System for Automatic Counting of Atlantic Walruses Using Neural Networks from UAV Aerial Photography Images

• Published in: 2024 26th International Conference on Digital Signal Processing and its Applications (DSPA) pp. 1 - 6
• Main Authors: Efremov, Vlad, Leus, Andrew, Gavrilov, Dmitry, Mangazeev, Daniil, Zuev, Viktor, Kholodnyak, Ivan, Vodichev, Nikita, Parshikov, Mikhail
• Format: Conference Proceeding
• Language: English
• Published: IEEE 27.03.2024
• Subjects: Atlantic walruses; Autonomous aerial vehicles; census; detector; interactive segmentation; Measurement; Neural networks; Prediction algorithms; Signal processing algorithms; Training; Wildlife
• DOI: 10.1109/DSPA60853.2024.10510031

Monitoring of natural areas using unmanned aerial vehicles is increasingly being used to track the state of wildlife. UAVs make it possible to obtain high-resolution aerial photographs, with the help of which reserve employees count individuals in the observed areas. Often, the calculation is carried out manually in conventional graphic programs without the use of automated data processing tools, which greatly slows down the calculation process and depletes cognitive human resources, and also makes it impossible to conduct analysis directly on expeditions. In our work, we propose an approach to automate the process of counting Atlantic walrus subspecies Odobenus rosmarus rosmarus on Matveev Island, Russia, using neural network algorithms. Comparative analysis showed that the best result in terms of the quality of recognition of individuals was achieved by the YOLOv8 architecture with a mAP50 metric of 98.32% on test data. To prepare data for training, a semi-automatic data labeling approach using interactive segmentation was used, which made it possible to significantly speed up and improve the quality of labeling. Based on the selected algorithm, a graphical user interface was developed for users with the ability to automatically count individuals and manually correct neural network predictions in order to simplify user interaction with the program code. The dataset for this article is available at link (https://zenodo.org/records/10803921).