Intelligent Road Surface Monitoring Based on Optical and Meteorological Measurements
Monitoring of road surface conditions plays an essential role in driving safety and road maintenance. Currently, surface condition monitoring is often performed by weather information or maintenance personnel recordings, which are limited in timeliness, objectivity, and require human interpretation. In this work, we propose an economical solution for intelligent monitoring of surface conditions based on road cameras and meteorological measurements. The number of ordinary road surface conditions is often limited, e.g., dry, wet, or snowy. Thus, surface condition monitoring can be formulated as a supervised machine learning problem. A system learns the representation of different surface conditions from training data using machine learning models, and then applies the learned knowledge to new surface data to predict the conditions. After watching approximately 50,000 images across different seasons, we found that most surface conditions can be categorized into four conditions: dry, wet, full snowy and partial snowy. Consequently, 9,349 road images and associated meteorological parameters have been collected and served as training data. Feature representation is an essential part in machine learning. In this work, a number of image features together with meteorological parameters were first derived. A feature selection approach was then proposed using F-Score and PCA. Through the selection approach, four groups of image features were identified to be highly relevant with surface conditions, including luminance/chrominance features, gradient features, edge features, and texture features. We also found that five meteorological parameters: road surface temperature, air temperature, dew point, relative humidity, and precipitation, have shown strong discriminability in surface conditions. Furthermore, surface conditions have temporal connection, e.g., current condition presents indication for succeeding condition. Therefore, the condition changes are also included. Then, an SVM (support vector machine) model has been proposed to perform the machine learning problem. The above SVM approach only classifies surface conditions within the four predefined categories. Another crucial road condition is icy, especially invisible one (e.g., black ice). In this case, optical images become unimportant, while the meteorological parameters are emphasized. To handle such situation, another icy detection algorithm has been proposed based on the meteorological parameters. This algorithm mimics the ice formation process based on meteorological parameters: surface temperature, air temperature, wind speed, dew point, relative humidity, precipitation, road and air temperature changes. Four types of icy conditions are detected: frost, packed snow, black ice, and freezing rain. The SVM model is also employed in icy detection, e.g., a snowy condition given by the SVM model can differentiate a packed snow condition from others. Simulative experiments have been conducted to evaluate the performance of the proposed system. In total 6,000 images randomly captured across different seasons in 2016 have been collected together with meteorological parameters. Then the system was employed to detect the surface conditions in these images, and compared with a manual classification performed by maintenance experts. The comparison result has demonstrated promising accuracy (87.4%) of the system, which confirms that using optical and meteorological sensors can be an accurate, economical solution for monitoring road surface conditions. The developed system is accessible online (http://insitu.cmr.no/roadweatheranalysis/).