| Summary: | With increasing urbanization and the number of cars on road, there are many global issues on modern transport systems. Autonomous driving and connected vehicles are the most promising technologies to tackle these issues. The so-called integrated technology connected autonomous vehicles (CAV) can provide a wide range of safety applications for safer, greener and more efficient intelligent transport systems (ITS). As computing is an extreme component for CAV systems, various mobile computing models including mobile local computing, mobile edge computing and mobile cloud computing are proposed. However it is believed that none of these models fits all CAV applications, which have highly diverse quality of service (QoS) requirements such as communication delay, data rate, accuracy, reliability and/or computing latency. In this thesis, we are motivated to propose a hybrid mobile computing model with objective of overcoming limitations of individual models and maximizing the performances for CAV applications. In proposed hybrid mobile computing model three basic computing models and/or their combinations are chosen and applied to different CAV applications, which include mobile local computing, mobile edge computing and mobile cloud computing. Different computing models and their combinations are selected according to the QoS requirements of the CAV applications. Following the idea, we first investigate the job offloading and allocation of computing and communication resources at the local hosts and external computing centers with QoS aware and resource awareness. Distributed admission control and resource allocation algorithms are proposed including two baseline non-cooperative algorithms and a matching theory based cooperative algorithm. Experiment results demonstrate the feasibility of the hybrid mobile computing model and show large improvement on the service quality and capacity over existing individual computing models. The matching algorithm also largely outperforms the baseline non-cooperative algorithms. In addition, two specific use cases of the hybrid mobile computing for CAV applications are investigated: object detection with mobile local computing where only local computing resources are used, and movie recommendation with mobile cloud computing where remote cloud resources are used. For object detection, we focus on the challenges of detecting vehicles, pedestrians and cyclists in driving environment and propose three methods to an existing CNN based object detector. Large detection performance improvement is obtained over the KITTI benchmark test dataset. For movie recommendation we propose two recommendation models based on a general framework of integrating machine learning and collaborative filtering approach. The experiment results on Netflix movie dataset show that our models are very effective for cold start items recommendation.
|
|---|
