汽车的智能化是增强汽车行驶性能的重要途径。目前，紧急制动系统（AEB）主要感知自车前方的障碍物，局限于纵向维度的追尾避撞。具有自动驾驶功能的汽车，可同时感知侧向车道的机动车、行人等信息，这使得动态交通场景下汽车横向避撞控制成为可能，具有进一步提升汽车行驶安全性的潜力。面向动态交通场景的主动避撞控制是一个典型的非线性、带约束的预测型控制问题，它通过滚动时域优化的方式，在线求解控制命令，然而现有求解器受实时性不足的制约，尚不能满足毫秒级的车载控制器计算需求。针对上述难题，本文以智能汽车的运动控制为研究对象，提出了一种在线计算高效且具有算力自适应能力的预测控制快速求解算法，用于动态交通场景的汽车主动避撞控制，为提升汽车行驶安全性奠定了基础。首先，将车辆横向主动避撞控制构建为有限时域、带约束的预测型最优控制问题。建立了考虑轮胎附着极限的自车非线性车辆动力学模型和周车运动学模型；设计了双圆轮廓描述的车辆预测型碰撞约束，以及考虑横摆角速度和轮胎侧偏角的自车稳定性约束；引入惩罚函数处理车辆安全性约束，设计了综合两类约束和轨迹跟踪误差的目标函数，完成对主动避撞控制问题的构建。其次，提出了具有高实时在线计算能力的循环模型预测控制算法（RMPC，Recurrent Model Predictive Control），解决了预测型汽车运动控制器计算效率不高的难题。利用循环函数近似不同预测步数控制问题的最优解，并根据主动避撞控制问题的目标函数设计离线训练循环策略的损失函数，将在线优化问题转化为循环策略参数的离线求解；利用万能近似定理和Bellman最优性原理，证明了该算法的收敛性和循环策略的最优性。循环策略的循环次数可由控制器的实际算力决定，实现了算力自适应的高实时在线计算。最后，为进一步验证汽车避撞控制算法的控制性能和计算效率，将算法部署至快速原型控制器，结合CarSim车辆模型进行大规模硬件在环（HIL，Hardware-In-the-Loop）仿真试验。通过1000次随机交通场景验证表明：随着预测步数的增加，最小安全车距由0.34m提升至1.38m，千次试验碰撞次数由44次下降到0；与常用预测控制求解器COBYLA、L-BFGS-B和SLSQP相比，RMPC在线计算效率提升超过900倍（当预测时域为20步时）。
Automated vehicle technology has great potential to enhance the vehicle driving performance. Autonomous Emergency Braking (AEB) system, which detects the obstacles ahead, is only suitable for the longitudinal collision avoidance. Attributing to the development in automated vehicle sensors, the information of surrounding vehicles and pedestrians can be detected as well. It is possible to achieve the lateral collision avoidance control in dynamic traffic scenarios. Such technique is anticipated to further improve driving safety. While active collision avoidance control for dynamic traffic scenarios is a typical nonlinear, constrained predictive control problem. The existing optimizing solvers cannot satisfy the millisecond-level requirements of on-board vehicle controllers. This study aims to develop a real-time computational algorithm for collision avoidance control of automated vehicles, which has adaptive capability for hardware computational resource.Firstly, the collision avoidance control problem is constructed as a finite-horizon predictive optimal control problem. The nonlinear vehicle dynamics model with tire adhesion limit and the surrounding vehicles’ kinematic model are established. The predictive collision constraints described by the double-circle, and the vehicle stability constraint considering the yaw rate and the tire sideslip angle are designed. Penalty functions are introduced to deal with the safety constraints, and the objective function considering the constraints and tracking errors is designed. Secondly, the recurrent model predictive control algorithm (RMPC) with real-time online computational ability is proposed, which aims to solve the challenge of computational efficiency of predictive vehicle collision avoidance controller. Use the recurrent function to approximate the optimal solution in different prediction horizon, and design the cost function for off-line training process with the objective function of the active collision avoidance control problem. Then the on-line optimization problem can be transformed into the off-line calculation problem of the recurrent strategy parameters. With the universal approximation theorem and Bellman optimality principle, the convergence and optimality of the algorithm are proved. The number of recurrent time is determined by the computational resource, which realizes the real-time on-line calculation with the ability to adapt computational resource.Finally, in order to verify the control performance and calculation efficiency of the proposed collision avoidance control algorithm, the algorithm is deployed to the rapid prototype controller. Large-scale Hardware-In-the-Loop (HIL) simulation tests are carried out in combination with the CarSim platform. Take the control verification of 1000 random traffic scenarios experiments as an example: With the increasing of prediction horizon, the minimum vehicle distance increases from 0.34m to 1.38m, and the number of collisions decreases from 44 to 0 for 1000 tests using the proposed RMPC algorithm. Compared with the typical MPC algorithms, such as COBYLA, L-BFGS-B and SLSQP, the computational efficiency is increased by more than 900 times (prediction horizon is 20).
