Machine vision based waterlogged area detection for gravel road condition monitoring
Michael Starke, Chris Geiger
Abstract
When assessing forest road conditions, information about waterlogged areas on gravel roads brings high practical value when used as an indicator for road wear. Around these perimeters, lowered binding forces of the construction material reduce the stability of the road, which induces accelerated road damage. When a road is actively used to access a logging site under humid weather or thawing conditions, road wear can build up fast and make further use of the road critical. In this study, a deep learning algorithm was trained to test the detection of a combined observation of waterlogged appearances on forest roads from video and image data, collected from a passing vehicle’s perspective. The training of a YOLO v5s model achieved an F1-score of 0.59 and shows the applicability of this approach with high confidence of detection. Evaluating further training characteristics such as precision, recall, and the object size-related detection confidence reveals challenges for a successful application in terms of undetected objects, variation of objects in the training step, the required amount of training data and the object distance focused.