Visual-Inertial RGB-D Mapping for Quadruped Locomotion

Abstract

Quadruped Locomotion can profit from a local map by determining optimal foot placement locations and planning the foot motion knowing the ground height at foot position. Rapid motion of the robot necessites a robust yet accurate method to estimate odometry able to deal with high amounts of motion blur. In this work, the Realsense D435i RGB-D camera equipped with an IMU (both accelerometer and gyroscope) is used in a dense visual-inertial odometry algorithm to infer the motion. The dense odometry algorithm is based on DVO, it uses depth and intensity information for odometry estimation. It is fused in a tightly-coupled manner in a Gauss-Newton optimization scheme with the preintegrated IMU data. The vision sensor and IMU are calibrated using the algorithm of Rehder et al. Dense visual descriptor learning is being evaluated to relax the brightness constancy assumption and to improve the robustness to motion blur. Different network architectures are trained to output features at different scales. Two approaches for training are being evaluated, one using pixel-wise loss functions and the other using two RGB-D images as input which are then aligned using a differentiable implementation of DVO. The effectiveness of the full algorithm is evaluated on the ETH3D benchmark which contains accurately calibrated and synchronized RGB-D and IMU data. It is shown that the algorithm is in principal suitable for the problem of quadruped locomotion. Its accuracy however highly depends on the environment and the depth sensor quality. Thus, several possible future research directions are proposed to better deal with these deficiencies.