PDF version of this entire document

Introduction

The ubiquity and ever-decreasing cost of mobile/portable devices has gradually increased interest in their usage inside cars. In order to guide cars' preferred routes on the road, programmers are able to harness and truly exploit computer vision methods. It is not clear, however, which ones work best and are also practical to run on mobile hardware with decreased performance capabilities in mind (relative to desktops). This project explores the question by implementing a system which alerts the driver about obstacles on the road, primarily other vehicles¹. These ideas and accompanying code are extensible in the sense that detecting more types of nearby objects is a task largely hinged on additional training of multiple classifiers, necessitating more crude manual work. Figure shows the program running under a physical Android device.

Figure: Top: the Car Navigation program running in idle/inactive mode (as one of several programs, with static preview); Bottom left: the program as positioned in the main Android workspace; Bottom right: the program in the Apps list.

In this document I outline implementation aspects of this project. A comparison of multiple methods should be possible, but it remains beyond the scope of this work due to time constraints. I will begin by presenting the development framework, the scientific methods in brief, and the hardware used in subsequent experiments.