Perceptual phantom-oriented automatic driving vehicle local path planning method

Abstract

The invention discloses an automatic driving vehicle local path planning method facing perception phantom. The method comprises the following steps: (1) describing local path planning as an MPC problem to establish a mathematical model; (2) establishing a local path planning constraint set; (3) establishing a cost function considering the existence probability of the perceptual phantom; (4) compensating the uncertainty of the MPC model in the local path planning process by using sparse Gaussian process regression; and (5) carrying out optimization solution, and outputting an optimal automatic driving vehicle path planning result. The method has the technical advantages that a safety error correction opportunity is provided for the vehicle by establishing constraints of the vehicle braking distance and the effective sensing distance; establishing a cost function considering the existence probability of the perception phantom vehicle, and enabling the vehicle to realize balanced planning; and the uncertainty of the MPC model is compensated by using sparse Gaussian process regression, so that the safety of local path planning at the phantom vehicle is improved.

Description

Technical Field

[0001] This invention relates to the field of local path planning for autonomous vehicles, and in particular to a local path planning method when a phantom is perceived. Background Technology

[0002] Autonomous vehicles, also known as driverless vehicles, are a revolutionary transportation solution that integrates artificial intelligence, high-precision sensing, the Internet of Things, and advanced computing technologies. They aim to partially or completely replace human drivers through intelligent systems, achieving safe, efficient, and convenient autonomous travel. Autonomous driving technology is the brain and soul of an autonomous vehicle; it is a complex set of algorithms, software, and computing systems. This "artificial intelligence" system endows autonomous vehicles with the ability to understand the world, think, judge, and act. Autonomous driving technology can be functionally divided into an environmental perception module, a high-level planning module, and a tracking and control module. The high-level planning module includes a behavior decision layer and a motion planning layer. The realization of autonomous driving technology relies on the high degree of coordination between these modules, a complex systems engineering project involving interconnected and precisely coordinated components. First, the perception module, by fusing data from multiple sensors such as cameras, LiDAR, and millimeter-wave radar, identifies key information around the vehicle, such as obstacles, lane lines, and traffic signs, in real time and with high accuracy. The high-level planning module searches for an optimal or near-optimal path as the global path based on the known environment. When the environment changes, the high-level planning module comprehensively analyzes real-time obstacle perception information, high-precision maps, and positioning to assess the feasibility of the global path and make behavioral decisions (such as following, overtaking, and stopping) to plan a safe, smooth, and traffic-compliant local path. Finally, the tracking control module precisely executes the planned trajectory commands through drive-by-wire, steering, and braking technologies to control the vehicle's speed and direction.

[0003] Local path planning is an indispensable core technology in the field of autonomous driving research. Its core idea is that the high-level planning module, based on the globally planned path and real-time perception data, avoids obstacles and generates a safe and smooth driving trajectory. Therefore, the accuracy and safety of local path planning heavily rely on the real-time environmental information provided by the perception module. However, in actual driving, problems such as noise, occlusion, and delay in sensor information can lead to perception uncertainty (reduced perception accuracy of the perception module), making it impossible to accurately determine whether the identified obstacle vehicle actually exists, i.e., the perception module experiences perception phantoms. When the high-level planning module receives a perception phantom containing phantom uncertainty information (the probability of phantom existence) from the perception module, its obstacle avoidance algorithm design goes far beyond simply "avoiding" the phantom obstacle vehicle. It requires a flexible probabilistic decision-making and planning framework based on the probability of phantom existence. The core of this framework is to abandon the traditional method of planning based on whether the perception phantom is completely present or not. Instead, it generates a trajectory with controllable collision risk, balancing safety and efficiency, and predictable vehicle state (position, yaw angle) through evaluation and trade-offs.

