Continuous and Realtime Road Condition Assessment Using Deep Learning

Abstract

Continuous and real-time monitoring of road quality conditions is essential for the maintenance of roads and to ensure the safety of drivers and their vehicles. However, the continuous monitoring of thousands of kilometers of roads and highways is a very tedious, time-consuming, error-prone, and expensive operation. A deep learning based approach that can automatically classify the road condition can help tremendously in cutting down the time, effort, accuracy, and cost for monitoring and maintenance of vast road infrastructure. This paper proposes a mechanism to continuously monitor deteriorating road conditions at the city or municipality level in real time and classify them into four different categories (good, medium, bad and unpaved) using custom-built and transfer learning from pre-trained deep learning models (VGG16 and MobileNetV2). The dataset is collected from different roads in the Kingdom of Saudi Arabia. The dataset is composed of close-up road images taken in real time (while driving the car) at regular intervals using an Android App. In the data capture model, the Android App helps to easily tag (label) the captured images for model training purposes. In the classifier mode, the Android app uses the developed deep learning model to classify the captured image and then transmits the medium, bad or unpaved road condition to the central server along with longitude and latitude information to update the centralized map of the city (or municipality). The proposed approach provides an accuracy of 98.6 % to classify the road condition based on images captured during real time driving of the vehicle.