Download files
Abstract
City streets are full of useful information pertaining to the environment, traffic, commerce and crime, which can benefit the public. Unfortunately, even in today's highly advanced society, most street level information that we desire is still collected manually. Recent advances in mobile phone technology makes it feasible for ordinary users to sense and contribute their ambient information with their smartphones. A new paradigm of PS has emerged, driven by the ubiquitous presence of smartphones. An important form of PS is image sensing, as ``a picture is worth a thousand words''.
In this thesis we design and implement a PS system called PetrolWatch that allows volunteers to automatically collect, contribute and share fuel price information. The fuel price is automatically collected from road-side price board images captured by a vehicle mounted smartphone. By leveraging a variety of embedded phone sensors such as GPS receiver and publicly available information in the form of GIS database, our system automatically captures a sequence of fuel price board images when it approaches a service station. These images are then transported to a central server where computer vision algorithms are implemented for board detection and fuel price extraction.
Initial road tests show challenges in image capturing and processing. To overcome these challenges, we develop an automatic image capturing system, including a camera control scheme and an image pre-selection algorithm, to capture high quality images from a moving vehicle. We further design sophisticated computer vision algorithms to extract price data from the collected images. Our global colour classification algorithm, based on machine learning techniques, significantly improves the board detection and fuel price extraction rates.
We develop a PetrolWatch prototype on a Nokia N95 mobile phone. Extensive driving experiments were conducted in suburban Sydney to validate and optimize our work. Experimental results show that our design achieves a board detection rate of 86%, a price character classification rate of 85%, and a low false positive board detection rate of 10%. The final price extraction rate is 73.1% which combines the board detection rate and price character classification rate.