随着汽车工业的发展与人工智能技术的进步，人们对自动驾驶汽车的憧憬正逐渐变为现实。但是，现阶段的自动驾驶技术还不成熟，当自动驾驶汽车离开实验室进入到公共交通时，必须具备处理紧急情况的能力。本研究针对高速紧急避撞这一典型危险工况，提出了自动驾驶汽车紧急避撞系统的设计方案，对运动控制与决策机制中所涉及的重要基础理论和关键技术问题开展了研究。 首先，在考虑系统非线性、不确定性及未知外界干扰的条件下，通过反演控制、变结构控制、自适应控制理论及神经网络技术，并采用分层控制构架的方案，设计了自动驾驶汽车的纵、横向运动控制策略，提高了闭环系统的鲁棒性和稳定性，在一定程度上解决了在极限工况下智能车辆的操控难题。 其次，基于五次多项式和边界条件得到避撞轨迹的初始表达式，并从运动学的角度，利用避撞轨迹方程推导出了理想的车辆横摆角速度公式，通过遗传算法得出理想横摆角速度的等效最大值表达式。同时，从动力学的角度，利用路面附着约束条件获得车辆横摆角速度的极限值。再通过合理地设计避撞轨迹的终点坐标，建立能够同时考虑车辆碰撞因素与失稳因素的风险评估模型，为自动驾驶汽车在紧急情况下的风险量化及有效评估提供了解决方案。 再次，为了产生初始训练数据，通过分析优秀驾驶员执行紧急避撞操纵的范例和轮胎动力学特性，设计了基于规则的行为决策系统。基于Softmax分类器的方案设计了模仿学习算法，并采用小批量随机梯度下降算法离线训练模仿学习的神经网络。提出了基于模型的值函数近似Q(λ)-学习驾驶行为决策方法，并利用模仿学习算法学得的策略模型作为自动驾驶汽车进行强化学习的初始策略，从而改善了驾驶行为决策的效率及效果，为智能车辆在小样本条件下在线序列决策提供了技术手段。 最后，利用联合仿真平台和基于线控执行器的硬件在环试验平台，分别对本文所提出的方案进行了仿真分析与试验研究。结果表明，本文所设计的系统能够有效地控制自动驾驶汽车执行紧急避撞操纵并维持车辆稳定。

With the development of automobile industry and artificial intelligence technology, autonomous vehicles are coming into reality. However, the autonomous driving technology is not mature at the current stage. When autonomous vehicles leave research laboratory and enter into public traffic, they must be able to deal with emergency situations. This paper proposed a collision avoidance system for autonomous vehicle in emergency situations at high speed. Meanwhile, the important basic theories and key technical involved in motion control and decision-making mechanism are researched. Considering system nonlinearity, uncertainty and unknown external disturbance, the longitudinal and lateral motion control strategies for autonomous vehicle are designed by combining backstepping control method, variable structure control scheme, adaptive control theory and neural network technology, and using hierarchical control framework. The proposed strategy can improve robustness and stability of the closed-loop system, and to some extent, it can also solve the problem that the vehicle is difficult to control at driving limits The fifth-order polynomial equation and boundary conditions are adopted to obtain the initial expression of collision-free trajectory. Then, from the aspect of kinematics, an ideal vehicle yaw rate formula is derived from the collision avoidance trajectory expression, and the genetic algorithm is used to obtain the equivalent maximum expression of the desired yaw rate. Meanwhile, from the aspect of dynamics, the maximum expression of vehicle yaw rate is derived from the constraint condition of road surface adhesion. Finally, through the reasonable design of terminal point coordinates for the collision-free trajectory, a risk assessment model which can simultaneously consider the risk associated with collision and destabilization is derived. The proposed scheme can provide an effective solution to quantify and evaluate the risk for autonomous vehicle in emergency situations. In order to generate initial training data, a rule-based behavior decision-making system is designed by analyzing an excellent driver's emergency collision avoidance manipulation and vehicle dynamics characteristics. Then, an imitative learning algorithm is designed with the Softmax classifier, and the neural network of the imitative learning is trained offline through the mini-batch stochastic gradient descent (MSGD) algorithm. A model-based value function approximation Q(λ)-learning algorithm is designed, and the policy model learned by the imitation learning algorithm is used as the initial strategy for the reinforcement learning. The proposed method can improve the efficiency and effectiveness of behavior decision-making, and it provides a technical means to online sequential decision under the condition of small sample. The simulation analysis and experimental study are performed using the co-simulation platform and the hardware-in-the-loop (HIL) system, respectively. According to the results, the proposed scheme can effectively perform an emergency collision avoidance maneuver while stabilising vehicle.