Probabilistic 3D Pose Reconstruction of Freely Moving Rats

Arne Monsees1, Kay-Michael Voit1, Edyta Leks2, Klaus Scheffler2, Jakob H. Macke3, Jason N. D. Kerr1

1 Research Center caesar
2 Max Planck Institute for Biological Cybernetics
3 Technical University of Munich

A major goal of neuroscience is to gain a mechanistic understanding of the brain by studying how neural activity gives rise to animal behavior and vice versa. This requires to accurately quantify the behavior of freely moving animals in terms of both, neural and behavioral data. While it is feasible to obtain neural data due to the use of established methods like miniaturized microscopes, measuring animal behavior is a less standardized procedure. Thus, pose reconstruction algorithms that allow for the quantitative analysis of animal behavior by estimating the positions and orientations of individual body parts at any given moment in time are needed.

Here, we combine video data from multiple cameras with high-resolution MRI scans to perform probabilistic pose reconstruction of freely moving rats. We consider anatomical as well as temporal constraints by enforcing realistic joint angle limits and smooth movement trajectories in time, which enable us to estimate animal poses with high spatial and temporal resolution. The estimated poses consist of connected joints that have assigned probability values, indicating how accurately their positions could be reconstructed. When body parts are occluded in the camera views for a long period of time, the low probability values of the corresponding joints allow for detecting poses with high reconstruction uncertainty.

The resulting pose information contains 3D positions and velocities of individual joints as well as inter-skeletal joint angles. We show that these quantities are suited descriptors of behavior as they capture the periodic gait pattern of walking rats.