Modelling edge computation offloading for automotive video analytics

Abstract

Intelligent vehicles are becoming more common and affordable, and with each new model come complex and resource-intensive applications, starting at simple sensors, into assistant AI, and more recently full vehicle automation. These applications can be mostly segmented into two categories, infotainment and driving assistance. The latter category requires strict adherence to time limits, lest they become useless or even dangerous to the driver, and is the focus of the present work. The spread and availability of powerful computation devices throughout city streets as a result of a variety of factors, including the emergence of new technologies that demand higher node density like the fifth generation of mobile networks, raises the question as to whether there is an effective way that vehicles can take advantage of this spread out capacity, instead of depending solely on the conventional on board computing unit (OBU). Furthermore, given the complexity of these systems, how can one model them and perform simulations that are both valid and credible, as well as liable to verification through real world experiments. According to IEEExplore, even though the study of vehicular ad-hoc networks (VANET) goes all the way back to 2005, computation offloading within VANETs is a much more recent focus of general study (~2017). Nevertheless, dozens of different approaches with respective algorithms have been proposed. In terms of communication, both vehicle-to-infrastructure (V2I) and vehicle-to-vehicle (V2V) communication have been considered, with the technologies in use ranging from mobile and wifi, to VANET specific such as Dedicated Short Range Communications (DSRC / 802.11p). The most differentiating factor is the chosen parameters for the algorithms, that can be categorized in communication (available/used bandwidth), load (size, delay requirements), computation (required CPU cycles) and car movement (cell stay time). The main goals of this work are twofold. First, to provide a realistic and verifiable simulation environment with mathematical models for the load (based on a real video stream) and for the computation (based on a simple object detection engine). Second, to provide a simple proof-of-concept computation offloading algorithm that takes advantage of the information in the models to perform sensible offloading decisions.