Endangered wildlife species, such as rhinos, are under constant threat by illegal poaching actions. Wildlife censuses therefore play a vital role in assessing and monitoring the health of animal populations, their grazing locations, and thus for actions against poaching. Traditionally, such censuses were conducted manually, which is dangerous and expensive to do. In this work, we attempt to solve this problem through automation, using drones and state-of-the-art machine learning and computer vision tools.
With the advent of deep learning, in particular Convolutional Neural Networks (CNNs), machine vision tasks like object detection and classification have seen tremendous progress and rise in various fields, such as person identification or autonomous driving. CNNs are commonly trained and employed on extremely large, annotated datasets that contain natural images of everyday objects. Their theoretical concept and versatility allows us to also use them for detecting objects, such as animals, from a different viewpoint— the aerial perspective. The introduction of drones, or Unmanned Aerial Vehicles (UAVs) have enabled large-scale acquisitions of areas with sub-decimeter resolution, allowing to localise animals, perhaps even classify their species. However, animals are typically a rare sight, which makes this task a challenging needle-in-the-haystack problem.
In this project, the student will implement, train, and evaluate state-of-the-art deep learning-based object detectors (Faster R-CNN, RetinaNet, etc.), applied to mammal detection in UAV images. Particular focus will be laid on allowing the model to recognise animals at different flying altitudes of the UAV, thus making it more versatile. Upon completion, the trained detector will provide not only coordinates of animals (as previously done), but also bounding boxes and possibly species labels, thus pushing the information gain from UAV images even further.
- Familiarise and successfully setup a CNN-based object detector for animal localisation in UAV images
- Investigate different detection paradigms (region proposal network, direct regression) and draw conclusions on their usefulness for the particular task at hand
- Provide a machine-estimated set of animal locations, potentially their species, along with an established accuracy score
- Kellenberger, Benjamin, Diego Marcos, and Devis Tuia. "Detecting mammals in UAV images: Best practices to address a substantially imbalanced dataset with deep learning." Remote sensing of environment 216 (2018): 139-153.
- Ren, Shaoqing, et al. "Faster R-CNN: Towards real-time object detection with region proposal networks." Advances in neural information processing systems. 2015.
- Lin, Tsung-Yi, et al. "Focal loss for dense object detection." Proceedings of the IEEE international conference on computer vision. 2017.
- Programming skills in Python (or high motivation for learning it)
- Some background in statistics and/or machine learning is an asset
Theme(s): Sensing & measuring, Modelling & visualisation