[0004] Model predictive control (MPC) optimizes vehicle control variables such as steering angle and speed in real time through rolling optimization, taking into full account various variable constraints. It then predicts the vehicle state based on the optimal control variables and the vehicle dynamics model, thus achieving local path planning. Therefore, MPC is widely used in local path planning for autonomous vehicles. When using MPC for local path planning, an internal mathematical model needs to be built based on the vehicle dynamics model. The accuracy of this mathematical model significantly affects the path planning effect. However, unavoidable external disturbances such as changes in road slope and adhesion, as well as unmodeled dynamic characteristics such as nonlinear tire forces during high-speed cornering, can reduce the accuracy of the mathematical model (i.e., MPC model uncertainty exists), causing the vehicle's actual trajectory to deviate from the trajectory predicted by the MPC model. When planning a path that requires closely following a phantom (to balance safety and efficiency), this trajectory deviation may lead to a collision between the autonomous vehicle and the phantom obstacle. Therefore, when a phantom is perceived, MPC model uncertainty is also a problem that the path planning module needs to address.

[0005] In summary, addressing the impact of perception phantom information input from the perception module on vehicle collision risk, ride comfort, driving efficiency, and the uncertainty of the MPC model within the high-level path planning module is of great significance for achieving safe and reliable local path planning for autonomous vehicles. Summary of the Invention

[0006] Traditional Model Predictive Control (MPC) local path planning methods solve an optimization problem that balances planning objectives, control smoothness, and safety constraints through online rolling solutions, outputting the optimal planning result. This invention, building upon traditional MPC local path planning, designs a constraint set and cost function specifically for perception phantoms, and uses sparse Gaussian process regression to learn historical MPC path planning data online. The aim is to improve the safety, ride comfort, and driving efficiency of autonomous vehicles operating with perception phantoms.

[0007] This invention provides a local path planning method for autonomous vehicles oriented towards perception phantoms. This method improves upon traditional MPC local path planning, and its steps are as follows:

[0008] (1) Describe local path planning as an MPC problem and establish a mathematical model;

[0009] (2) Establish a local path planning constraint set, and provide sufficient safety error correction opportunities for the vehicle by establishing constraints on vehicle braking distance and effective perception distance, thereby controlling the risk of vehicle collision;

[0010] (3) Establish a cost function that takes into account the probability of the existence of perceived phantoms, and weight the probability of the existence of perceived phantoms on the cost function to avoid path planning being too conservative or too aggressive.

[0011] (4) Use sparse Gaussian process regression to compensate for the uncertainty of the MPC model in the local path planning process;

[0012] (5) Optimize the solution and output the optimal autonomous vehicle path planning result, including vehicle speed and steering wheel angle;

[0013] This enables autonomous vehicles to achieve balanced local path planning while controlling collision risks.

[0014] The technical advantages of this invention are as follows: This invention provides sufficient safety error correction (full braking to stop) opportunities for the vehicle by establishing constraints between the vehicle braking distance and the effective perception distance; it establishes a cost function that considers the probability of the presence of a phantom vehicle, enabling the vehicle to achieve balanced planning; and it uses sparse Gaussian process regression to compensate for the uncertainty of the MPC model in the local path planning process, thereby improving the safety of local path planning at the phantom vehicle location. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of the path planning method of the present invention;

[0016] Figure 2 This is a schematic diagram illustrating the constraints between vehicle braking distance and effective sensing distance.

[0017] Figure 3A schematic diagram of path planning considering the probability of perceiving the existence of phantoms;

[0018] Figure 4 A schematic diagram illustrating the principle of compensating for uncertainties in the MPC model through sparse Gaussian process regression.

[0019] Figure 5 A speed comparison chart showing the planning speeds of the traditional MPC method and the method proposed in this invention. Detailed Implementation

[0020] The present invention will be described in detail below with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention. These embodiments are merely illustrative and should not be construed as limiting the present invention in any way. It should be noted that any modifications and improvements made by those skilled in the art without departing from the concept of the present invention are within the scope of protection of the present invention.

[0021] This invention provides a local path planning method for autonomous vehicles oriented towards perception phantoms. This method improves upon traditional MPC local path planning, and its specific steps are as follows:

[0022] (1) Describe local path planning as an MPC problem and establish a mathematical model;

[0023] (2) Establish a local path planning constraint set, and provide sufficient safety error correction opportunities for the vehicle by establishing constraints on vehicle braking distance and effective perception distance, thereby controlling the risk of vehicle collision;

[0024] (3) Establish a cost function that takes into account the probability of the existence of perceived phantoms, and weight the probability of the existence of perceived phantoms on the cost function to avoid path planning being too conservative or too aggressive.

