Deep-Visual-SLAM

This repository implements a deep learning-based Monocular Visual SLAM. The key idea is to integrate a pre-trained depth estimation model and a camera optical flow estimation algorithm to construct the Visual SLAM Front-end. Using self-supervised learning, we enable camera ego-motion estimation, similar to Monodepth2. In the future, once the training results of each module (depth, flow, pose) are stabilized, we aim to integrate SLAM Backend algorithms as well.

Features

• Deep-based Visual SLAM (VO/VIO)
  • Monocular Depth Estimation
  • Optical Flow Estimation
  • Visual Odometry (with inertial sensors)

Monocular Depth Estimation

Train and Evaluate Mono-Depth

Optical Flow Estimation

Future work

Visual Odometry

http://redwood-data.org/indoor_lidar_rgbd/download.html

The focal length is 525 for both axes and the principal point is (319.5, 239.5)