Community Based Repository for Georeferenced Traffic Signs

Student: Hélder Novais
Supervisor: António Ramires

Abstract

In traffic environments, road signs have a key role to control, warn, and command or prohibit the driver of certain actions. Traffic sign maintenance is essential to prevent negative events. In order for these traffic signs to play the role they were designed for, periodic on-site inspections are essential and followed out to determine if signs are in good condition and visible, both during the day and night. However, periodic inspections are time and cost consuming.

Another issue is related to the drivers’ awareness to the traffic signs on the road. Many factors, both internal and external to the driver, may potentially contribute to him missing a sign. Given the purpose of this dissertation, we will focus primarily on the external factors such as the sign being damaged or occluded, or distractions caused by the many gadgets inside the vehicle. Due to all these extraneous influences, a traffic sign recognition system may help the driver to respect these signs and increase significantly their safety, as well as the others around them.

Some high-end vehicles already have such a warning system, at least for danger signs. However, drivers with these vehicles represent a small fraction of the total driving force. This dissertation aims at bringing such a system to a much broader audience.

Smartphones are one of the most used devices by society today, mostly due to the many functionalities they provide in day to day life and their relative accessible monetary value. The increased computational power and cameras’ quality improvement of these devices over the years make them good candidates to support the access to this kind of technology to all. In other words, smartphones of this day and age have the necessary resources to be used as instruments for sign recognition.

Hence, we propose a dual purpose community based approach. On the one hand, each driver can use his mobile device to detect, recognize and geolocate traffic signs, contributing to the traffic sign central repository. Detection is performed using Cascade Classifiers, while a Convolutional Neural Network supports the recognition phase. The repository, based on the information received from the clients, can be used to provide sign status reports and to enable more direct and timely inspection instead of relying on prescheduled global inspections. On the other hand, drivers would have access to the database of traffic signs, therefore being able to receive real-time notifications regarding traffic signs such as speed limit signs, school proximity, or road construction signs. Hence, allowing the system to perform its function even if the recognition phase is not active when used in a low computational power device.

Thesis Download (PDF)

Paper in ICGI 2017 (Best Paper Award) LINK