This is a report on Institute of Control Sciences (ICS) team participation in AEROBOT-2021 UAV challenge. It is a contest of autonomous navigation by multirotor unmanned aerial vehicles in an unfamiliar GPS-denied environment. AEROBOT was held for the 3rd time, and ICS team successfully finished the course. We describe the approaches we used to solve competition tasks. In particular, the visual navigation that works in the actual arena is selected. Our flight strategies work reliably even with visual odometry errors. An original black lane segmentation algorithm using exponential context histogram and priority direction estimation is presented. The on-site adaptation of neural network models for detecting objects has been tested in practice.