Fast Lane Filtering for Autonomous Vehicle

Lane filtering is a necessary process applied after lane detection in autonomous vehicle applications. The unprocessed result from lane detection can usually be noisy. The length and position of the detected lanes are often changing abruptly across frames due to imperfect detection, which would introduce noise to downstream processes. In order to obtain steady lane detection result, we develop a new method to filter the raw output of detection. We first perform a prepossessing to filter out large obvious inconsistency. A compact lane representation is designed, to convert the various length into fixed-dimension vector representation. The general shape of the lanes is kept while a low computational complexity is maintained. We then apply the Kalman filter to perform filtering in temporal domain, and estimate the location of the lanes. Qualitative and quantitative experiments are conducted on real data collected from vehicle driving in urban area, showing improved results compared to unprocessed lane detection results.