Open AccessAbsolute Scale Estimation Approach for Monocular Visual Odometry
Author(s) -
Aldo Diaz,
Paulo Roberto Gardel Kurka
Publication year - 2021
Language(s) - English
Resource type - Conference proceedings
DOI - 10.52591/lxai2021062515
Subject(s) - visual odometry , monocular , artificial intelligence , computer vision , computer science , absolute scale , motion estimation , ground truth , scale (ratio) , metric (unit) , pose , odometry , subpixel rendering , robot , pixel , mobile robot , engineering , operations management , physics , thermodynamics , quantum mechanics
Monocular visual odometry is an effective motion estimation technique that requires to solve for the challenging problem of absolute (metric) scale estimation. Current approaches use information such as the camera height or size of known objects to estimate the scene scale. In this paper, we propose a novel prediction-correction method to estimate the absolute scale of motion using camera height and flat ground assumption. Prediction is provided by a robust relative scale estimation strategy that exploits redundancy in depth information. Correction implements ground patch correlation using subpixel search refinement. The proposed method is tested using the public KITTI benchmark. As result, we derive analytical expressions to determine the absolute scale using a monocular camera. The empirical results shows the effectiveness of the proposed absolute scale estimation strategy in reducing the scale drift in monocular visual odometry.