Open Access
Geng, Junyi
Graduate Program:
Aerospace Engineering
Master of Science
Document Type:
Master Thesis
Date of Defense:
July 11, 2016
Committee Members:
  • Jacob Willem Langelaan, Thesis Advisor
  • Joseph Francis Horn, Committee Member
  • George A Lesieutre, Committee Member
  • Small UAV
  • Obstacle avoidance
  • Occupancy gird
  • Quasi-polar frame
  • Monocular vision
  • Stereo vision
The motivation behind the research described in this thesis is simplifying the process of small unmanned aerial vehicles (UAVs) navigation through an obstacle field: typical methods follow a two step process of map generation followed by trajectory computation; the method proposed here directly computes the likelihood that a given path is collision-free. Trajectory generation simply consists of selecting the path that is likely to be free over the longest distance. While the approach described here is independent of the choice of sensor, this thesis focuses on vision- based obstacle avoidance. This thesis proposes a quasi-polar (turn rate-time) local occupancy grid approach for obstacle avoidance. It uses GPS and inertial navigation combined with a vision system to map sensor data directly onto dynamically feasible paths, so that path planning consists simply of selecting the path with lowest likelihood of collision. This approach brings many new challenges besides the usual challenges posed by vision-based obstacle mapping (i.e. noisy vision measurements and highly non-linear governing equations of system lead to significant uncertainties). First, the mapping problem is not performed in the traditional Cartesian or polar coordinate frame. Second, the motion update is difficult since the cells in the quasi-polar map have different shapes and size. The boundary of grid cells can be straight lines or arcs, hence classical image processing methods cannot be used for the motion update. These challenges are addressed by: (1) inverse sensor models that map directly onto the quasi-polar grid; (2) a numerical approach for motion updates. Three exteroceptive sensor models (wide field monocular vision, pushbroom stereo, and pushbroom stereo combined with wide field monocular) are presented in this context. Two path selectors considering the occupied probability differently are created. Simulations of flight through a two dimensional environment consisting of both forest and urban terrain are used to demonstrate the utility of this approach. Preliminary work towards hardware mainly for vision system is presented to show the feasibility of sensor models used in simulation.