Water-Puddle Segmentation Using Deep Learning in Unstructured Environments

Unstructured environments have many anomalies on the road such as potholes, water puddles, etc., that need to be detected for safe driving. The detection of water-puddles bear a major challenge due to the dynamic and unstructured nature of the puddles. The complexity of the problem of detection and segmentation of the water-puddle is high since it can have any shape and size that gets dynamically affected by its surroundings and weather conditions. In this paper we propose a novel frame-work for detection and segmentation of water-puddles on road in an unstructured environment from a dash camera feed. We also create a novel dataset using a camera placed on the dashboard of a car. These videos are recorded as the car moves in the unstructured environment. We have annotated the dataset and also augmented the dataset using various augmentation techniques. We use deep learning for detection and segmentation of water-puddles under various weather condition and time of day in an unstructured environment. Our experimental results show that we are able to achieve accuracy of 63% at mAP50 and the system works well in various conditions.