a hybrid approach to condition assessment of paved roads
abstract
road reconstruction is drastically more expensive than an appropriate maintenance program.
performing targeted maintenance requires a comprehensive inventory of road conditions. to create
and maintain a road condition inventory, many cities use expensive, purpose-built road profiling
vehicles. with the abundance of smartphones, and the variety of sensors they hold, there is an
opportunity to collect road condition data more pervasively and at a lower cost. this study shows
that smartphones can be used to collect road surface data for the purpose of creating an inventory
of road conditions. the data sources within the phone are its onboard accelerometer and camera.
readings from these sensors are taken simultaneously while the phone is mounted to the
windshield of a vehicle driving over paved city streets. the collected data was used for measuring
road surface roughness and automatically identifying specific defects, such as cracks and potholes.
road roughness is quantified using the range between acceleration values and pairing them with
readings from the smartphone’s gps antenna to create roughness intensity maps. defect
identification is done by a convolutional neural network that can label the pixels of an image it
receives into one of several defect pixel label classes. the neural networks used for this study were
trained using a manually labelled image set of several hundred images. the novel hybrid approach
used in this study combines the accelerometer data with the camera footage. the data are
synchronized so that frames can be extracted from the video footage based on anomalies in the
acceleration data. peaks in the acceleration data are used as a trigger to extract and analyze images
from the video footage. this approach to defect identification can be done at a much lower
computational burden than if all the video footage was processed by a convolutional neural
network.
the defect identification results show strong performance in terms of identifying locations of
defects, as determined by the acceleration data, and identifying the defects at the identified
locations themselves. our study found that an unmodified vehicle with a smartphone mounted to
its windshield can generate data that our system, developed in matlab, can find over 60% of
road defects. this is achieved in just one pass over the road surface.