| dc.description.abstract | In Norway, approximately two million sheep and lambs are released on open ranges to graze each summer. Locating the sheep after each grazing period is a tedious and challenging process. An experiment conducted by IDI at Norwegian University of Science and Technology (NTNU) shows that thermal imaging and Unmanned Aerial Vehicle (UAV) can be an effective tool for locating sheep.
This thesis covers two modules. The first module is to develop an algorithm used for generating coverage paths for acquiring forward-looking infrared (FLIR) images. The algorithm is capable of generating efficient routes for any user-defined area, with the possibility of omitting lakes to reduce flight time. Using a UAV with an attached \acrshort{flir} camera and an onboard auto flight system, images can be acquired autonomously using the generated route.
The second module is to develop and test different image processing methods for detecting sheep in FLIR images. The primary focus was on developing an algorithm using classic image processing. This method was evaluated against two machine learning methods: Convolutional Neural Network and RetinaNet. The results showed that classic image processing produced a better result than the state of the art machine learning methods. The algorithm was able to detect 83.3% of the sheep in the test set, which included images with a severe amount of background clutter. In areas with dense forest canopy, the algorithm was able to detect most animals with a low rate of false positives. It was found that the available dataset was too sparse to be used for machine learning based object detection. | |
