AI and coastal highway monitoring

With the recent federal legislation approval of the funding of the United States’s infrastructure, investment into transportation research has prompted the implementation of new technologies to advance the data collection and evaluation methods for transportation asset management. As departments of transportation (DOTs) upgrade their equipment to utilize these technologies, a higher standard for the quality of highway infrastructure across the states has raised concerns for an improvement in traffic safety and efficient maintenance and rehabilitation of the roadways. This is essential for pavements in locations where surface distresses are developed at a higher rate, such as coastal regions due to rapid development and extreme weather phenomena being more frequent and intensive. State DOTs have adopted the use of automated pavement data collection and condition evaluation. However, the inadequate accuracy of the existing automated technology has led to erroneous distress measurements and inconsistent manual intervention approaches for pavement engineers to assess the surface damage of the pavements after data collection to establish more reliable analyses of the performances. With the recent advancements in artificial intelligence and deep learning, progress towards more accurate and efficient detection algorithms has provided possibilities of better data quality to monitor the surface condition of the pavements. Due to these new technologies, 2D/3D pavement images can be analyzed with improved accuracy for more efficient detection of pavement distresses such as cracks caused by various factors like environment, weather, age, and traffic loading. In this study, 2D/3D images were collected from pavements in the coastal regions in the states of Louisiana, Mississippi, and Texas. The proximity of these states to the Gulf of Mexico and the presence of rapid population growth and economic development in this region, unique distress image data were acquired to train and evaluate models such as convolution neural networks (CNN) and a visual transformer after manual annotations were made. 

This research makes two contributions: the establishment of a novel multi-modal pavement distress dataset utilizing the combination of high-resolution 2D/3D imagery across inland and coastal regions and a systematic evaluation of state-of-the-art detection models. The images consisted of high quality 4096x2048 pixel resolution scans that were 47 feet in the longitudinal direction and 14 feet in the transverse direction on both asphalt and concrete surfaces. Deep learning models such as the You Only Look Once (YOLO) line of detection models and Real-Time Detection Transformer (RT-DETR) were developed by training, validation, and testing on the pavement image data collected in this study to more efficiently evaluate the pavement surface condition of the coastal regions. Rigorous labelling and revisions within the dataset were conducted to further optimize the detection accuracy of models as selection continued. Comparisons between the inland and coastal data in asphalt pavements were conducted to determine model capabilities for expansion as the sections near the Gulf coast had lacked sufficient instances of a few classes. Detection capabilities of the selected model faced many issues that would be reflected in real-world conditions, such as noise, distress severity, distress density, and abnormal distress appearances. These factors would heavily affect the accuracy for each distress, with joints on asphalt surfaces reaching a peak mAP50 of 0.928 while asphalt patches would reach a peak mAP50 of 0.981. Despite retaining lower scores on some classes, this study displays the effectiveness of deep learning detection models on pavement datasets with distresses containing a large variety of distresses and common objects.