Training a Neural Network for Lane Demarcation Detection in the Infrared Spectrum

Abstract

The retro-reflective characteristics of lane demarcations on roadways can potentially provide robust detection in the infrared spectrum even in poor lighting and weather conditions. This paper explores this idea by training a convolutional neural network using Darknet with YOLO to detect 9 classes of road lines from the Berkeley Deep Drive Dataset (BDD). Although BDD is composed of conventional colour images, they were converted to greyscale prior to training as a solution to the scarcity of datasets in the infrared spectrum. The trained model was evaluated on road scenes acquired by the infrared sensor of an Intel-Realsense camera. From the experimental results, it is concluded that object detection techniques primarily developed for localization and classification of objects in the form of bounding boxes are inherently unsuitable for detecting line shaped objects such roadway lane demarcations. In addition, despite the sub-optimal training and detection approach, the performance showed potential for robust lane detection using infrared images.