semantic_slam

Semantic SLAM

Author: Xuan Zhang, Supervisor: David Filliat

Research internship @ENSTA ParisTech

Semantic SLAM can generate a 3D voxel based semantic map using only a hand held RGB-D camera (e.g. Asus xtion) in real time. We use ORB_SLAM2 as SLAM backend, a CNN (PSPNet) to produce semantic prediction and fuse semantic information into a octomap. Note that our system can also be configured to generate rgb octomap without semantic information.

Project Report & Demo:

• Demo: [Youtube] [Youku]
• Project report: [Google] [Baidu]

Acknowledgement

This work cannot be done without many open source projets. Special thanks to

• ORB_SLAM2, used as our SLAM backend.
• pytorch-semseg, used as our semantic segmantation library.
• octomap, used as our map representation.
• pcl library, used for point cloud processing.

License

This project is released under a GPLv3 license.

Overview

Dependencies

• Openni2_launch

sudo apt-get install ros-kinetic-openni2-launch

• ORB_SLAM2

We use ORB_SLAM2 as SLAM backend. Please refer to the official repo for installation dependencies.

• PyTorch 0.4.0

• For other dependencies, please see semantic_slam/package.xml

Installation

Build ORB_SLAM2

After installing dependencies for ORB_SLAM. You should first build the library.

cd ORB_SLAM2
./build.sh

Install dependencies

rosdep install semantic_slam

Make

cd 
catkin_make

Run with camera

Launch rgbd camera

roslaunch semantic_slam camera.launch

Run ORB_SLAM2 node

roslaunch semantic_slam slam.launch

When the slam system has finished initialization, try to move the camera and check if the camera trajectory in the viewer is reasonable, reset SLAM if not.

Run semantic_mapping

You can now run the semantic_cloud node and the octomap_generator node. You will have to provide trained models, see links below.

roslaunch semantic_slam semantic_mapping.launch

This will also launch rviz for visualization.

You can then move around the camera and construct semantic map. Make sure SLAM is not losing itself.

If you are constructing a semantic map, you can toggle the display color between semantic color and rgb color by running

rosservice call toggle_use_semantic_color

Run with ros bag

If you want to test semantic mapping without a camera, you can also run a rosbag.

Download rosbag with camera position (tracked by SLAM)

demo.bag

Run semantic_mapping

roslaunch semantic_slam semantic mapping.launch

Play ROS bag

rosbag play demo.bag

Trained models

• Model trained on ade20k dataset
• Model fine tuned on SUNRGBD dataset

Run time

Evaluation is done on a computer with 6 Xeon 1.7 GHz CPU and one GeForce GTX Titan Z GPU. Input image size is 640×480 recorded by a camera Asus Xtion Pro.

When our system works together, ORB-SLAM works at about 15 Hz (the setting is 30 Hz). Point cloud generation alone can run at 30 Hz. Semantic segmentation runs at about 2 to 3 Hz. Octomap insertion and visualization works at about 1 Hz. Please refer to section 3.6.2 of the project report for more analysis of run times.

Citation

To cite this project in your research:

    @misc{semantic_slam,
      author = {Xuan, Zhang and David, Filliat},
      title = {Real-time voxel based 3D semantic mapping with a hand held RGB-D camera},
      year = {2018},
      publisher = {GitHub},
      journal = {GitHub repository},
      howpublished = {\\url{https://github.com/floatlazer/semantic_slam}},
    }

Configuration

You can change parameters for launch. Parameters are in ./semantic_slam/params folder.

Note that you can set octomap/tree_type and semantic_cloud/point_type to 0 to generate a map with rgb color without doing semantic segmantation.

Parameters for octomap_generator node (octomap_generator.yaml)

namespace octomap

• pointcloud_topic
  • Topic of input point cloud topic
• tree_type
  • OcTree type. 0 for ColorOcTree, 1 for SemanticsOcTree using max fusion (keep the most confident), 2 for SemanticsOcTree using bayesian fusion (fuse top 3 most confident semantic colors). See project report for details of fusion methods.
• world_frame_id
  • Frame id of world frame.
• resolution
  • Resolution of octomap, in meters.
• max_range
  • Maximum distance of a point from camera to be inserted into octomap, in meters.
• raycast_range
  • Maximum distance of a point from camera be perform raycasting to clear free space, in meters.
• clamping_thres_min
  • Octomap parameter, minimum octree node occupancy during update.
• clamping_thres_max
  • Octomap parameter, maximum octree node occupancy during update.
• occupancy_thres
  • Octomap parameter, octree node occupancy to be considered as occupied
• prob_hit
  • Octomap parameter, hitting probability of the sensor model.
• prob_miss
  • Octomap parameter, missing probability of the sensor model.
• save_path
  • Octomap saving path. (not tested)

Parameters for semantic_cloud node (semantic_cloud.yaml)

namespace camera

• fx, fy, cx, cy
  • Camera intrinsic matrix parameters.
• width, height
  • Image size.

namespace semantic_pcl

• color_image_topic
  • Topic for input color image.
• depth_image_topic
  • Topic for input depth image.
• point_type
  • Point cloud type, should be same as octomap/tree_type. 0 for color point cloud, 1 for semantic point cloud including top 3 most confident semanic colors and their confidences, 2 for semantic including most confident semantic color and its confident. See project report for details of point cloud types.
• frame_id
  • Point cloud frame id.
• dataset
  • Dataset on which PSPNet is trained. "ade20k" or "sunrgbd".
• model_path
  • Path to pytorch trained model.