Multiple flying bird and bird keypoint detection toolbox for processing bird videos

Details

Supervisors

Ronsse, Renaud

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

Multiple Bird detection is challenging due to their diversity, their sizes and their overlaps in flock, but is essential to derive information on birds in many application fields. A multiple bird detection toolbox is proposed with several deep learning models with different accuracies and speeds, trained on a bird dataset. A new dataset, constructed by adding non-bird images, is investigated to increase the robustness of the models on non-bird objects. In addition to that, an algorithm to improve the bird detection on large images, is tested. Finally, a very simplified keypoint detection method is explained. The detection results show that some deep learning models have very good performance and are able to detect slightly overlapping birds. Some models obtain good precision while being fast which is a real advantage on large images. Adding the non-bird dataset reduces false detections and gives promising detection results.