participants while ensuring safety and efficiency. To this end, the module models the current driving interaction scenario as a multi-agent game model. In this model, the vehicle and the target traffic participant i are the players in the game. The module finds the optimal driving behavior of the vehicle by solving a total social utility function U. social The expression is:

[0093] U social (s,a ego ,a i )=U ego (s,a ego ,a i )+λ·U i (s,a ego ,a i );

[0094] In the aforementioned expression for the total social utility function, the symbols are defined as follows:

[0095] s is the current joint state; a ego This is a candidate behavior for this vehicle; a i It is a predicted behavior of the target traffic participant i.

[0096] U ego This vehicle was involved in joint action (a ego ,a i The utility value is defined by the vehicle's own driving objectives (such as safety, efficiency, and comfort).

[0097] U i Is the target traffic participant i in the joint behavior (a ego ,a i The utility value is calculated using the weight vector w passed from the mental inference engine as the weight vector of the mental model. i The defined reward function R i .

[0098] λ is an online configurable, dimensionless social factor with a value in the range [0,1], used to adjust the weight of the consideration of the utility of the target traffic participants in the vehicle's decision-making.

[0099] The module uses a numerical optimization or search algorithm (e.g., Monte Carlo tree search) to find the function U that maximizes the total social utility. social This vehicle's behavior This optimal behavior This is defined as a socialized driving strategy and is used as the final decision output by this module.

[0100] Whether it is the detection behavior output by the active cognitive detection mode or the socialized driving strategy output by the socialized game planning mode, it ultimately manifests as a specific, time-sensitive trajectory or behavioral instruction, which is transmitted from the output of the decision planning module to the vehicle control actuator to generate underlying control commands for the vehicle.

[0101] U social (s,a ego ,a i )=U ego (s,a ego ,a i )+λ·U i (s,a ego ,a i );

[0102] In the aforementioned expression for the total social utility function, the symbols are defined as follows:

[0103] s is the current joint state; a ego This is a candidate behavior for this vehicle; a i It is a predicted behavior of the target traffic participant i.

[0104] U ego This vehicle was involved in joint action (a ego ,a i The utility value is defined by the vehicle's own driving objectives (such as safety, efficiency, and comfort).

[0105] U i Is the target traffic participant i in the joint behavior (a ego ,a i The utility value is calculated using the weight vector w passed from the mental inference engine as the weight vector of the mental model. i The defined reward function R i .

[0106] λ is an online configurable, dimensionless social factor with a value in the range [0,1], used to adjust the weight of the consideration of the utility of the target traffic participants in the vehicle's decision-making.

[0107] The module uses a numerical optimization or search algorithm (e.g., Monte Carlo tree search) to find the function U that maximizes the total social utility. social This vehicle's behavior This optimal behavior This is defined as a socialized driving strategy and is used as the final decision output by this module.

[0108] Whether it is the detection behavior output by the active cognitive detection mode or the socialized driving strategy output by the socialized game planning mode, it ultimately manifests as a specific, time-sensitive trajectory or behavioral instruction, which is transmitted from the output of the decision planning module to the vehicle control actuator to generate underlying control commands for the vehicle.

[0109] The vehicle control actuator is an interface module connecting the upper-level planning and the vehicle's physical actuators. Its input is connected to the output of the aforementioned decision-making and planning module, receiving detection behavior commands or socialized driving strategy commands from that module. The actuator's output communicates directly with the vehicle's lower-level control units, such as the Electronic Power Steering (EPS), Electronic Stability Control (ESC), and Powertrain Control Unit (PCU), via an onboard bus (e.g., Controller Area Network (CAN)).

[0110] The function of the vehicle control actuator is to transform the relatively macroscopic behavioral instructions received from the decision-making and planning module into a physically feasible, temporally continuous driving trajectory for the vehicle, and further generate low-level control commands that can accurately track this trajectory. This function is accomplished collaboratively by a trajectory generation unit and a trajectory tracking controller.

[0111] Upon receiving a behavioral instruction, the trajectory generation unit first operates. This unit parses the received instruction (e.g., a target state point, a target velocity, or a small lateral offset) into one or more path constraints. Subsequently, within a finite time domain, the unit generates a time-parameterized, smooth driving trajectory that satisfies vehicle kinematic and dynamic constraints by solving a constrained optimization problem. The trajectory It consists of a dense series of trajectory points, each defining the vehicle's expected pose, velocity, and acceleration at a future moment. Specifically, this trajectory generation unit minimizes a cost function. To obtain the optimal trajectory. The mathematical expression of this cost function is:

[0112]

[0113] In the aforementioned expression for the cost function, the symbols are defined as follows:

[0114] Tf It is the time domain endpoint of the trajectory planning.