[0025] (4) Use sparse Gaussian process regression to compensate for the uncertainty of the MPC model in the local path planning process;

[0026] (5) Optimize the solution and output the optimal autonomous vehicle path planning result, including vehicle speed and steering wheel angle;

[0027] In step (2), to address the issue that autonomous vehicles may lose sufficient space and time for safety correction when a phantom vehicle becomes a real obstacle vehicle, leading to a collision, this invention provides sufficient safety correction (full braking to a stop) opportunities for the vehicle by establishing constraints between the vehicle's braking distance and the effective perception distance, thus controlling the risk of vehicle collisions. In step (3), to address the problem that traditional local path planning is too aggressive (ignoring the phantom vehicle) when a phantom is perceived, resulting in a collision between the autonomous vehicle and the phantom vehicle, or too conservative (assuming the phantom vehicle is 100% present), resulting in a low planning speed that affects driving efficiency, this invention establishes a cost function that considers the probability of the presence of a perceived phantom vehicle, enabling the vehicle to achieve a balanced planning of safety and driving efficiency. In step (4), to address the problem that vehicle collisions are caused by the coupling of MPC model uncertainty and perceived phantom, this invention uses sparse Gaussian process regression to compensate for the uncertainty of the MPC model in the local path planning process, thereby improving the safety of local path planning at the phantom vehicle location.

[0028] Figure 1 This is a flowchart illustrating the path planning method of the present invention, applicable to local path planning for autonomous vehicles oriented towards perception phantoms. The specific steps are as follows:

[0029] 1) Establish a mathematical model for MPC local path planning

[0030] In the MPC local path planning framework, the core action of each control cycle is as follows: First, based on the current vehicle state and environmental information, the optimal control quantity (vehicle speed and front wheel steering angle) that satisfies the constraint set of the system within a finite future time domain is optimized and output to the vehicle actuators (powertrain, braking system, steering wheel). Then, the MPC controller, through its internal vehicle dynamics model and this optimal control sequence, iteratively generates a local predicted path (position curve) corresponding to the current decision, moving backward from the current vehicle state. The next cycle uses the real state at the new sampling time as the initial condition for that cycle, triggering a new round of optimization and state iteration. Through this rolling optimization mechanism, continuous local predicted paths are linked end-to-end, forming a continuous local planning path. The mathematical model of MPC local path planning can be described as follows:

[0031]

[0032] Where x is the state variable, including vehicle position and yaw angle; u is the control variable, including vehicle speed and front wheel steering angle; J is the cost function (the deviation between the predicted local path and the global path); and h is the system constraint set for local path planning. The purpose of MPC optimization is to find the optimal control variable that minimizes the cost function while satisfying the system constraints.

[0033] 2) Establish a local path planning constraint set

[0034] In traditional MPC local path planning, the constraint set mainly includes two categories: first, obstacle avoidance constraints to prevent autonomous vehicles from colliding with obstacles on the road; and second, road boundary constraints to prevent vehicles from deviating from the road boundary. However, these constraints can only solve local path planning constraints under deterministic perception information. They have significant limitations when facing highly uncertain obstacle vehicles such as "phantom vehicles" input by the perception module. That is, when a suspected phantom vehicle is suddenly identified as a real obstacle vehicle, the autonomous vehicle may already be at a dangerous following distance, losing sufficient decision space and time for a safe response, thus leading to a sharp increase in the collision risk of the autonomous vehicle.

[0035] To address this critical issue, this invention proposes an innovative constraint mechanism based on the constraints of traditional MPC local path planning: a novel safety error correction constraint is constructed by establishing a constraint relationship between the vehicle's braking distance and the effective perception distance. The core of this constraint lies in dynamically calculating the braking distance at the planned vehicle speed during the planning phase, ensuring it is always less than the effective perception distance to the phantom vehicle ahead. Within this effective perception distance, the confidence level of the information output by the perception module to the higher-level planning module is 100%. Specifically, this invention introduces adaptive safety boundary conditions in the motion planning layer, enabling the planner to consider not only geometric feasibility when generating a path, but also prioritizing ensuring that the vehicle still has sufficient distance and time to perform a full braking stop in the event that the phantom suddenly transforms into a real obstacle, thus securing a crucial safety error correction margin for the system.

