Real-time Quadrotor Navigation Through Planning in Depth Space in Unstructured Environments

TL;DR

This paper presents a real-time, vision-based quadrotor navigation method that plans trajectories in depth image space using onboard stereo camera data, enabling rapid obstacle avoidance in unstructured environments.

Contribution

It introduces a novel depth image space planning approach with collision prediction and a switching strategy for obstacle avoidance, validated through simulations and hardware tests.

Findings

Effective real-time obstacle avoidance demonstrated in simulations.

Successful hardware validation of the depth space planning approach.

Improved trajectory safety and efficiency in unstructured environments.

Abstract

This paper addresses the problem of real-time vision-based autonomous obstacle avoidance in unstructured environments for quadrotor UAVs. We assume that our UAV is equipped with a forward facing stereo camera as the only sensor to perceive the world around it. Moreover, all the computations are performed onboard. Feasible trajectory generation in this kind of problems requires rapid collision checks along with efficient planning algorithms. We propose a trajectory generation approach in the depth image space, which refers to the environment information as depicted by the depth images. In order to predict the collision in a look ahead robot trajectory, we create depth images from the sequence of robot poses along the path. We compare these images with the depth images of the actual world sensed through the forward facing stereo camera. We aim at generating fuel optimal trajectories inside the depth image space. In case of a predicted collision, a switching strategy is used to aggressively deviate the quadrotor away from the obstacle. For this purpose we use two closed loop motion primitives based on Linear Quadratic Regulator (LQR) objective functions. The proposed approach is validated through simulation and hardware experiments.