This is the second blog in a series on “Computer Vision” by Joe Duncan, our Data Science Researcher.
In the quest to find opportunities for a computer vision solution, experts from across our business have been brainstorming ideas. We have identified four main categories of features to detect:
- traffic signs e.g. speed limit, regulatory (give way, stop), or advisory signs (curves signs)
- road markings e.g edge or centrelines
- hazards e.g. structures, poles, or ditches
- asset information e.g. road roughness, barrier damage, or missing marker posts
Weighing up the value of each idea against its technical complexity, we decided to pursue the identification of traffic signs on highway corridors. By automating the detection of traffic signs, we aim to:
- reduce the manual workload associated in undertaking signage surveys along a corridor
- improve safety and reduce risks associated with surveying live traffic lanes
- provide a tool to assist for further application
Detecting traffic signs from imagery has been researched heavily over the last decade. One of the key challenges is that traffic signs differ across the world, making effective global solutions difficult to achieve. Thankfully, this challenge does not prevent us from using existing research to aid in our endeavor, and then tailor the architecture to suit our specific application.
The solution consists of two key components: a feature extractor and a classifier. The feature extractor is responsible for detecting the presence of a traffic sign. It does this by automatically understanding distinct elements within the image that are associated with signs. There are many existing models developed for this purpose that we can leverage.
The classifier component uses output from the extractor to determine what category a sign falls into. A classifier can be a simple as determining whether a feature is or is not a traffic sign. We aim to produce a more complex solution that will also classify signs in a manner that identifies their message or purpose, while noting there are hundreds of unique traffic signs along New Zealand roads.
This leads to further questions. How do we group traffic signs together in a meaningful way? Is there value in knowing the difference between the types of animal warning signs or can they be treated equally? These questions among others will inform how we prepare the data. We will cover the data preparation process in more detail in our next blog.
Matthew Noon
Associate Director, Safe System