[0036] By embedding this constraint into the rolling optimization framework of MPC, this invention achieves feedforward control of vehicle collision risk, fundamentally improving the active safety and decision robustness of autonomous vehicles under uncertain perception information.

[0037] 3) Establish a cost function that considers the probability of perceiving the existence of phantoms.

[0038] Traditional MPC local path planning algorithms treat phantom vehicles as either completely nonexistent or 100% present, leading to a dilemma where "overly aggressive planning increases the risk of collisions" or "overly conservative planning reduces driving efficiency." Furthermore, overly aggressive or overly conservative path planning can cause significant fluctuations in the vehicle speed curves obtained through MPC optimization, impacting the driving comfort of autonomous vehicles. To address this, this invention proposes an innovative probabilistic fusion cost function design method. Breaking away from the traditional black-and-white obstacle modeling approach, it introduces phantom confidence as a weighting coefficient, constructing a J-shaped cost function. C =(1-ω)J A +ωJ BThe weighted cost function. This function dynamically harmonizes the cost J for the complete non-existence of phantoms through the phantom existence probability ω. A The cost of 100% existence of phantom obstacle vehicles J B The contribution ratio allows the planner to generate a trajectory that achieves an optimal balance between safety and driving efficiency without having to make hard decisions in advance.

[0039] This invention further embeds this probabilistic cost function into the MPC framework, enabling dynamic real-time adjustment of the safety margin. As the vehicle approaches the target or the perception information is updated, the system refreshes the ω value online, thereby smoothly adjusting the conservatism of the path: when the probability of a phantom increases, the path automatically increases the avoidance distance; when the probability decreases, it seamlessly reverts to an efficient driving path. This probability-based continuous adaptive mechanism fundamentally avoids the performance mutations or failures caused by fixed thresholds in traditional methods, significantly improving the robustness and traffic efficiency of autonomous vehicles under uncertain perception information.

[0040] Furthermore, in traditional high-level planning module architectures, the behavior decision-making layer needs to pre-output abstract instructions such as "ignore" or "avoid," which the motion planning layer then executes accordingly. This process suffers from instruction discretization and information attenuation. This invention achieves deep integration of the behavior decision-making layer and the motion planning layer in the high-level planning module by establishing a decision cost function that considers phantom probabilities within the motion planning layer, demonstrating significant architectural advantages. By transforming decision rules into continuous probability weights and incorporating them into the cost function, the planning layer directly digests uncertain information during optimization, simultaneously completing behavior decision-making and path generation. This integration mechanism not only avoids delays and errors in instruction conversion between layers, achieving millisecond-level policy response, but also ensures the global optimality of decision-making and planning behavior through the consistency of optimization objectives, enhancing the system's comprehensive ability to cope with complex and uncertain scenarios.

[0041] 4) Use sparse Gaussian process regression to compensate for the uncertainty of the model in the local path planning process.

[0042] Traditional MPC local path planning algorithms do not consider the impact of external disturbances and unmodeled vehicle dynamics on the accuracy of the MPC mathematical model. However, these influencing factors are unavoidable in actual driving, causing uncertainty in the MPC model and ultimately leading to a large deviation between the actual driving route of the autonomous vehicle and the planned route. When the planning module plans a path that requires closely following the phantom, this deviation in trajectory may cause the autonomous vehicle to collide with the phantom obstacle vehicle.

[0043] To address the issue of collisions in autonomous vehicles caused by the coupling of uncertainty in the MPC model and perception phantoms, this invention uses Gaussian Process Regression (GPR) to compensate for the uncertainty in the MPC mathematical model during local path planning. GPR is a numerical fitting method that constructs a fitting function using a prior Gaussian distribution and data sampling points. Under a reasonable prior distribution, the Gaussian process can provide the best linear unbiased posterior distribution for intermediate values. The principle of GPR in learning MPC path planning lies in using GPR to learn and replace the system dynamics that are difficult to model accurately in MPC. Its core idea is to learn from a large amount of historical MPC planning data (such as vehicle state, control input, and optimal output trajectory) to construct a mapping relationship from the current state to the optimal control quantity, and finally output the predicted optimal path point.

