Applying Deep Learning to the Safety of Autonomous Ground Vehicles

最近，自主地面车辆的研究已经走出实验室，进入到当地的交通甚至高速公路。虽然自动驾驶越来越成为现实，但自动驾驶汽车的道路安全更受到制造商、政策制定者和公众的关注。目前，汽车交通事故每年都会造成大量人员伤亡，造成巨大的经济损失。据统计，事故发生的主要原因有三个：驾驶员失误、环境因素和车辆部件故障或退化。从理论上讲，自动驾驶汽车在几乎所有的事故场景中都可以胜过人类驾驶员，并导致更少的车祸。然而，它需要大量的研究工作，以实现其在道路安全方面的全部潜力。来自不同学科的许多技术已经被开发并应用于自动驾驶汽车。在过去的十年中，在另一个研究领域，我们见证了深度学习（DL）技术的蓬勃发展，包括深度强化学习（DRL），深度自动编码器，深度卷积神经网络，长短期记忆和深度信念网络。这些方法最近极大地推动了机器人、计算机视觉和控制等不同领域的技术发展。这些深度网络在包括自动驾驶汽车设计和操作在内的许多研究领域都有很大的潜力，可以实现比传统技术更好的性能。鉴于自动驾驶汽车的优势以及人工智能或深度学习的快速发展，这些技术的结合具有很大的潜力，通过将深度学习技术应用于许多具有挑战性的驾驶场景，可以实现近乎无事故的自动驾驶体验。我的五年计划是开发和应用新型深度学习算法和深度神经网络架构，用于无碰撞路径规划、故障检测和故障诊断以及道路紧急情况检测和控制。本文将沿着以下三个研究方向展开研究：1）动态障碍物下的无碰撞路径规划（DRL）; 2）基于传感器融合的道路紧急事件检测（DL）; 3）故障检测与诊断（DL）。 我相信这些项目的成功完成将大大提高自动驾驶地面车辆的性能和道路安全性，并为几乎无事故的自动驾驶体验铺平道路。它将节省人类生命和车祸造成的费用。这对汽车制造商、保险公司和公众也将大有裨益。

Recently, the research of autonomous ground vehicles has moved out of the lab to local traffic and even highways. While autonomous driving has more and more become the reality, road safety with autonomous vehicles is more of a concern for manufactures, policy makers and the public. At present, vehicle accidents result in many human fatalities and injuries every year, and significant economic loss. According to statistics, accidents occur mainly due to three critical reasons: driver errors, environment factors and vehicle components' failure or degradation. Autonomous vehicles, in theory, can excel over human drivers in almost all accident scenarios and result in much fewer car accidents. However, it needs a significant amount of research work to realize its full potential for road safety. Many techniques from different disciplines have been developed and applied to autonomous vehicles. In the past decade, on another research front, we have witnessed a boom of deep learning (DL) techniques, including deep reinforcement learning (DRL), deep auto-encoders, deep convolutional neural networks, long short-term memory, and deep belief networks. These methods have recently dramatically pushed forward the state of the art in diverse domains such as robotics, computer vision and control. These deep networks have a great potential to achieve a much better performance than traditional techniques in many research fronts including autonomous vehicles design and operation. Given the advantages of autonomous vehicles and the rapid advances of artificial intelligence or deep learning, the combination of these technologies has a great potential to achieve a near accident-free autonomous driving experience by applying deep learning techniques to many challenging driving scenarios. My five-year plan is to develop and apply novel deep learning algorithms and deep neural network architectures for collision-free path planning, fault detection and fault diagnosis, and road emergency detection and control. I will investigate along the following three research directions: 1) Collision-free path planning with dynamic obstacles using DRL, 2) Sensor fusion based road emergency detection using DL, and 3) Fault detection and fault diagnosis using DL. I believe that the successful completion of these projects will greatly improve the performance of autonomous ground vehicles and the road safety, and pave the way for a near accident-free autonomous driving experience. It will save human lives and costs resulting from the car accidents. It will also be of great benefit to vehicle manufacturers, insurance companies and the general public.