Vehicles are very important in our daily life. They are often on the move at significant speeds and are equipped with significant embedded computing systems that control and manage different functions including providing various types of assistance to its driving function. The computational capabilities of automobile systems provide opportunities and challenges to remedy various known issues caused by modern vehicles, such as environmental pollution, road congestion and fatalities.

To address these problems, we propose a system design principle that utilizes various sensing, computational and control capabilities of modern vehicles to monitor, assist or even replace in-vehicle human drivers in improving driving performance and experience with low-cost hardware and small deployment efforts. Under such design principle, the systems sense and model vehicle dynamics from vehicle parameters and third-party sensors, based on which they provide guidance and control to assist or even replace in-vehicle human drivers and achieve better driving performance and experience.

We start by deploying a smartphone application called RideSense to enhance existing approaches for whitespace determination. RideSense senses vehicle dynamics by built-in sensors and provides feedback to drivers on aggressive events to improve their driving safety awareness. Sensing vehicle dynamics by smartphone sensors require coordinate alignment between the smartphone and the car. We found that even gentle road slopes may cause severe coordinate misalignment and acceleration over/underestimation. To resolve these problems, we propose slope-aware coordinate alignment algorithm and linear acceleration estimation method to reduce alignment training time and improve linear acceleration estimation accuracy. Given the improved vehicle dynamics knowledge and the understanding of human control limitations, we then focus on extending the sensing capability to sense more comprehensive vehicle parameters and control vehicle based on well-tuned models and algorithms. We present a vehicle sensing and control system module called EcoDrive. EcoDrive models instant fuel consumption based on vehicle parameters collected from On-board diagnostics (OBD) port. According to the model, it controls the gas pedal position sensor to adjust fuel injection rate according to road segment distance and speed limit. By using careful control of fuel injection rate, it is able to improve fuel efficiency comparing to human drivers.

In the last part of the thesis, we discuss the limitations of self-driving systems and propose a remote monitoring and control system to augment self-driving systems. The intuition is that a human operator can control the vehicle remotely, when self-driving system failures occur, which may be due to bad weather, malfunction, a contradiction in sensory inputs, and other such conditions. We present a live streaming and remote control system to handle self-driving system failures, called RTDrive. RTDrive consists of a context-aware video encoding method and a live streaming protocol. The context-aware video encoding method can improve video streaming quality by adjusting encoding parameters according to vehicle dynamics. We also implement a consistent-latency view mechanism to smooth the video frames, under which the remote driver can have more precise control over the vehicle.

We believe that our driving analytics application, economic driving system, and remote control system are useful in enhancing driving performance and experience. Furthermore, such a design principle can also be applied to broader applications and systems that can assist or even replace human drivers. We also discuss how such a design principle can be utilized for our future work.