Distributed Variable-Baseline Stereo SLAM from two UAVs

Abstract

Visual-Inertial Odometry (VIO) has been widely used and researched to control and aid the automation of navigation of robots especially in the absence of absolute position measurements, such as GPS. However, when observable landmarks in the scene lie far away from the robot's sensor suite, as it is the case at high altitude flights, the fidelity of estimates and the observability of the metric scale degrades greatly for these methods. Aiming to tackle this issue, in this article, we employ two Unmanned Aerial Vehicles (UAVs) equipped with one monocular camera and one Inertial Measurement Unit (IMU) each, to exploit their view overlap and relative distance measurements between them using Ultra-Wideband (UWB) modules onboard to enable collaborative VIO. In particular, we propose a novel, distributed fusion scheme enabling the formation of a virtual stereo camera rig with adjustable baseline from the two UAVs. In order to control the UAV agents autonomously, we propose a decentralized collaborative estimation scheme, where each agent hold its own local map, achieving an average pose estimation latency of 11ms, while ensuring consistency of the agents' estimates via consensus based optimization. Following a thorough evaluation on photorealistic simulations, we demonstrate the effectiveness of the approach at high altitude flights of up to 160m, going significantly beyond the capabilities of state-of-theart VIO methods.Finally, we show the advantage of actively adjusting the baseline on-the-fly over a fixed, target baseline, reducing the error in our experiments by a factor of two.

VIRAL SLAM: Tightly Coupled Camera-IMU-UWB-Lidar SLAM

“…For e.g. it has been integrated with monocular VIO for drift correction [5], [6], or can be used as a variable baseline for cameras on different UAVs [16]. In recent years, VIO, UWB and lidar have also been used in a loosely coupled manner for relative localization [17]- [22].…”

“…where ρ H (•) and ρ A (•) are the Huber and arctan loss functions used to reduce the effects of outliers, N m L ∈ N is the number of FMM coefficients extracted from the CFC F m , N k U ∈ N is the number of UWB samples obtained in the interval (t m−1 , t m ], N k V ∈ N is the number of visual features that are tracked on the sliding window from t w to t k , and C i refers to the set of cameras that observe the visual feature f i , excluding C a , λi can be either the state estimate λi or the marginalized inverse depth λi of the MMM features. The cost function (16) summarizes the coupling of each sensor's factor with the state estimate.…”

“…where c 1 and c 2 > 1 are some user-defined parameters, then we can start adding the range factors r 3 (•) to (18). After this BA result, we can obtain the estimates of L W T and b r which can be used for the fusion of UWB factors in (16).…”