Estimation of a Road Mask to Improve Vehicle Detection and Tracking in Airborne Imagery

Abstract

This research proposes a method to estimate a road mask in airborne imagery to improve vehicle detection and tracking. The road mask can remove false detections and reduce computation time. The road mask was estimated on the basis of smoothed (interpolated) positions of the vehicles tracked by the short-term and long-term trackers in the software tool “tracking and registration of airborne video image sequences” (TRAVIS) in a fixed number of frames (i.e., one cycle). First, raw vehicle positions that were obtained from TRAVIS were smoothed with a Kalman filter. Second, the slope of the roadway was estimated based on the smoothed positions of these vehicles. Third, falsely detected vehicles were removed if the slope of the lines that connected two vehicles in the same lane were close to the slope of the roadway. Finally, the road mask was generated according to the minimal and maximal intercepts of the approximated roadway edges in one cycle. Experimental results showed that the estimation of the road mask was accurate and that the improvement in the correct detection and tracking rate and in the computation time was significant.