Processing Traffic Data Collected by Remote Sensing

Video data are being used more often to study traffic operations. However, extracting vehicle trajectories from video by current methods is a difficult process, typically resulting in many errors. The process requires extensive labor to correct the trajectories manually. This paper proposes a method to process video data from traffic operations. Instead of detecting a vehicle in each picture of the video separately, the video data are transformed so that the trajectories of the vehicles (their position over time) become visible in a single image. In this single image, the trajectories can be found by detecting lines. The difference from other methods is that trajectories rather than vehicles are detected. Trajectory (line) detection is more robust than vehicle (rectangle) detection; with this method, about 95% of the trajectories are detected correctly and, more important, the segments of each trajectory are much longer compared with results from other methods in the literature. Also, the detection is a quick process because only a single image is required to be analyzed. For a data set 5 min long, transforming costs several minutes, and automatically detecting and tracking costs 40 to 50 min per lane. Manual correction is then necessary, which costs about 10 min per lane. In contrast, with a different method the total processing time for analyzing traffic operations costs about 1 week for all lanes together.