Monocular Event Camera Odometry Using Deep Learning

Abstract

Event cameras, with their high temporal resolution, low latency, and dynamic range, provide a promising alternative to frame-based systems. However, they face unique challenges in performing odometry due to their sparse and asynchronous data output. This thesis addresses these challenges by presenting a novel monocular event camera odometry framework that leverages deep learning techniques. The proposed approach, TartanEVO, integrates optical flow prediction from an event camera with a pose estimation network to produce incremental motion estimates. This thesis also introduces the Tartanair-v2 - Event Camera, the largest public event camera dataset for odometry and SLAM; this dataset features large and diverse scenes with challenging viewpoints, varying lighting, and diverse motion patterns. Extensive evaluations demonstrate the robustness of TartanEVO in diverse environments where traditional odometry algorithms fail. In low lighting and rapid motion, our method even outperforms frame-based algorithms. This work highlights the potential of event cameras and deep learning in advancing robust odometry systems.