FarSight

Abstract

Maintaining an adequate separation from the vehicle in front is key to safe driving. While LIDAR and RADAR sensors could be used for ranging, cost considerations and the huge installed base of vehicles that lack these sensors, especially in developing regions, call for a low-cost yet robust alternative. To this end, we present FarSight, a system that performs vehicle ranging using a smartphone mounted on the windshield or the dashboard. FarSight uses the smartphone's camera to identify and draw a bounding box around vehicles in front, based on which ranging is performed. Unlike prior smartphone-based work, FarSight does not depend on any infrastructure support such as standard-width lane markers and works with a heterogeneous mix of vehicles, both of which are characteristics of developing regions. We develop a novel hybrid approach for vehicle detection and tracking, which balances accuracy and speed by combining deep neural network based vehicle detection with vision-based object tracking in a pipelined manner. We also devise data augmentation techniques to improve the effectiveness of vehicle detection, thereby increasing the ranging distance. We show that FarSight can range accurately in both daytime and nighttime conditions and up to distances of 90 m. We have implemented FarSight as an Android-app and tested it across various phones. Further, we present two ranging-based applications built on FarSight.