Overlapping and Non-overlapping Camera Layouts for Robot Pose Estimation

TL;DR

This paper compares overlapping and non-overlapping camera configurations for robot pose estimation, analyzing their advantages, disadvantages, and performance through experiments using Kalman filtering.

Contribution

It provides a comparative analysis of camera layouts for robot ego-motion estimation, highlighting their respective benefits and challenges with experimental validation.

Findings

Non-overlapping cameras offer larger field of view but face scale ambiguity.

Stereo pairs provide higher accuracy at increased computational cost.

Both approaches are validated with synthetic and real-world experiments.

Abstract

We study the use of overlapping and non-overlapping camera layouts in estimating the ego-motion of a moving robot. To estimate the location and orientation of the robot, we investigate using four cameras as non-overlapping individuals, and as two stereo pairs. The pros and cons of the two approaches are elucidated. The cameras work independently and can have larger field of view in the non-overlapping layout. However, a scale factor ambiguity should be dealt with. On the other hand, stereo systems provide more accuracy but require establishing feature correspondence with more computational demand. For both approaches, the extended Kalman filter is used as a real-time recursive estimator. The approaches studied are verified with synthetic and real experiments alike.