Sara Zarifi

Advisors

Matthias Zürl (M. Sc.), Philipp Stoll (M. Sc.), Prof. Dr. Björn Eskofier

Duration

09 / 2023 – 02 / 2024

Abstract

To gain a better understanding of wildlife and how to protect it more effectively, one of the main tools is to monitor different kinds of animals and collect detailed information. Technology has an important role in gathering wildlife data since it can reduce the required time and effort of associated tasks compared to manual monitoring processes. These manual processes typically involve direct on-site observation of wildlife subjects in their natural habitats during case studies and manual data recording. Tracking animals’ movements provides helpful information for their behavior analysis. However, in order to ensure accurate tracking of movements, it is essential to differentiate between individuals of a species [1, 2]. If individual animals cannot be distinguished from each other, it can be difficult to analyze their behavior accurately, and there is a risk of double-counting animals. All of these problems can lead to an overestimate of population size and potential inaccuracies in wildlife management decisions. Traditionally, wild animals are monitored by using various types of transmitters attached to them and these conventional approaches have some drawbacks. In addition to the invasive characteristics of these approaches, they also suffer from sensor failure and also they are difficult and expensive to implement on a large scale [3]. Photo trap images are an important tool for wildlife conservation. They can be used to track the movements of animals, identify individual animals, and assess population sizes. However, manually sorting through thousands of photo trap images is a time-consuming and labor-intensive task. In recent years, computer vision and deep learning have been used with increasing success for animal tracking. This has made it possible to automate the sorting and distinguishing of photo trap images, which can save time and resources. In particular, re-identification (re-ID) is a computer vision technique that can be used to match individuals in different images. Reidentification is vital for accurately counting wildlife populations and tracking their movements. It is a non-invasive method that relies on the ability of computer vision algorithms to extract and analyze visual features from images and videos. These features can be used to distinguish between individuals or objects, even if they are not in the same pose or lighting conditions. This makes it possible to track the movements of individual animals over time and to count the number of individuals in a population. Therefore, wildlife re-identification techniques based on the camera trap is the method that has the potential to provide precise data for supporting conservation biology studies [4]. However, wildlife re-identification through this approach has some challenges due to the presence of different viewpoints, illumination changes, or even the change of their appearance over time because of maturation. Additionally, the used data for re-identification of animals in the wild may be limited, noisy and low-resolution. There is a probability of imbalanced data from animals as it is difficult to collect a large and balanced dataset of images and videos of animals in the wild. Another challenge of wildlife re-identification is that animals may be moving quickly which leads to motion blur or being partially obscured [5, 6, 7, 8]. In this thesis, our objective is to perform re-identification for lynx. The dataset we are working with consists of 37 individual lynxes. These raw images have been captured by one or two cameras placed in different territories and spanning several years. We encounter various challenges inherent to wildlife re-identification mentioned above. For instance, there are images that have been captured under varying light conditions which has an adverse impact on image quality, or images that have been affected by motion blur due to the swift movement of animals. These factors further complicate the re-identification process. To overcome these challenges and enhance the accuracy of our results, we will employ techniques such as data preprocessing and data augmentation. These techniques will address the data limitations and noise which are frequently seen in wildlife datasets, and improve the robustness of models to variations of appearance in the images. We will also apply transfer learning to save time and improve the performance of our model by using a pre-trained model of Resnet50, Alexnet, and inceptionV3 on a large dataset, such as ImageNet, and fine-tune its classifier based on our smaller dataset of lynx images. This allows us to transfer the knowledge learned from the large dataset to our smaller dataset, which can help to improve the performance of our model. The training is modified by adding cross-entropy loss and for evaluating the performance of our network, we will calculate accuracy, precision, recall, F1 score, rank-k score and mean average accuracy (mAP). Finally, we will test the ability of our model to correctly re-identify individuals by implementing it on other datasets with similar characteristics, such as the Amur Tiger Re-identification in the Wild (ATRW) dataset [3]. By carefully considering all of the aforementioned factors, we aim to develop a robust and effective re-identification approach for lynx.

References

[1] M. S. Norouzzadeh, A. Nguyen, M. Kosmala, A. Swanson, M. S. Palmer, C. Packer, and J. Clune, “Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning,” Proceedings of the National Academy of Sciences, vol. 115, no. 25, pp. E5716–E5725, 2018.
[2] P. C. Ravoor and T. Sudarshan, “Deep learning methods for multi-species animal reidentification and tracking–a survey,” Computer Science Review, vol. 38, p. 100289, 2020.
[3] S. Li, J. Li, H. Tang, R. Qian, and W. Lin, “Atrw: a benchmark for amur tiger re-identification in the wild,” arXiv preprint arXiv:1906.05586, 2019.
[4] Z. Zheng, Y. Zhao, A. Li, and Q. Yu, “Wild terrestrial animal re-identification based on an improved locally aware transformer with a cross-attention mechanism,” Animals, vol. 12, no. 24, p. 3503, 2022.
[5] S. Schneider, G. W. Taylor, S. Linquist, and S. C. Kremer, “Past, present and future approaches using computer vision for animal re-identification from camera trap data,” Methods in Ecology and Evolution, vol. 10, no. 4, pp. 461–470, 2019.
[6] V. Miele, G. Dussert, B. Spataro, S. Chamaill´e-Jammes, D. Allain´e, and C. Bonenfant, “Revisiting animal photo-identification using deep metric learning and network analysis,” Methods in Ecology and Evolution, vol. 12, no. 5, pp. 863–873, 2021.
[7] M. Zuerl, R. Dirauf, F. Koeferl, N. Steinlein, J. Sueskind, D. Zanca, I. Brehm, L. v. Fersen, and B. Eskofier, “Polarbearvidid: A video-based re-identification benchmark dataset for polar bears,” Animals, vol. 13, no. 5, p. 801, 2023.
[8]M. Ye, J. Shen, G. Lin, T. Xiang, L. Shao, and S. C. Hoi, “Deep learning for person reidentification: A survey and outlook,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021.