WildPose: Long-Range 3D Motion Tracking of Cheetahs in the Wild Using Multi-Sensor Fusion

Thesis / Dissertation


Permanent link to this Item
Journal Title
Link to Journal
Journal ISSN
Volume Title
In many fields of research, it is often desirable to estimate the 3D pose of a subject - human, animal, or otherwise. Methods for obtaining accurate 3D pose data of a subject are broad in their applications; they inform the design of bio-mimetic robots, they aid greatly in bio-mechanical research, and they are used commonly in the study of animal neuroscience. Currently, robust methods for long-range tracking of subjects in the wild are few and far between, given the rarity of specific training data as well as the generally challenging nature of the associated computer vision problems. This thesis describes the design, implementation, and testing of both a hardware and software component to a method for the 3D motion capture of cheetahs in the wild, dubbed WildPose. The method makes use of multi-sensor fusion including lidar, RGB and IMU sensors to compensate for situations where pure vision-based techniques perform inadequately. To increase robustness, the software design makes use of previously successful trajectory optimisation techniques to yield accurate pose data in adverse conditions that would otherwise be extremely difficult to obtain. The method is extendable to other species with minimal variations. We demonstrate the method's efficacy through experimental validation on challenging cheetah locomotion datasets collected in the wild, presenting both qualitative and quantitative analyses for varied movements, environments, and lighting conditions. Through the shown effectiveness of these techniques in our specific use case, we aim to prove that the methods used perform admirably even under the trickiest of reconstruction problems. Thereby, we present our findings on cheetahs as a promising blueprint for the 3D motion capture of other species.