Files
Rixen_14632000_Troch_07252000_2026.pdf
Open access - Adobe PDF
- 45.47 MB
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- This thesis presents the development of a computer vision architecture to automatically detect, track, and recognise individual chimpanzees in video footage. The project was designed to support the doctoral research of Calogero Montedoro, which studies how environmental conditions influence the social development of juvenile chimpanzees. By reducing the amount of manual identification required to process long videos, the goal is to make behavioural studies more scalable and less time-consuming. The proposed system is modular and combines body detection, face detection, multi-object tracking, and identity assignment. It builds on YOLOv8 for body detection, ByteTrack for motion-based tracking, and a ChimpUFE-based face embedder fine-tuned for the study group. In addition, this thesis introduces a three-step annotation workflow and a HOTA-based evaluation pipeline to support both dataset production and reproducible comparison between methods. To improve recognition, a three-tier face dataset was assembled from curated identity images, student annotations, and automatically mined face crops from annotated videos. Building on this model, an auto-labelling pipeline was developed to assign names to tracklets and refine them through clustering and temporal post-processing. Although some limitations remain, especially under long occlusions, dense crossings, and difficult viewing conditions, the resulting pipeline can already be used in a semi-automatic way. Corrected videos and new annotations can in turn be reused to strengthen the training data and further improve the system in future work.