Dynamic Wide-Baseline Stereo Vision on a Transformable UAV

Abstract

Stereo vision is used extensively in robotics for applications such as 3D mapping and obstacle detection. It can be employed on Unmanned Aerial Vehicles (UAVs) for these purposes. However, a large baseline is often necessary to obtain depth estimates that are sufficiently distant for a UAV to react to obstacles in time. A wide-baseline stereo vision setup that takes advantage of the full wingspan of a fixed-wing UAV is a natural solution to this problem. However, the dynamic nature of such a system, especially due to wing deformation during flight, can present a challenging problem to solve. Ideally, the stereo baseline should be dynamic to account for these deformations. The work done in this area of robotics can be applied to an even more dynamic scenario: a transformable UAV with the capabilities of both fixed-wing and quadrotor flight. This paper addresses the problem of dynamic wide-baseline stereo vision on a transformable UAV platform. It details the design, implementation, and testing of a system that dynamically adapts to changes in both relative camera orientations and baseline. The system utilizes a model-based kinematics approach, incorporating the known geometry of the UAV and the current state of its servo-actuated joints to estimate the dynamic orientation and baseline of the cameras. This approach is then fused with a visionbased approach that involves estimating the essential matrix from SIFT features for a more robust recalibration approach in real-world scenarios. The ROS 2 middleware facilitates communication between the various software components of this project. All testing was conducted in a Gazebo simulation environment. The results demonstrate the system’s ability to generate disparity maps even with significant changes in the baseline and orientation of the cameras, validating the proposed dynamic stereo vision approach.