Using color profiles for street detection in low-altitude UAV video
Joshua Candamo, Rangachar Kasturi, Dmitry B. Goldgof · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2009
This paper describes a vision-based street detection algorithm to be used by small autonomous aircraft in low-altitude urban surveillance. The algorithm uses Bayesian analysis to differentiate between street and background pixels. The color profile of edges on the detected street is used to represent objects with respect to their surroundings. These color profiles are used to improve street detection over time. Pixels that do not likely originate from the "true" street are excluded from the recurring Bayesian estimation in the video. Results are presented comparing to a previously published Unmanned Aerial Vehicle (UAV) road detection algorithm. Robust performance is demonstrated with urban surveillance scenes including UAV surveillance, police chases from helicopters, and traffic monitoring. The proposed method is shown to be robust to data uncertainty and has low sensitivity to the training dataset. Performance is computed using a challenging multi-site dataset that includes compression artifacts, poor resolution, and large variation of scene complexity.