Vision-Based Eco-Oriented Driving Strategies for Freeway Scenarios Using Deep Reinforcement Learning

The rapid development of sensor technologies and machine learning algorithms has prompted the automotive industry to take big steps towards autonomous driving. However, complex driving scenarios make the task extremely challenging. The current technologies are sensors-hungry, and the systems are comprehensive to achieve robust behavior in the real world, which causes an autonomous driving system to be unwillingly expensive. Leveraging the computer vision techniques, the camera as a relatively cheaper sensing option is becoming crucial in the system. Using the emerging game engine-based simulation platform, the development of vision-based autonomous driving achieved by deep reinforcement learning becomes practical and more efficient. Besides, energy consumption and greenhouse gas emissions have been a major concern in recent years due to the increased travel demand. Environmental sustainability of our transportation system has become a significant topic for researchers and engineers. Motivated by all these, this thesis presented an end-to-end solution for advanced driving assistance strategies for freeway scenarios using vision-based deep reinforcement learning technique. The system is designed and tested in an advanced driving simulation platform based on game engine and can learn optimal control actions directly from raw input images captured by the front on-board cameras. The trained vehicle has robust and safe driving behavior following a preceding vehicle or cruising on the straight freeway. The system can be well adaptive to different speed trajectories of the preceding vehicle and run in real-time. Novel adaptive gap-based and energy model-based reward functions are designed for eco-oriented driving. The method achieved 7.2 – 21.7% energy savings compared with other methods led by strict gap-based reward or force-based reward in the car following strategy