Skip to content

Using pre-trained DL models and Transformations for generating occupancy maps. Includes some other basic deep learning tasks. Feel free to contribute.

License

Notifications You must be signed in to change notification settings

qureshinomaan/Generate-Occupancy-Maps

Folders and files

NameName
Last commit message
Last commit date

Latest commit

0a4ff44 · Jul 15, 2020

History

85 Commits
Jun 7, 2020
Jun 7, 2020
Mar 28, 2020
Jun 7, 2020
Jul 15, 2020
Jun 7, 2020
May 6, 2020
Feb 3, 2020
Jun 7, 2020
Jun 7, 2020
Mar 28, 2020
May 4, 2020
Apr 26, 2020

Repository files navigation

Generate-Occupancy-Maps

Using pre-trained Deep Learning models and Transformations for generating occupancy maps.

Note

If you find any difficulty in using some notebook or script, please feel free to create an issue. This is a work in progress and I will keep making changes to the repository for a while.

Occupancy Maps

Occupancy Grid Mapping refers to a family of computer algorithms which address the problem of generating maps from noisy and uncertain data.

Algorithm

The system takes a stereo pair and generates a depth map(using PSMNet) and instance segmented scene(using maskrcnn). We then use these to get a 3D Model of the scene. This 3d model is projected to the ground to get the occupancy grid. "Algorithmic Pipeline"

Depth Image

In CV, a depth image contains information about depth of surfaces presents in the image. Some methods to get the depth image.

  • Stereo Methods.
  • Monocular Methodss.
  • Network (PSMNET)
    • PSMNet actually gives disparity map, which can be converted to a depth map.

Instance Segmentation

We identify each instance of each object featured in the image instead of categorizing each pixel like in semantic segmentation.

  • We can use pretrained models of detectron2 model-zoo.

Input

The inputs are

  • Image or scene of a self driving car.
  • Resolution Size for occupancy grid(to be done).
  • Size of image input.

Output

The output consists of

  • Semantic Segmentation in one folder.
  • Depth Image in another folder.
  • Occupancy Grid in another.
  • The output images should have a proper name.

Dataset

  • KITTI Dataset is used for testing in the project.

Tasks

  • Debug PSMNet to get the depth image.
  • Using detectron2 to get the instance segmented image(without text).
  • Using depth image to obtain a 3d visualisation.
  • Getting occupancy maps from 3d visualisation.
  • Format in the input/output format mentioned above.
  • Making the occupancy map better.
  • Writing scripts so that the system is easily usable.
  • Adding the Monocular depth maps.
  • Stitching the point clouds to get a map of the environment(a little ambitious for now).
    • If you are a beginner and want to team up on this, please contact.

Other Tasks

Resources

Mentor

Shashank Srikanth

About

Using pre-trained DL models and Transformations for generating occupancy maps. Includes some other basic deep learning tasks. Feel free to contribute.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published