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:
@ARTICLE{9045978,
author={Bavle, Hriday and De La Puente, Paloma and How, Jonathan P. and Campoy, Pascual},
journal={IEEE Access},
title={VPS-SLAM: Visual Planar Semantic SLAM for Aerial Robotic Systems},
year={2020},
volume={8},
number={},
pages={60704-60718},
doi={10.1109/ACCESS.2020.2983121}}
Video
How do I set it up?
First install g2o following these instructions (Tested on Kinetic and Melodic Distributions):
- sudo apt-get install ros-$ROS_DISTRO-libg2o
- sudo cp -r /opt/ros/$ROS_DISTRO/lib/libg2o_* /usr/local/lib
- sudo cp -r /opt/ros/$ROS_DISTRO/include/g2o /usr/local/include
Install OctopMap server for map generation capabilities:
- sudo apt install ros-$ROS_DISTRO-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:=${HOME}/Downloads/entire_lab_3_rounds.bag show_rviz:=true
Using Docker Image
If the code is giving problems with you local machine, you can try the docker image created with the repo and the required settings.
Download Docker from: Docker
Follow the commands to run the algorithm with the docker
docker pull hridaybavle/semantic_slam:v1
docker run --rm -it --net="host" -p 11311:11311 hridaybavle/semantic_slam:v1 bash
cd ~/workspace/ros/semantic_slam_ws/
source devel/setup.bash
roslaunch semantic_SLAM ps_slam_with_snap_pose_bucket_det_lab_data_with_octomap.launch bagfile:=${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.
Published TFs
-
map to odom transform: The transform published between the map frame and the odom frame after the corrections from the semantic SLAM.
-
base_link to odom transform: The transform published between the base_link (on the robot) frame and the odom frame as estimated by the VO/VIO algorithm.
36 Jan 3, 2023
![[ICCV 2021 (oral)] Planar Surface Reconstruction from Sparse Views](https://github.com/jinlinyi/SparsePlanes/raw/main/docs/static/architecture.jpg)
89 Jan 5, 2023
48 Dec 15, 2022
101 Dec 13, 2022
24 Oct 26, 2022
37 Dec 22, 2022
247 Dec 28, 2022
66 Nov 8, 2022
149 Dec 15, 2022
![[semantic_graph_slam_node-9] process has died](https://avatars.githubusercontent.com/u/40681527?v=4)
Imagino que se deben a incompatibilidades entre versiones de opencv, podrías confirmarme esto último? Sería posible trabajar con una versión de opencv diferente?
Gracias, un saludo!
7 Aug 07, 2022
124 Dec 20, 2022
8k Jan 04, 2023
1.6k Jan 08, 2023
40 Dec 12, 2022
6 Feb 12, 2022
0 Dec 13, 2021
185 Dec 26, 2022
531 Dec 20, 2022
68 Dec 16, 2022
153 Dec 22, 2022
19 Oct 07, 2022
1.1k Jan 01, 2023
135 Dec 02, 2022
8 Jan 02, 2023
155 Dec 17, 2022
125 Jan 03, 2023
5 Oct 02, 2022
89 Jan 04, 2023
2 Feb 16, 2022