Centralized Design And Implementation Of Intelligent Vehicle Planning And Control System Based On Fast Predictive Control

Intelligent vehicles can comprehensively enhance the industry’s basic capabilities,improve traffic efficiency,and ensure life safety,make it a significant strategic choice for China to transform from a major automobile country to a powerful automobile country.Motion planning and motion control are the two key technologies for intelligent vehicles to achieve autonomous driving,and their performance determines the safety,riding comfort,and vehicle stability.To further improve the safety and tracking accuracy of automated driving planning,motion planning and motion control are designed together.When performing centralized design,it is necessary to consider the nonlinear and strongly coupled motion characteristics of vehicles,and simultaneously meet the requirements of multi-objective control such as obstacle avoidance and lane change,speed tracking,as well as constraints such as vehicle actuators and road boundaries.Nonlinear model predictive control is used for centralized motion planning and motion control design because it can handle nonlinear,multi-objective,and multi-constraint problems.However,nonlinear predictive control is difficult to optimize and solve quickly,and motion planning increases the overall prediction time.In fast dynamic scenes,intelligent vehicles have higher requirements for real-time controller performance.In response to the problem of realtime optimization of the controller,this paper investigates control optimization design,real-time optimization algorithms,and acceleration experimental verification from three aspects.The main work of this paper is as follows:1.To meet the safety requirements of autonomous driving,a centralized design approach for motion planning and motion control is proposed,and the computational efficiency of prediction controller is improved through the variable prediction step size method.Firstly,a simplified nonlinear vehicle model is established based on the tire brush model.Then,using the artificial potential field method,the road boundary model,lane line model,and centerline model are respectively established in structured roads,and obstacles are modeled.Secondly,a nonlinear prediction controller with variable prediction step size is designed according to the different sampling requirements of motion control and motion planning.Finally,different long and short-term schemes are built in MATLAB and verified through joint simulation with CarSim under different working conditions.The offline experimental results validate the effectiveness of the controller design and the acceleration effect of the variable prediction step size scheme.2.A parallel optimization method based on calculation graph is proposed to achieve fast optimization solution for nonlinear predictive control of motion planning and control.Firstly,the overall optimization goal of the controller is decomposed into individual time-domain objective functions,and the vehicle’s future dynamic prediction is carried out through forward propagation,and all time-domain vehicle states are obtained recursively,with their relevant information being stored in matrices.Secondly,in all single time-domain objective functions,combined with the stored values from forward propagation,the gradients of the state variables and control variables are calculated through back propagation.Lastly,combining dynamic programming and back propagation,the recursive equation of the overall objective function gradient with respect to each time-domain control variable is obtained,and relevant calculations are performed with the stored gradient values through gradient descent optimization.Since there is no coupling relationship between the gradient calculation process of each time-domain,parallel calculation can be performed.Offline simulation experiments are conducted under the condition of a single stationary obstacle,which verifies the effectiveness and robustness of the algorithm optimization solution,and has a significant improvement in computational performance compared with sequential quadratic programming algorithms.3.To further accelerate the real-time performance of the algorithm in the in-vehicle controller,a FPGA-based algorithm acceleration implementation plan is proposed.Firstly,the algorithm is decomposed in detail,and since there is no coupling relationship between the backward propagation of each time domain in the algorithm,loop unrolling can be used to accelerate it.In addition,there are many matrix calculations in the forward propagation,backward propagation,and dynamic programming of the algorithm,which can be accelerated by loop pipelining.Then,suitable clock frequencies are selected in the high-level synthesis tool,and comprehensive experiments are conducted with loop unrolling,loop pipelining,and loop pipelining combined with loop unrolling based on the results of algorithm decomposition.The acceleration performance of the algorithm is analyzed based on experimental results.Finally,a heterogeneous design is performed based on the selected synthesis plan that meets the resource and timing requirements,and a hardware-in-the-loop experiment is conducted on the built real-time platform to verify the effectiveness and real-time performance of the algorithm.