Cross-Layer System Design for Autonomous Driving

Abstract

Autonomous driving has gained tremendous popularity and becomes one of the most emerging applications recently, which allows the vehicle to drive by itself without requiring help from a human. The demand of this application continues to grow leading to ever increasing investment from industry in the last decade. Unfortunately, autonomous driving systems remain unavailable to the public and are still under development even with the recent considerable advancement achieved in our community. Several key challenges are observed across the stack of autonomous driving systems and must be addressed to bridge the gap.

This dissertation investigates cross-layer autonomous driving systems from hardware architecture, software algorithms to human-vehicle interaction. In the hardware architecture layer, we investigate and present the design constraints of autonomous driving systems. With an end-to-end autonomous driving system framework we built, we accelerate the computational bottlenecks identified and thoroughly investigate the implications and trade-offs across various accelerator platforms. In the software algorithm layer, we propose an accelerating technique for object recognition, which is one of the critical bottlenecks in autonomous driving systems. We exploit the similarity across frames in streaming videos for autonomous vehicles and reuse the intermediate outputs computed in the algorithm to reduce the computation required and improve the performance. In the human-vehicle interaction layer, we design a conversational in-vehicle interface framework which enables drivers to interact with vehicles by using natural human language to improve the usability of autonomous driving features. We also integrate this framework into a commercially available vehicle and conduct a real-world driving study.