[0044] MPC path planning based on Gaussian Process Regression (GPR) involves training the GPR model and predicting test points. However, the training and prediction times of GPR increase exponentially with the amount of historical data. Therefore, it is necessary to approximate GPR to implement Sparse Gaussian Process Regression (SGPR) to reduce the computation time of local path planning. This invention selects historical data points in the MPC path planning process as induction points to perform sparsification processing on GPR, achieving Sparse Gaussian Process Regression. This reduces the complexity of finding the inverse matrix in the Gaussian regression process and improves the timeliness of path planning.

[0045] 5) Output the optimal path planning result for the autonomous vehicle.

[0046] After completing the above steps, the optimization solver solves the optimization problem that satisfies the above system constraint set in real time, and calculates the optimal control quantity that minimizes the above cost function. This control quantity is specifically decomposed into two key execution commands: target speed and target steering wheel angle, and output to the vehicle's underlying control system.

[0047] The target speed is primarily achieved through the vehicle's longitudinal control system. This system coordinates actuators such as the throttle and brakes to precisely adjust the drive torque and braking torque, ensuring that the vehicle's actual speed tracks the target value given by the planning module, thereby enabling acceleration and deceleration.

[0048] The target steering wheel angle is tracked by the lateral control system. The steering controller receives the steering angle command, drives the steering mechanism via a motor, adjusts the front wheel steering angle, thereby controlling the vehicle's direction of travel and enabling the vehicle to accurately track the desired trajectory.

[0049] In summary, this invention establishes constraints and cost functions that take into account perceived phantoms in the high-level planning module, and uses sparse Gaussian process regression to improve the accuracy of the MPC model, enabling autonomous vehicles to achieve balanced local path planning while controlling collision risks.

[0050] Figure 2 This invention establishes a constraint between the braking distance and the effective sensing distance at the point of perception of a phantom. In the figure, 1 represents the autonomous vehicle; 2 represents the phantom vehicle, the exact existence of which is unknown; and 3 represents other vehicles outside the effective sensing distance. The purpose of establishing a constraint between the braking distance and the effective sensing distance is to ensure that the vehicle can apply full braking to a stop within the effective sensing distance and enter a safe state when the phantom is suddenly identified as a real obstacle vehicle.

[0051]

[0052] Among them, S vis S represents the effective sensing distance output by the sensing module to the planning module. min The minimum safe distance that must be maintained between two vehicles to prevent a collision, S brake The braking distance of a vehicle at the planned speed can be calculated using the following formula:

[0053]

[0054] Among them, v S The current speed of the autonomous vehicle. To brake and decelerate the vehicle.

[0055] Figure 3 This diagram illustrates a path planning scheme that considers the probability of perceiving the presence of a phantom vehicle. Scheme A represents the path planning scheme when the phantom vehicle 2 is considered to be completely nonexistent in conventional MPC path planning; Scheme B represents the path planning scheme when the phantom vehicle 2 is considered to be fully present in conventional MPC path planning; and Scheme C represents the path planning scheme that achieves the optimal balance between safety and driving efficiency. Scheme C is implemented based on the cost function considering the probability of perceiving the presence of a phantom vehicle proposed in this invention. This cost function has the form J... C =(1-ω)J A +ωJ B The weighted cost function is dynamically harmonized with the cost J for the complete non-existence of phantoms through the probability ω of phantom existence. A The cost of 100% existence of phantom obstacle vehicles J B The contribution ratio allows the planner to generate a trajectory that achieves an optimal balance between safety and driving efficiency. The cost function for considering the probability of perceiving the presence of a phantom is constructed as follows:

[0056] The cost function J of scheme A A for:

[0057]

[0058] Cost function J of Option B B for:

[0059]

[0060] Among them, e l e o e ψ and u represent the longitudinal error, lateral error, heading error, and control variable between the locally planned position and the globally planned position, respectively; ql, qo, qψ, and qu are the weighting coefficients for the costs of longitudinal error, lateral error, heading error, and control variable, respectively; Np represents the prediction time domain; J obs To avoid the cost of obstacles.

