Fast region convolutional neural network lane and road defect detection for autonomous vehicle application

Abstract

Road markers are the main landmarks and provide critical information for traffic guidance and safety for drivers. These markers are especially crucial for an autonomous vehicle as this type of vehicle needs to make automated decisions to ensure the safety of other road users. Therefore, for an autonomous vehicle to avoid traffic accidents, these markers should be detected and localized accurately. Additionally, road defects may also present a hazard to autonomous vehicles. This paper proposes a Faster Region Convolutional Neural Network (FRCNN) to detect and localize road markers and potholes on the bird's eye view images of roads. Data were collected using the Point Grey Blackfly camera mounted on the roof of a car. The Inverse Perspective Mapping (IPM) algorithm was used to transform the images into a bird's eye view perspective. The bird's eye view perspective is particularly important as it is easier to detect objects of interest from the top view. The data was then used to train an FRCNN on MATLAB R2018a. The results indicate that the FRCNN was successful in the detection of the objects of interest with some overlapping issues, which will be addressed in future works.