Artificial intelligence empowered autonomous vehicles on-demand system for smart cities

Abstract

An effective intelligent transportation system is a core part of a modern smart city to support various human activities. One of the state-of-the-art technologies to provide significant improvement to the next generation transportation system is the Autonomous Vehicle (AV). AVs are vehicles that can autonomously navigate on the road without the control of a human driver. AV technology can shorten the distance between vehicles in vehicle platooning, relieve drivers from intensive driving, and enhance road safety. However, the impact of AVs is far beyond the above-mentioned features if the system is properly designed. Therefore, it is promising to investigate and exploit the unique properties of AVs and to develop advanced AV systems with various additional benefits. In this thesis, we propose a novel AV system called Autonomous Vehicles on Demand (AVoD) for future smart cities. AVoD is designed for advanced transportation systems with AVs using numerous artificial intelligence techniques. It focuses on exploiting the autonomy of AVs to enhance transport efficiency and social benefits. The AVs support highly efficient on-demand transport and vehicle-to-grid services by using artificial intelligence techniques to forecast the future travel demand, modify the traffic signaling of road lanes and intersections, and optimize the routing and scheduling of AVs, based on real-time and historical data. The corresponding AVoD strategies are presented in different chapters in the thesis. First, we propose state-of-the-art forecasting deep learning models for traffic data to improve the prediction accuracy. We propose a novel spatial-temporal deep learning model called Multi-Scale Convolutional Long Short-Term Memory network for predicting future travel demand which is useful for the AVoD strategies. For graph-structured data, we developed a semi-supervised graph convolutional network for identifying the waiting time level at public transport stations based on proxy data and a small amount of labelled data. Second, AVs, as transportation service providers managed by our proposed optimal routing and scheduling method, can pick up the passengers who request transport services. After dropping off the passengers, the battery-powered empty AVs without new requests can park at the high travel demand location for reducing the unnecessary waiting time of potential passengers using the predicted traffic data, or at electric vehicle charging stations for vehicle-to-grid support using our designed re-balancing strategy. Finally, we propose the road infrastructure adaptation for transportation systems with AVs to enhance the road throughput. We propose dynamically reversible lane which is designed for AV system based on real-time on-road traffic to increase the road throughput. At the road intersections, we propose an end-to-end traffic signal control scheme using deep reinforcement learning that can adapt to real-time traffic to mitigate the stop-and-go traffic caused by inappropriate traffic signal control timing. The above AVoD strategies were tested in corresponding testbeds and simulators to demonstrate the potential of AVoD over the traditional transportation system.