Awesome

Semantic SLAM

This package can perform optimization of pose estimated from VO/VIO methods which tend to drift over time. It uses planar surfaces extracted from object detections in order to create a sparse semantic map of the environment, thus optimizing the drift of the VO/VIO algorithms.

In order to run this package you will need two additional modules

• A VO/VIO algorithm: ROVIO, VINS_MONO, OKVIS etc.
• An Object Detector: Yolo, Shape Color Detector

Currently it can extract planar surfaces and create a semantic map from from the following objects:

• chair
• tvmonitor
• book
• keyboard
• laptop
• bucket
• car

Related Paper:

VPS-SLAM: Visual Planar Semantic SLAM for Aerial Robotic Systems, Hriday Bavle, Paloma de la Puente, Jonathan How, Pascual Campoy, Journal of IEEE Access.

Video

<a href="https://vimeo.com/368217703" target="_blank"><img src="semantic_slam.png" alt="Semantic SLAM" width="320" height="240" border="10" /></a>

How do I set it up?

First install g2o following these instructions (assuming ROS kinetic distribution):

- sudo apt-get install ros-kinetic-libg2o
- sudo cp -r /opt/ros/kinetic/lib/libg2o_* /usr/local/lib
- sudo cp -r /opt/ros/kinetic/include/g2o /usr/local/include

Install OctopMap server for map generation capabilities:

- sudo apt install ros-kinetic-octomap*

Try a simple example with pre-recorded VIO pose and a blue bucket detector:

Create a ros workspace and clone the following packages:

• Download the rosbag:

    wget -P ~/Downloads/ https://www.dropbox.com/s/jnywuvcn2m9ubu2/entire_lab_3_rounds.bag

• Create a workspace, clone the repo and compile:

    mkdir -p workspace/ros/semantic_slam_ws/src/ && cd workspace/ros/semantic_slam_ws/src/    
    git clone https://github.com/hridaybavle/semantic_slam && git clone https://bitbucket.org/hridaybavle/bucket_detector.git   
    cd .. && catkin build --cmake-args -DCMAKE_BUILD_TYPE=Release

• Launch and visualize

    source devel/setup.bash
    roslaunch semantic_SLAM ps_slam_with_snap_pose_bucket_det_lab_data_with_octomap.launch bagfile:=${HOME}/Downloads/entire_lab_3_rounds.bag show_rviz:=true  

**Note: In the example above octomap is used for mapping. Right now at loop closure the complete map is not optimized and will be done in the future.

Using Docker Image

If the code is giving problems with you local machine, you can try the docker image created with the repo and the required settings.

Download Docker from: Docker

Follow the commands to run the algorithm with the docker

  docker pull hridaybavle/semantic_slam:v1 	
  docker run --rm -it --net="host" -p 11311:11311 hridaybavle/semantic_slam:v1 bash
  cd ~/workspace/ros/semantic_slam_ws/
  source devel/setup.bash
  roslaunch semantic_SLAM ps_slam_with_snap_pose_bucket_det_lab_data_with_octomap.launch bagfile:=${HOME}/Downloads/entire_lab_3_rounds.bag show_rviz:=false  

Open a new terminal and rviz in local machine

  cd ~/Downloads/ && wget https://raw.githubusercontent.com/hridaybavle/semantic_slam/master/rviz/graph_semantic_slam.rviz
  rviz -d graph_semantic_slam.rviz	

Subsribed Topics

• /SQ04/snap_vislam/vislam/pose (geometry_msgs/PoseStamped)
  The default snapdragon VIO pose published in NED in frame. This message can be remapped remapped to any other VO pose message publishing in NED frame. (See frame conventions)

• /rovio/odometry (geometry_msgs/PoseStamped)
  The VIO odometry published in ENU frame. Can be remapped to the desired topic name in the launcher.

• /depth_registered/points (sensor_msgs/PointCloud2) The point cloud required for planar surface extraction.

• /darknet_ros/bounding_boxes(darknet_msgs_ros/BoundingBoxes)
  The detection bounding boxes published by yolo if using the yolo detector ros package.

• /image_processed/bounding_boxes(ShapeColor_ObjectDetection/DetectedObjects) The detection bounding boxes if using the bucket detector. It can be downloaded from the link above.

Published Topics

• robot_pose(geometry_msgs/PoseStamped) The pose of the robot estimated by the algo.

• robot_path(nav_msgs/Path)
  The path of the robot psoe for visualization.

• keyframe_poses(geometry_msgs/PoseArray)
  The poses of the keyframe being added to the g2o graph.

• mapped_landmarks(visualization_msgs/MarkerArray)
  The mapped semantic planar surfaces.

• detected_landmarks(visualization_msgs/MarkerArray)
  The detected landmarks in the current frame.

The configurations of the algorithms can be found inside the cfg folder in order to be changed accordingly.