University of Twente Student Theses

Login

Feasibility of real-time UAV flight path planning for urban monitoring

Hu, Shichen (2021) Feasibility of real-time UAV flight path planning for urban monitoring.

Link to full-text:https://ezproxy2.utwente.nl/login?url=https://library.itc.utwente.nl/login/2021/msc/gfm/hu.pdf
(only accessible for UT students and staff)
Abstract:The unmanned aerial vehicles (UAVs) have become popular in many research fields in recent years. The UAV has the advantage of small size, simple structure, high mobility, and low cost. The flexibility of UAVs provides a new possibility of environment exploration, especially for those dangerous places where are difficult to reach. However, UAVs also have limitations due to their small size. The small size means limited payload and flight time. For applications that the target of interest is not the whole scene but specific objects within the scene, the traditional way is to fly twice. Use the first flight to obtain the location of the target, then fly another mission. However, the result of UAV applications also depends on environmental conditions. For instance, wind speed and illumination conditions also play a vital role in the data acquisition process. It is hard to keep the environmental factor being consistent between two flights. This research proposed an updatable flight path according to the target of interest in one flight. The algorithm was divided into three parts: global planner, local planner, and object detector. The global planner is a pre-defined strip flight according to the simulation scene. The local planner is a circular flight path around the target with a lower flight height. The trigger between these two planners is an object detector. This object detector is a YOLO v3 detector with a Squeeze net backbone. The detector was first trained on the Aeroscape dataset and then performed a transfer learning process on the synthetic image dataset created manually. The whole algorithm was developed and tested using the MATLAB and Simulink platform. The simulation environment was established using Unreal Engine from Epic Games. The final output of this model is nadir view images along the whole flight path with object detection results added. The training result on Aeroscape dataset achieved an average precision of 90%, and after the transfer learning process, the average precision increased to 100%. The object detector performs better at higher flight height in the simulation environment. The entire model runs stable in the Simulink and can have a target inspection from all directions.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:https://purl.utwente.nl/essays/88645
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page