Region Proposal Technique for Traffic Light Detection Supplemented by Deep Learning and Virtual Data

Paper #:
  • 2017-01-0104

Published:
  • 2017-03-28
Abstract:
Traffic light detection is critical for safe behavior in a world where technology on vehicles is growing more complex. In this work we outline a deep learning based solution for traffic light detection that leverages virtual data for affordable and efficient supervised learning. Using Unreal Engine, we generated a virtual dataset by moving a virtual camera through a variety of intersection scenes while varying parameters such as lighting, camera position and angle. Using the automatically generated bounding boxes around the illuminated traffic lights themselves, we trained an 8-layer deep neural network (DNN), without pre-training, for classification of traffic light signals (green, amber, red). After training on virtual data, we tested the network on real world data collected from a forward facing camera on a vehicle. Using color space conversion and contour extraction, we identified candidate regions by filtering based on color, shape and size. These candidate regions are fed to the DNN. This combination of deep learning, classic computer vision techniques and virtual data all lead to a high performance on our test set. Furthermore, based on our analysis of the virtual dataset we intelligently generated a second virtual dataset addressing the weakness observed from the first analysis, which improved overall performance of the detector. Our solution has applications for many levels of autonomy, from driver assistance technology (level 2) to fully automated vehicles (level 5).
Access
Now
SAE MOBILUS Subscriber? You may already have access.
Buy
Attention: This item is not yet published. Pre-Order to be notified, via email, when it becomes available.
Select
Price
List
Download
$22.00
Mail
$22.00
Members save up to 36% off list price.
Share
HTML for Linking to Page
Page URL

Related Items

Technical Paper / Journal Article
2003-10-27
Technical Paper / Journal Article
2003-10-27
Technical Paper / Journal Article
2003-10-19
Article
2016-03-01