[0115] jerk(t) is the jerk of the trajectory at time t (i.e., the rate of change of acceleration), and minimizing its squared term aims to ensure the smoothness of the trajectory.

[0116] dev(t) is the lateral and directional deviation of the trajectory at time t relative to the road centerline or reference path. Minimizing its squared term aims to maintain lane centering.

[0117] coll(t) is the collision risk cost between the trajectory and surrounding obstacles at time t. This value increases as the distance between the predicted position of the vehicle and the predicted position of the obstacle decreases.

[0118] c1, c2, and c3 are preset normal weights used to balance various performance indicators.

[0119] Upon receiving a behavioral instruction, the trajectory generation unit first operates. This unit parses the received instruction (e.g., a target state point, a target velocity, or a small lateral offset) into one or more path constraints. Subsequently, within a finite time domain, the unit generates a time-parameterized, smooth driving trajectory that satisfies vehicle kinematic and dynamic constraints by solving a constrained optimization problem. The trajectory It consists of a dense series of trajectory points, each defining the vehicle's expected pose, velocity, and acceleration at a future moment. Specifically, this trajectory generation unit minimizes a cost function. To obtain the optimal trajectory. The mathematical expression of this cost function is:

[0120]

[0121] In the aforementioned expression for the cost function, the symbols are defined as follows:

[0122] T f It is the time domain endpoint of the trajectory planning.

[0123] jerk(t) is the jerk of the trajectory at time t (i.e., the rate of change of acceleration), and minimizing its squared term aims to ensure the smoothness of the trajectory.

[0124] dev(t) is the lateral and directional deviation of the trajectory at time t relative to the road centerline or reference path. Minimizing its squared term aims to maintain lane centering.

[0125] coll(t) is the collision risk cost between the trajectory and surrounding obstacles at time t. This value increases as the distance between the predicted position of the vehicle and the predicted position of the obstacle decreases.

[0126] c1, c2, and c3 are preset normal weights used to balance various performance indicators.

[0127] In generating the optimal trajectory Then, the trajectory tracking controller begins operation. This controller employs a Model Predictive Control (MPC) algorithm. In each control cycle k, the controller acquires the vehicle's current actual state (obtained through onboard sensors or a state estimator) and uses the aforementioned generated optimal trajectory. This serves as a reference trajectory for the next N time steps. The model predictive controller calculates the optimal control input sequence {u} for the next N steps by solving an optimization problem online. k ,u k+1 ,…,u k+N-1 The optimization problem aims to minimize the difference between the predicted vehicle trajectory and the reference trajectory. The error between the two conditions must simultaneously satisfy the vehicle's dynamic constraints and the physical limitations of the actuators (such as maximum steering angle, maximum acceleration / deceleration, etc.). This control input vector u k The expression is:

[0128] u k =[δ k ,a k ];

[0129] Where, δ k It is the expected front wheel steering angle during control period k, while a k It is the desired longitudinal acceleration.

[0130] After obtaining the optimal control input sequence, the controller only applies the first control input u in the sequence. k Applied to vehicles. This control input u k This is converted into specific electronic control commands, such as the desired front wheel steering angle δ. k The desired longitudinal acceleration a is sent to the EPS controller. k This is converted into a request for throttle opening or brake pressure and sent to the PCU or ESC. In the next control cycle k+1, the process of acquiring the state, solving the optimization problem, and applying control will be repeated to achieve rolling time-domain optimization tracking of the planned trajectory.

[0131] Working Principle: This method first performs an environmental perception step, acquiring and processing data through an onboard multimodal sensor array to generate a series of time-ordered physical states for the target traffic participants in the scene. Based on this sequence of physical states, the method then executes a complete mental inference process. This process first analyzes the historical behavior of the participants using an inverse reinforcement learning algorithm to construct a mental model that quantifies their inherent driving decision preferences. Next, using this mental model and the current state, the method calculates the belief distribution of all higher-order intentions of the target through a probabilistic inference model, and finally uses information theory principles to calculate a quantified intention uncertainty metric from this distribution. Subsequently, the method compares this uncertainty metric with a preset threshold and determines the subsequent planning mode based on the comparison result: if the uncertainty metric exceeds the threshold, active detection planning is executed, aiming to maximize the expected information gain and generating a detection behavior instruction with minimal physical perturbation to reduce uncertainty; if it does not exceed the threshold, social game planning is executed, generating a socialized driving strategy instruction that takes into account the interests of all parties by solving a social utility function that integrates the utility of the participant and the utility of others. Ultimately, regardless of whether the generated commands are for probing behavior or socialized driving strategies, this method transforms them into a smooth driving trajectory through a trajectory generation and control execution step. It then uses model predictive control to generate and apply precise low-level vehicle control commands. This complete process, as a decision-making and control cycle, is executed cyclically at a preset frequency, thereby achieving continuous response and adaptive control to the driving environment.

