detect-slam icon indicating copy to clipboard operation
detect-slam copied to clipboard

reimplementation of paper Detect-SLAM: Making Object Detection and SLAM Mutually Beneficial

NOTE

  1. The result of this repo is not reliable。
  2. I just modify Tracking.cc and Frame.cc, and I can not make sure my code is correct, I upload this project just for course project.
  3. I can not provide help if you want to reimplement detect-slam because I am not doing job about slam now. SORRY FOR THAT.

What it is

This a reimplementation of the paper: Detect-SLAM Making Object Detection and SLAM Mutually Beneficial.

How it works

We have only reimplemented the "moving object removal" part due the time constraints.

We use Yolo3 pretrained on the dataset VOC and COCO to detect the moving person on each frame(image) from dataset fr3/walking_xyz and fr3/walking_halfsphere on TUM webpage.

We modified the tracking thread of ORB-SLAM 2 to do Moving Probability Updating and Moving Probability Propagation. And then move the feature points of moving object before pose estimation.

What we got

We did experiments on fr3/walking_xyz and fr3/walking_halfsphere. Below are the estimated trajectory.

fr3/walking_xyz

fr3/walking_halfsphere

For comparasion, below are the results of original Orb-SLAM 2.

fr3/walking_xyz

fr3/walking_halfsphere