[0061] Option C considers the cost function J of the probability of perceiving the existence of a phantom. C The following can be calculated:

[0062] J C =(1-ω)J A +ωJ B

[0063] Where ω is the probability of the existence of the perception phantom input from the perception module to the planning module.

[0064] Figure 4 This diagram illustrates the principle of MPC path planning based on sparse Gaussian process regression, which mainly includes a historical data acquisition phase, an offline training phase, and an online prediction phase. The historical data acquisition phase collects vehicle state variables (position, yaw angle) and optimal control variables (speed, front wheel steering angle) predicted by the MPC model, as well as the vehicle state variables during actual driving controlled by the optimal control variables, by running a traditional MPC. The offline training phase filters inducement points from this historical data and uses these inducement points and the original historical data to train the SGPR model. The online prediction phase uses the SGPR model trained in the offline training phase and the currently acquired vehicle state to quickly solve for the current optimal control variable and output it to the vehicle actuators (steering, braking, or drive system). The entire process of the online prediction phase is continuously executed in a high-frequency closed-loop manner, thereby achieving real-time response and stable control of the vehicle in dynamic environments.

[0065] The learning-based MPC problem can be described as follows:

[0066]

[0067] In this mathematical model, fest The prediction function established for this invention after adding SGPR, To compensate for the uncertainty of the MPC model using SGPR, f norm The prediction function is established for the traditional MPC model. Z is the input to the GP model. These are the selection matrices for state variables and control variables, respectively. It is optimized by a learning-based path planner. By nominal model Control is achieved. x* and u* are the predicted optimal vehicle state and control variables at the current moment.

[0068] Figure 5 A comparison chart showing the velocity curves obtained by the planning module using the traditional MPC method and the method proposed in this invention. Figure 5 It can be seen that the method proposed in this invention uses phantom probability to make decisions on safety and driving efficiency, resulting in a smoother speed curve and smaller speed change amplitude in the solved control variables. This reduces the frequency of rapid acceleration or deceleration of the vehicle, which is more conducive to the smooth driving of autonomous vehicles.

[0069] The beneficial effects of this invention are as follows:

[0070] The local path planning method proposed in this invention can solve the decision dilemma of local path planning when the perception phantom exists and control the problem of collisions caused by the coupling between the uncertainty of the planning algorithm model and the perception phantom.

[0071] The local path planning method proposed in this invention can solve the path decision planning problem in the presence of perception phantoms, and avoid the risk of vehicle collision caused by the coupling between model uncertainty and perception phantoms by effectively controlling the uncertainty of the planning algorithm model.

[0072] Improving safety: This invention proposes adding the constraint between braking distance and effective perception distance to the traditional MPC path planning constraint set. This can ensure that when a phantom vehicle is suddenly identified as a real obstacle vehicle, the vehicle can perform full braking and stop within the effective perception distance, entering a safe state, thus improving the safety of local path planning when a phantom vehicle is perceived.

[0073] Improving the driving stability of autonomous vehicles: By using the probability of perceiving the existence of phantoms in empirical detection data to weight the cost function in the local path planning process, neutral path planning is achieved, avoiding large changes in the speed of autonomous vehicle planning, thereby ensuring the driving stability of autonomous vehicles.

[0074] Improving the accuracy of path planning models: This invention proposes to use sparse Gaussian process regression to learn historical data of state variables in the MPC path planning process, thereby compensating for model uncertainties. This can improve the accuracy of path planning in observable environments and control the problem of collisions caused by the coupling of uncertainty in the planning algorithm model and perception phantoms.

[0075] Improved efficiency: This invention utilizes sparsity techniques to sparsify the training data for Gaussian process regression, thereby reducing the data complexity of Gaussian process regression and improving the online computational efficiency of local path planning.

