Pavement distress instance segmentation using deep neural networks and low-cost sensors

Road degradation and deterioration cause more accidents in the transportation sector. Municipalities worldwide monitor and repair roads to solve this issue. However, this method is expensive. Recent research has focused on cost-effective alternatives, notably deep neural networks (DNNs), which can identify road distress kinds and locations from low-cost camera photographs. The lack of extensive road distress datasets, which DNN models mainly rely on, hinders their road condition monitoring performance. Because photographs typically contain noise that severely degrades DNN models, previous attempts to gather such datasets have failed. Our work presents a unique road distress dataset of 1040 photographs from a low-cost camera, covering six categories. For the training of DNN models, each incidence of distress in the dataset is rigorously annotated with bounding boxes, distress segmentation, and distress kind. We trained two DNN models to predict road distress instance segmentation and carefully compared their accuracy with 92.2% and 88.2% mean average precision for longitudinal and transverse cracks for a mere 17-ms inference time making it suitable for working with low-cost smartphone cameras with the highest frames per second (60 FPS). The distress instance segmentation estimation result is sent to a module that calculates the reliable Pavement Condition Index (PCI) of road health. Our unique approach to real-time online PCI computation uses instance segmentation on smartphone-collected road photographs. This application would access a cloud platform using DNN models. This strategy promises to change road monitoring and maintenance, making transport networks safer and more efficient.