Thumbnail Image

Publication or External Link





Animals are remarkable at navigation, even in extreme situations. Through motion perception, animals compute their own movements (egomotion) and find other objects (prey, predator, obstacles) and their motions in the environment. Analogous to animals, artificial systems such as robots also need to know where they are relative to structure and segment obstacles to avoid collisions. Even though substantial progress has been made in the development of artificial visual systems, they still struggle to achieve robust and generalizable solutions. To this end, I propose a bio-inspired framework that narrows the gap between natural and artificial systems.

The standard approaches in robot motion perception seek to reconstruct a three-dimensional model of the scene and then use this model to estimate egomotion and object segmentation. However, the scene reconstruction process is data-heavy and computationally expensive and fails to deal with high-speed and dynamic scenarios. On the contrary, biological visual systems excel in the aforementioned difficult situation by extracting only minimal information sufficient for motion perception tasks. I derive minimalist/purposive ideas from biological processes throughout this thesis and develop mathematical solutions for robot motion perception problems.

In this thesis, I develop a full range of solutions that utilize bio-inspired motion representation and learning approaches for motion perception tasks. Particularly, I focus on egomotion estimation and motion segmentation tasks. I have four main contributions: 1. First, I introduce NFlowNet, a neural network to estimate normal flow (bio-inspired motion filters). Normal flow estimation presents a new avenue for solving egomotion in a robust and qualitative framework. 2. Utilizing normal flow, I propose the DiffPoseNet framework to estimate egomotion by formulating the qualitative constraint in a differentiable optimization layer, which allows for end-to-end learning. 3. Further, utilizing a neuromorphic event camera, a retina-inspired vision sensor, I develop 0-MMS, a model-based optimization approach that employs event spikes to segment the scene into multiple moving parts in high-speed dynamic lighting scenarios. 4. To improve the precision of event-based motion perception across time, I develop SpikeMS, a novel bio-inspired learning approach that fully capitalizes on the rich temporal information in event spikes.