[0076] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A local path planning method for autonomous vehicles oriented towards perception phantoms, comprising the following steps: (1) Describe local path planning as an MPC problem and establish a mathematical model; (2) Establish a local path planning constraint set, and provide sufficient safety error correction opportunities for the vehicle by establishing constraints on vehicle braking distance and effective perception distance, thereby controlling the risk of vehicle collision; (3) Establish a cost function that takes into account the probability of the existence of perceived phantoms, and weight the probability of the existence of perceived phantoms on the cost function to avoid path planning being too conservative or too aggressive. (4) Use sparse Gaussian process regression to compensate for the uncertainty of the MPC model in the local path planning process; (5) Optimize the solution and output the optimal autonomous vehicle path planning result, including vehicle speed and steering wheel angle; This enables autonomous vehicles to achieve balanced local path planning while controlling collision risks.

2. The method according to claim 1, characterized in that, During the planning phase, the vehicle braking distance at the planned speed is dynamically calculated and ensured to be always less than the effective perception distance of the phantom vehicle ahead. Within this effective perception distance, the information output by the perception module to the high-level planning module has a confidence level of 100%. Adaptive safety boundary conditions are introduced in the motion planning layer, so that when the planner generates a path, it not only considers the feasibility of geometric space, but also prioritizes ensuring that the autonomous vehicle still has enough distance and time to perform a full braking stop operation in the event that the phantom suddenly becomes a real obstacle.

3. The method according to claim 1, characterized in that, Deep integration of the behavioral decision-making layer and the motion planning layer in the high-level planning module is achieved by establishing a decision cost function that considers the probability of phantom existence in the motion planning layer. By transforming decision rules into continuous probability weights and incorporating them into the cost function, the planning layer directly digests uncertainty information during the optimization process, simultaneously completing behavioral decisions and path generation. The weighted cost function takes the form J. C =(1-ω)J A +ωJ B This function dynamically harmonizes the cost J of the complete non-existence of phantoms through the probability ω of phantom existence. A The cost of 100% existence of phantom obstacle vehicles J B The contribution ratio allows the planner to generate a trajectory that achieves an optimal balance between safety and driving efficiency without having to make hard decisions in advance. As the vehicle approaches the target or the perception information is updated, the system refreshes the ω value online, thereby smoothly adjusting the conservatism of the path: when the probability of a phantom increases, the path automatically increases the avoidance distance, and when the probability decreases, it seamlessly returns to an efficient driving path.

4. The method according to claim 1, characterized in that, Points from historical data in the MPC path planning process are selected as induction points to perform sparse processing on GPR, thereby achieving sparse Gaussian process regression, which reduces the complexity of finding the inverse matrix in the Gaussian regression process and improves the timeliness of path planning.

5. The method according to claim 2, characterized in that, The constraint between braking distance and effective sensing distance is: Among them, S vis S represents the effective sensing distance output by the sensing module to the planning module. min The minimum safe distance between two vehicles to prevent a collision, S brake The braking distance of a vehicle at the planned speed is calculated using the following formula: Among them, v S The current speed of the autonomous vehicle. To brake and decelerate the vehicle.

6. The method according to claim 3, characterized in that, The phantom has no cost. A for: The cost of 100% existence of phantom obstacle vehicles J B for: Among them, e l e o e ψ and u represent the longitudinal error, lateral error, heading error, and control variable between the locally planned position and the globally planned position, respectively; ql, qo, qψ, and qu are the weighting coefficients for the costs of longitudinal error, lateral error, heading error, and control variable, respectively; Np represents the prediction time domain; J obs To avoid the cost of obstacles.

7. The method according to claim 4, characterized in that, The MPC path planning based on sparse Gaussian process regression includes a historical data collection stage, an offline training stage, and an online prediction stage. The historical data collection stage collects the vehicle state variables and optimal control variables predicted by the MPC model and the vehicle state variables during actual driving controlled by the optimal control variables by running the traditional MPC. The offline training stage selects guidance points from the historical data and uses the guidance points and the original historical data to train the SGPR model. The online prediction phase utilizes the SGPR model trained in the offline training phase and the actual vehicle state collected at the current moment to quickly solve for the current optimal control quantity and output it to the vehicle actuator. The entire process of the online prediction phase is continuously executed in a high-frequency closed-loop manner, thereby realizing the real-time response and stable control of the vehicle in the dynamic environment.

Images