Using Stereo Vision and Transfer Learning to Identify Potholes on the Road
MetadataShow full metadata
Identification of dangerous road distress on road surfaces is essential to many applications such as improved driving comfort, safety, the country’s economy and better traffic efficiency. For these reasons, research around the world has comprehensively explored strategies for identification of road distress. In this study, two different state-of-the-art approaches have been implemented and compared for pothole detection. The first approach, focuses on identification of road distress using stereo vision. Single and multiple 3D frame reconstruction techniques are performed for 3D plane fitting to model the road surface using a digital elevation model. Then a road manifold is constructed and further investigated for the detection of major road distress such as potholes. As potholes might be dry, or snow or water filled, the second approach focuses on identifying potholes under different weather conditions. Transfer learning based techniques using Mask R-CNN and YOLOv2 convolutional neural networks focus on improved pothole identification. The pre-trained convolution layers of Mask RCNN and YOLOv2 are trained to identify common natural features in an image such as edges or corners. This knowledge is adapted and transferred to the task of pothole identification using transfer learning. Potholes are identified with promising accuracy using a transfer learning based model. This study also serves the purpose of providing pointers to different datasets recorded on different days and light conditions.