ROAD CONDITION ASSESSMENT FROM AERIAL IMAGERY USING DEEP LEARNING

Abstract

Terrestrial sensors are commonly used to inspect and document the condition of roads at regular intervals and according to defined rules. For example in Germany, extensive data and information is obtained, which is stored in the Federal Road Information System and made available in particular for deriving necessary decisions. Transverse and longitudinal evenness, for example, are recorded by vehicles using laser techniques. To detect damage to the road surface, images are captured and recorded using area or line scan cameras. All these methods provide very accurate information about the condition of the road, but are time-consuming and costly. Aerial imagery (e.g. multi- or hyperspectral, SAR) provide an additional possibility for the acquisition of the specific parameters describing the condition of roads, yet a direct transfer from objects extractable from aerial imagery to the required objects or parameters, which determine the condition of the road is difficult and in some cases impossible. In this work, we investigate the transferability of objects commonly used for the terrestrial-based assessment of road surfaces to an aerial image-based assessment. In addition, we generated a suitable dataset and developed a deep learning based image segmentation method capable of extracting two relevant road condition parameters from high-resolution multispectral aerial imagery, namely cracks and working seams. The obtained results show that our models are able to extraction these thin features from aerial images, indicating the possibility of using more automated approaches for road surface condition assessment in the future.