CN-121979206-A - Automatic driving vehicle local path planning method oriented to perception phantom

CN121979206ACN 121979206 ACN121979206 ACN 121979206ACN-121979206-A

Abstract

The invention discloses a local path planning method of an automatic driving vehicle facing to a perception phantom, which comprises the following steps of (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 perception phantom, (4) compensating uncertainty of the MPC model in the local path planning process by using sparse Gaussian process regression, and (5) optimizing and solving to output an optimal automatic driving vehicle path planning result. The method has the technical advantages that safety error correction opportunities are provided for the vehicle by establishing constraint of the vehicle braking distance and the effective perception distance, a cost function considering the existence probability of the phantom vehicle is established, so that the vehicle realizes balanced planning, uncertainty of the MPC model is compensated by using sparse Gaussian process regression, and the safety of local path planning at the phantom vehicle is improved.

Inventors

WANG LIFEN
JIAN XIONG
WANG YAOHUI

Assignees

北京交大思诺科技股份有限公司

Dates

Publication Date: 20260505
Application Date: 20260114

Claims (7)

1. A method for planning a local path of an automatic driving vehicle facing to perception phantom comprises the following steps: (1) Describing local path planning as an MPC problem, and establishing a mathematical model; (2) Establishing a local path planning constraint set, and providing enough safety error correction opportunities for the vehicle by establishing the constraint of the vehicle braking distance and the effective sensing distance to control the collision risk of the vehicle; (3) Establishing a cost function considering the existence probability of the sensing phantom, weighting the cost function by the existence probability of the sensing phantom, and avoiding too conservation or too aggressive path planning; (4) Compensating uncertainty of the MPC model in the local path planning process by using sparse Gaussian process regression; (5) Optimally solving, and outputting an optimal automatic driving vehicle path planning result, wherein the optimal automatic driving vehicle path planning result comprises the speed of a vehicle and the steering wheel angle; the automatic driving vehicle realizes balanced local path planning on the premise of controlling collision risk.
2. The method according to claim 1, characterized in that the vehicle braking distance at the planned speed is calculated dynamically during the planning phase and is ensured to be always smaller than the effective perceived distance of the front phantom vehicle, the confidence level of the information output by the perceived module to the high-level planning module is 100% within the effective perceived distance, and the self-adaptive safety boundary conditions are introduced in the motion planning layer, so that the planner not only considers the feasibility of the geometric space when generating the path, but also ensures that the automatic driving vehicle still has enough distance and time to execute the full-braking parking operation under the condition that the phantom is suddenly changed to a real obstacle.
3. The method of claim 1, wherein the depth fusion of the motion planning layer and the action decision layer in the high-level planning module is realized by establishing a decision cost function considering the existence probability of the phantom in the motion planning layer, the uncertainty information is directly digested by the planning layer in the optimization process by converting a decision rule into a continuous probability weight and fusing the continuous probability weight into the cost function, the action decision and the path generation are synchronously completed, the weighted cost function is in the form of J C =（1-ω）J A +ωJ B , the function dynamically reconciles the contribution proportion of the total absence cost J A of the phantom and the 100% existence cost J B of a phantom obstacle vehicle through the existence probability omega, the planner can generate a track realizing optimal balance between safety and driving efficiency without making hard decisions in advance, and the system refreshes omega values online as the vehicle approaches to update the target or perception information, so as to smoothly adjust the conservation degree of the path, namely the path automatically increases the avoidance distance when the phantom probability increases, and the path is recovered to the efficient driving path without sense when the probability decreases.
4. The method of claim 1, wherein points of historical data in the MPC path planning process are selected as the induction points to perform sparse processing on the GPR to achieve sparse gaussian process regression, thereby reducing complexity of an inversion matrix in the gaussian process and improving timeliness of path planning.
5. The method of claim 2, wherein the constraint between braking distance and effective perceived distance is: Wherein S vis represents the effective sensing distance output to the planning module by the sensing module, S min is the minimum safe distance between two vehicles for preventing the collision of the vehicles, S brake is the braking distance of the vehicles at the planned speed, and is calculated by the following formula: wherein v S is the current running speed of the automatic driving vehicle, Braking deceleration of the vehicle.
6. A method according to claim 3, wherein the total absence of phantom cost J A is: The phantom obstacle vehicle 100% present cost J B is: The method comprises the steps of determining a local path planning position, wherein e l 、e o 、e ψ and u are longitudinal errors, lateral errors, heading errors and control quantities between the local path planning position and the global planning position, ql, qo, qψ and qu are weight coefficients of the longitudinal errors, the lateral errors, the heading errors and the control quantity cost, np is a prediction time domain, and J obs is obstacle avoidance cost.
7. The method of claim 4, wherein the MPC path planning based on sparse Gaussian process regression comprises a historical data acquisition stage, an offline training stage and an online prediction stage, wherein the historical data acquisition stage is used for collecting vehicle state quantity and optimal control quantity predicted by an MPC model and vehicle state quantity in the actual running process of a vehicle controlled by the optimal control quantity by running a traditional MPC, the offline training stage is used for screening induction points from historical data and training an SGPR model by using the induction points and original historical data, the online prediction stage is used for rapidly solving the current optimal control quantity by using the SGPR model trained in the offline training stage and the actual state of the vehicle acquired at the current moment and outputting the current optimal control quantity to a vehicle executing mechanism, and the whole process of the online prediction stage is continuously and circularly executed in a high-frequency closed-loop mode, so that real-time response and stable control of the vehicle in a dynamic environment are realized.

