Fast pasture growth makes pastoral farming low-cost, sustainable and efficient for New Zealand. However, in recent years, sheep farming has declined due to the extra labor needed in handling these farms. One of the main tasks that need manual involvement is farm animal counting. Currently, most of the stock is counted manually with a gap of weeks, usually during a major stock movement or other events. The conventional method is time-consuming, tiring, and prone to errors. Also, this inability to constantly monitor stock numbers gives farm rustlers plenty of time to steal animals, hence causing a significant financial loss annually. Farmers desire to minimize their losses by having a daily count of their stock with the highest possible accuracy.

The research work presented in this thesis focused on solving livestock (sheep) counting problems using an unmanned aerial vehicle (UAV) and deep learning. It provides a solution to detect, count, and track livestock in a paddock using advanced methods, and gives extremely accurate information to farmers in a minimum time. The proposed system uses a UAV for counting farm stock while observing the code of conduct for UAV usage. It involves no disruption to the animals as a UAV will hardly be noticeable from the allowable height. Different deep learning algorithms were investigated in this regard and recorded videos were processed to provide the estimated sheep count in the UAV images. The next step of the system is livestock tracking, meaning that each sheep in the full paddock video should be assigned a virtual identification number to keep a track of movement in the video and to count the herd size correctly. Fence detection was one of the main steps to avoid detecting sheep in nearby paddocks.

Stock counting using a UAV is a promising research area but offers various technical challenges, such as non-uniform illumination in images, occlusions, object scaling, rotation, noise, and the challenges in identifying objects from various visual perspectives. Previous studies have shown that there is a lot of potential in this field as not much work has been done so far.

For this research, a full dataset was created from different paddocks to make the proposed research work scalable to accommodate a variety of backgrounds and perimeters. The discussed network and the proposed method gave promising results and is crucial step forward towards a fully-automated livestock tracking system in sheep farms.