[0132] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A driving assistance system based on artificial intelligence, characterized in that, include: The environmental perception module is configured to acquire the physical state sequence of at least one target traffic participant in the scene in real time; The mental inference engine, whose input is connected to the output of the environment perception module, is configured to receive the physical state sequence and perform the following operations: (a) Construct a quantitative mental model for the target traffic participant, which is used to characterize the participant's intrinsic driving decision preferences; (b) Combining the mental model with the current physical state, infer a probability distribution of a set of higher-order intentions of the participant, and calculate an intention uncertainty metric based on the probability distribution; The decision planning module, whose input is connected to the output of the mental inference engine, is configured to receive the mental model, the probability distribution of the higher-order intention, and the intention uncertainty metric, and adaptively switch between two operating modes based on a comparison of the intention uncertainty metric with a preset threshold. (i) Proactive cognitive detection mode: activated when the metric exceeds the threshold, used to generate and execute a detection behavior with the primary goal of reducing the uncertainty of the intent; (ii) Socialized game planning mode: activated when the metric does not exceed the threshold, used to generate and execute a socialized driving strategy that takes into account the interests of itself and others.

2. The driving assistance system based on artificial intelligence according to claim 1, characterized in that, When constructing the mental model, the mental inference engine uses an inverse reinforcement learning algorithm to model the intrinsic reward function of the target traffic participant as a weighted linear combination of a set of preset driving behavior features. By analyzing the received physical state sequence, it reverse-engineers the weight vector of the combination and uses the weight vector as the quantized mental model.

3. The driving assistance system based on artificial intelligence according to claim 2, characterized in that, The combination of preset driving behavior characteristics includes at least: efficiency characteristics for measuring traffic speed and time, safety characteristics for measuring headway and lateral distance, and comfort characteristics for measuring longitudinal and lateral acceleration.

4. The driving assistance system based on artificial intelligence according to claim 1, characterized in that, When reasoning about the higher-order intention, the mental inference engine uses a Bayesian belief network to calculate the posterior probability distribution of the higher-order intention, taking the mental model, the current physical state, and the current traffic environment context as input evidence.

5. A driving assistance system based on artificial intelligence according to claim 4, characterized in that, When calculating the intent uncertainty metric, the mental inference engine uses the Shannon entropy formula to calculate the posterior probability distribution.

6. The artificial intelligence-based driving assistance system according to claim 1, characterized in that, In the active cognitive detection mode, the decision planning module is further configured as follows: Generate a set of candidate detection behaviors that conform to social norms; Each candidate detection behavior is internally simulated to assess its expected information gain, and finally the detection behavior with the greatest expected information gain is selected and executed.

7. A driving assistance system based on artificial intelligence according to claim 6, characterized in that, The detection behavior is a driving action with minimal physical impact but perceptible to a human driver, selected from the group consisting of: slight lateral deviation within the current lane, a brief and gentle speed pulse, and a momentary single or multiple flashes of the turn signal.

8. A driving assistance system based on artificial intelligence according to claim 1, characterized in that, In the aforementioned social game planning model, the decision planning module is further configured as follows: The current driving scenario is modeled as a multi-agent game model involving the vehicle itself and the target traffic participants; The game model is solved based on a total social utility function, which is a weighted sum of the vehicle's own utility and the utility of the target traffic participant predicted using the mental model.

9. A driving assistance system based on artificial intelligence according to claim 1, characterized in that, In the total social utility function, the weight of the target traffic participant's utility is determined by an online adjustable social factor parameter, which is used to dynamically balance egoism and altruism in driving decisions.

10. A driving assistance system based on artificial intelligence according to claim 1, characterized in that, The system also includes a human-computer interaction interface, which is connected to the mental inference engine and the decision planning module, and is configured as follows: Present the current high-order intention inference results regarding the target traffic participant to the driver in a visual manner; When the system is in the active cognitive detection mode, it informs the driver that the system is performing a detection action and its purpose.

Images