Estimating curvilinear self-motion from optic flow with a biologically inspired neural system

Area of Science:

• Computational neuroscience
• Robotics
• Computer vision

Background:

• Optic flow is crucial for animals to perceive self-motion and guide movement.
• Existing models focus on linear paths, neglecting complex curvilinear motion crucial for natural navigation.
• Primate visual cortex area MSTd is linked to heading perception, inspiring biologically plausible models.

Purpose of the Study:

• To develop a computational model for accurate self-motion estimation using optic flow.
• To investigate if MSTd-like sensors tuned to diverse optic flow patterns can support linear and curvilinear motion estimation.
• To enable robots to perform robust vision-based self-motion estimation.

Main Methods:

• Developed a computational model with a population of MSTd-like sensors tuned to various optic flow patterns.
• Employed deep learning to decode self-motion parameters from sensor signals.
• Utilized synthetic and naturalistic videos of simulated self-motion for testing.

Main Results:

• Accurate estimation of curvilinear path curvature and direction (clockwise/counterclockwise).
• Precise estimation of gaze direction relative to the path tangent.
• Stable estimates over time with rapid adaptation to dynamic motion changes.

Conclusions:

• A coupled system of biologically inspired and artificial neural networks shows promise for robust robot self-motion estimation.
• The model successfully extends self-motion estimation beyond linear paths to naturalistic curvilinear trajectories.
• This approach offers a pathway for enhancing robotic navigation capabilities using visual cues.