We propose a novel edge-based visual–inertial fusion approach to address the problem of tracking aggressivemotionswithreal-timestateestimates.Atthefront-end, our system performs edge alignment, which estimates the relative poses in the distance transform domain with a larger convergence basinandstrongerresistancetochanginglightingconditionsorcameraexposurescomparedtothepopular direct dense tracking. At the back-end, a sliding-window optimization-based framework is applied to fuse visual and inertial measurements. We utilize efficient inertial measurementunit(IMU)preintegrationandtwo-waymarginalization to generate accurate and smooth estimates with limited computational resources. To increase the robustness of our proposed system, we propose to perform an edge alignment self check and IMU-aided external check. Extensive statistical analysis and comparison are presented to verify the performanceofourproposedapproachanditsusabilitywith resource-constrained platforms. Comparing to state-of-theart point feature-based visual–inertial fusion methods, our approach achieves better robustness under extreme motions or low frame rates, at the expense of slightly lower accu racy in general scenarios. We release our implementation as open-source ROS packages.

EB00000003710K

Available

• Edge alignment-based visual–inertial fusion for tracking of aggressive motions