Crowdsourcing from the True crowd: Device, vehicle, road-surface and driving independent road profiling from smartphone sensors

Abstract

Existing smartphone-based systems for detection of road surface events such as speed-breakers, potholes, broken road patches, etc. have been developed primarily for the use in four-wheeler vehicles such as cars with perfect driving maneuvers over a road with occasional irregularities. However, our experiments on a 673 km road trail in a suburban city of India, where an overall road condition is poor, show that such event detection accuracy drops to less than 80% for speed-breakers and less than 70% for potholes, when the crowdsourcing data is collected from different vehicles such as two-wheeler (bike or scooty), three-wheeler (auto-rickshaw) or four-wheeler (car) or with different smartphones kept at different positions (in-pocket, in-dashboard, vehicle-mounted).The aim of this work is to develop a system that detects three road events — speed-breakers, potholes and broken road patches, with improved accuracy even under varying conditions and rough roads. The proposed RoadSurP system works in two phases. The first phase runs in a smartphone and identifies the candidate signatures for road anomalies using robust auto-orientation and auto-tune thresholding algorithms to make it almost invariant of position, placement, vehicle, and smartphone type. The second phase works in a server and uses a decision tree based classifier to reduce the false-negative and false-positive instances caused due to the impact of different driving maneuvers, vehicle suspensions, etc. Finally, we apply a k-medoids clustering to geo-localize detected events from multiple trails over a map service. RoadSurP is implemented as an Android application, and tested over a 26 km road using five two-wheeler, seven three-wheeler and three four-wheeler vehicles with six different smartphone types under varying placement and position of the smartphones. After being thoroughly trained, the mean accuracy of RoadSurP is found to be 98% for speed-breakers and 92% for potholes over a smooth road (a road with occasional irregularities) and 92% for speed-breakers and 90% for potholes over the rough road. The developed application can be used as an effective crowdsourcing system for road quality monitoring.