Abstract
Video-log images are often used by state DOTs to manually or automatically extract roadway infrastructure information, including roadway geometry and signs. Poor quality images with lens debris are unacceptable and need to be identified before state DOTs accept the collections of video-log images from contractors. However, manually reviewing the tens of millions of video-log images to detect lens debris deficiencies is labor-intensive and time-consuming. Therefore, automatic lens debris detection in video-log images is needed. Based on joint domain-range representation of lens debris candidates that are obtained from dark channel prior depth map estimation and Canny lens debris edge detection, a nonparametric Kernel Density Estimator (KDE) model for lens debris detection has been developed for the first time. The model detects the lens debris areas in the video-log images using recursive bandwidth selection and hysteresis updating strategy. An experimental test, using 13,007 video-log images provided by the Alberta DOT, was conducted to validate the proposed algorithm. Test results show that the proposed algorithm can detect lens debris in video-log images with a detection rate greater than 84%. The proposed algorithm is promising for improving video-log image data quality control and assurance.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call