Description

Automatic driving vehicle local path planning method oriented to perception phantom Technical Field The invention relates to the field of automatic driving vehicle local path planning, in particular to a local path planning method aiming at sensing the existence of phantom. Background An automatic driving vehicle, also called an unmanned driving vehicle, is a revolutionary traffic solution integrating artificial intelligence, high-precision sensing, internet of things and advanced computing technology. The intelligent travel system aims to partially or completely replace a human driver through an intelligent system and realize safe, efficient and convenient autonomous travel. Autopilot technology is the brain and soul of an autopilot vehicle, which is the sum of a complex set of algorithms, software and computing systems that gives the autopilot vehicle the ability to understand the world, think judgment and make actions. The automatic driving technology can be divided into an environment sensing module, a high-level planning module and a tracking control module according to functions, wherein the high-level planning module comprises a behavior decision layer and a motion planning layer. The implementation of autopilot technology relies on a high degree of coordination between the modules, which is a complex system engineering of close fit, endless. Firstly, the perception module accurately recognizes key information such as obstacles, lane lines, traffic signs and the like around a vehicle in real time by fusing various sensor data such as a camera, a laser radar, a millimeter wave radar and the like. The high-level planning module searches an optimal or suboptimal path as a global path according to the known environment. When the environment changes, the high-level planning module comprehensively analyzes real-time obstacle sensing information, a high-precision map and positioning, evaluates the feasibility of the global path and makes action decisions (such as following, overtaking and stopping) so as to plan a safe, smooth and local path conforming to traffic rules. Finally, the tracking control module precisely executes the planned track instruction through the drive-by-wire, steering and braking technology to control the speed and direction of the vehicle. The local path planning is an indispensable core technology in the field of automatic driving research, and the core idea is that a high-level planning module avoids obstacles according to real-time perception data on the basis of a global planning path to generate a safe and smooth driving track. Thus, the accuracy and safety of local path planning is severely dependent on the real-time environmental information provided by the perception module. However, problems such as noise, shielding, delay and the like of sensor information in an actual driving process can cause perception uncertainty (the perception accuracy of a perception module is reduced), and whether the identified obstacle vehicle exactly exists cannot be accurately judged, namely, the perception module has perception phantom. When the high-level planning module receives the sensing illusion containing the illusion uncertainty information (illusion existence probability) input by the sensing module, the obstacle avoidance algorithm is designed to avoid the illusion obstacle vehicle far from simple, and a flexible probability decision and planning framework based on the illusion existence probability is needed. The core of the framework is to abandon the traditional method of planning according to the complete existence or complete absence of the perception phantom, and generate a track which has controllable collision risk, safety and efficiency and predictable vehicle state (position and yaw angle) through evaluation and balance. The Model Predictive Control (MPC) can optimize control quantities such as steering angle and vehicle speed of the vehicle in real time by a rolling optimization mode under the condition of fully considering various variable constraints, and predicts the vehicle state according to the optimal control quantity and the vehicle dynamics model so as to realize local path planning. Thus, MPC is widely used for autonomous vehicle local path planning. When the MPC is used for local path planning, an internal mathematical model is required to be established according to a vehicle dynamics model, the accuracy of the mathematical model obviously influences the path planning effect, but unavoidable external disturbances such as road gradient and adhesive force changes and unmodeled dynamics such as nonlinear tire force during high-speed over-bending can cause the accuracy of the mathematical model to be reduced (namely, model uncertainty exists in the MPC model), so that the actual motion track of the vehicle deviates from the track predicted by the MPC model. When a path is planned that needs to be closely tied to the phantom (to balance safety and efficiency), this de