Towards Dynamic Monocular Visual Odometry Based on an Event Camera and IMU Sensor

Abstract

Visual odometry (VO) and visual simultaneous localization and mapping (V-SLAM) have gained a lot of attention in the field of autonomous robots due to the high amount of information per unit cost vision sensors can provide. One main problem in VO techniques is the high amount of data that a pixelated image has, affecting negatively the overall performance of such techniques. An event-based camera, as an alternative to a normal frame-based camera, is a prominent candidate to solve this problem by considering only pixel changes in consecutive events that can be observed with high time resolution. However, processing the event data that is captured by event-based cameras requires specific algorithms to extract and track features applicable for odometry. We propose a novel approach to process the data of an event-based camera and use it for odometry. It is a hybrid method that combines the abilities of event-based and frame-based cameras to reach a near-optimal solution for VO. Our approach can be split into two main contributions that are (1) using information theory and non-euclidean geometry to estimate the number of events that should be processed for efficient odometry and (2) using a normal pixelated frame to determine the location of features in an event-based camera. The obtained experimental results show that our proposed technique can significantly increase performance while keeping the accuracy of pose estimation in an acceptable range.