Location

Virtual

Start Date

5-10-2021 11:10 AM

End Date

5-10-2021 11:20 AM

Description

We present a method for creating 3D obstacle maps in real-time using only a monocular camera and an inertial measurement unit (IMU). We track a large amount of sparse features in the image frame. Then, given scale-accurate pose estimates from a front-end visual-inertial odometry (VIO) algorithm, we estimate the inverse depth to each of the tracked features using a keyframe-based feature-only bundle adjustment. These features are then accumulated within a probabilistic robocentric 3D voxel map that rolls as the camera moves. This local rolling voxel map provides a simple scene representation within which obstacle avoidance planning can easily be done. Our system is capable of running at camera frame rate on a laptop CPU.

Available for download on Tuesday, May 10, 2022

Share

COinS
 
May 10th, 11:10 AM May 10th, 11:20 AM

Sparse Monocular Scene Reconstruction Using Rolling Voxel Maps

Virtual

We present a method for creating 3D obstacle maps in real-time using only a monocular camera and an inertial measurement unit (IMU). We track a large amount of sparse features in the image frame. Then, given scale-accurate pose estimates from a front-end visual-inertial odometry (VIO) algorithm, we estimate the inverse depth to each of the tracked features using a keyframe-based feature-only bundle adjustment. These features are then accumulated within a probabilistic robocentric 3D voxel map that rolls as the camera moves. This local rolling voxel map provides a simple scene representation within which obstacle avoidance planning can easily be done. Our system is capable of running at camera frame rate on a laptop CPU.