Toward Real-World Implementation of Deep Learning for Smartphone-Crowdsourced Pavement Condition Assessment

In recent years, acrlong MCS has emerged as an effective way of collecting essential information about our urban infrastructure integrity conditions. As such, smartphone sensor data obtained while driving vehicles has been widely investigated for road condition monitoring, hoping it can be used as a cost-effective alternative to conventional methods which uses high-cost system, including inertial profilers. However, currently available smartphone methods require expensive signal processing to address various practical uncertainty issues, such as unknown mechanical characteristics of vehicles, variable driving speed, and sensor location. Hence, only a precisely calibrated setup of the vehicle and smartphone could be used for road condition monitoring under a limited environment. The authors' prior study has developed a deep-learning-based road roughness monitoring method that estimates the international roughness index (IRI) from the anonymous vehicles' vibrations at any driving speeds, measured by smartphones. The feasibility of the proposed method has been numerically validated. Building on prior efforts, this study investigates its full-scale experimental validation under a real-world environment, addressing associated critical challenges in the real-world implementation of the proposed method. A fully convolutional neural network (CNN) architecture, called IRI-Net, is developed in this study. A new data management strategy is proposed to deal with the low-resolution characteristics of crowdsourced smartphone GPS data. A large-scale data set is collected with total 29 different vehicles under a real-world uncertain environment. The performance of the proposed method is validated by comparing to the reference IRI data obtained from a conventional inertial profiler, showing a great step toward real-world practical applications.