• Head of Laboratory: Dr. Tamás Zsedrovits
  • Members: Nawar AL-HEMEARY, Lóránt DAUBNER
  • The UAV Vision Lab aims to research topics at the intersection of unmanned aerial vehicles and computer vision. In the case of unmanned aerial vehicles, an unavoidable research question is how to solve the computations needed to solve a given problem with the lowest possible power consumption and preferably with small and affordable tools. One of the main topics is camera-based collision avoidance, where we want to detect and track distant aircraft on one or more camera images. In addition to better detection and tracking, the aim is to have a collision avoidance system with the lowest possible power consumption and size since the calculations must be done onboard the aircraft in real time.
  • Another area of research is ergonomic control of aircraft, where we want to develop control methods that allow aircraft to be piloted in a more natural, faster-learning, or even more accurate way than is currently possible. The third major area is indoor applications, where we will also rely on machine vision to investigate what tasks can be done well without external reference systems, using only sensors on board the aircraft. For instance, in 3D cave mapping, we would like to use cameras, IMU and LiDAR to produce a more accurate 3D map of cave passages than is currently available, thus facilitating the exploration of new paths or helping cavers. In addition, students will develop a simulator for collision avoidance, ergonomic drone control, virtual drone fence, neuromorphic collision avoidance, visual indoor navigation, and a video annotation database.

 

Experiment with a biologically motivated vision system for autonomous obstacle avoidance by drones. The image shows the drone flying along a clear path between display panels in a complex environment. The system consists of a pre-processing based on a mammalian retinal model and a U-net for the vision task. The video is captured and streamed by a DJI Tello drone to a laptop running ROS Ubuntu for drone control and communication, and the neural network is implemented using PyTorch-based implementation on the laptop's dedicated GPU.