3DSemanticMapping JINT 2020 Save

Extending Maps with Semantic and Contextual Object Information for Robot Navigation
[Project Page] [Paper arXiv]

This repository provides a code base to evaluate and test the semantic object mapping from the paper Extending Maps with Semantic and Contextual Object Information for Robot Navigation (arXiv version). The project core method is an object segmentation that leverages a CNN-based object detector (Yolo) with a 3D model-based segmentation technique (RANSAC) to perform instance semantic segmentation, and to localize, identify, and track different classes of objects in the scene over time.

The package is implemented on ROS and we provide pre-trained Yolo model weights and some ROS bags datasets for testing. Please follow the install and setup instructions to run the semantic object mapping system with a rosbag recorded data. This is research code, expect that it can change and any fitness for a particular purpose is disclaimed.

Installation

The install steps were tested on Ubuntu 16.04, ROS Kinetic and Python 2.7. The package requires ROS, Catkin and CUDNN properly installed. We divide instructions to meet these requirements in two parts, including some intermediate sanity check tips:

• ROS and Catkin workspace setup (if not set already): Please go to INSTALL_ROS.md for detailed instructions.
• CUDA/CUDNN installation (if not set already): Please refer to INSTALL_CUDNN.md for detailed instructions.

Build Mapping Project

Once the package requirements are met, you can download and build the package uisng the following instructions:

cd ~/code/catkin_ws/src
git clone --recurse-submodules https://github.com/verlab/3d-object-semantic-mapping.git

Before compiling the project, you might need to change line 23 in file "~/code/catkin_ws/src/3d-object-semantic-mapping/darknet_ros/darknet_ros/CMakeLists.txt" to assert the cuda compute capability to your GPU. For a compute capability of 5.0, change this line to -gencode arch=compute_50,code=[sm_50,compute_50] (please have a look at the instructions at INSTALL_CUDNN.md):

vim 3d-object-semantic-mapping/darknet_ros/darknet_ros/CMakeLists.txt

And then build the packages in the following order:

cd ~/code/catkin_ws/
catkin config --default-devel-space --default-build-space --default-install-space --default-source-space --cmake-args -DCMAKE_BUILD_TYPE=Release -DCMAKE_C_COMPILER=/usr/bin/gcc-6 -DCMAKE_CXX_COMPILER=/usr/bin/g++-6
catkin build darknet_ros_msgs
catkin build custom_msgs
catkin build
source ~/.bashrc

Getting Started: Package Testing

The provided code can run in two modes: offline and online.

Offline Testing

We provide pre-trained weights to detect objects such as "door", "water fountain", "chair", "person", "fire extinguisher". Three datasets containing data streams recorded with a mobile robot are also available for testing. Please have a look on the project website for futher information of these sequences.

First plese download the weights yolo-custom5.weights and the rosbag sequence3_kinect.bag for testing:

sudo apt install python-pip python-numpy python-scipy
pip install gdown
source ~/.bashrc
roscd darknet_ros/yolo_network_config/weights/
gdown https://drive.google.com/uc?id=1Q1oik-R7yPJn6Lh9iMjj5ZMYx13xwliE
roscd auto/../datasets/sequence3-kinect/files
gdown https://drive.google.com/uc?id=1u5wpCdpMP81VrADteEpRVhvWpbrq2Y1Q

The package running requires three terminal windows (shells). In the first terminal (shell) start darknet_ros and rviz:

roslaunch auto quick_start.launch

In terminal 2 go to dataset folder and start playing the dataset bag with RATB-SLAM:

roscd auto/../datasets/sequence3-kinect
./rtab-play.sh

Then press SPACE key on keyboard on terminal 2 to start publishing the sensor topics. Then open terminal 3 and start the object localizer node:

roslaunch auto obj_positioner.launch

You should have the augmented map display in the rviz window on the fly, such as the following images for sequence3-kinect:

The display of the point cloud superposed with the semantic map is done by marking the option on rviz *MapCloud* to on.

Customized configuration

Few modificateasy to use with your own bags and other network weights. In order to configure the stack to work on different datasets and classes, some configuration files need to be adjusted:

Object detection configuration files

Configuration files related to darknet_ros and yolo detector itself. This includes the following:

• auto/launch/yolo_detector.launch : Change the parameters file.
• darknet_ros/darknet_ros/config : Add the new parameters file for new network. Also, update ros.yaml topic names.
• darknet_ros/darknet_ros/yolo_network_config/cfg : Add new network configuration file.
• darknet_ros/darknet_ros/yolo_network_config/weights : Add new network weights.

Object localization node configuration

Configuration files related to the 3D projector and the filter. This includes:

• auto/param/object_positioner.yaml : Every node configuration for the object projection module is there, and it includes topic names, class detection list, frames of reference, camera parameters, etc.

Online Testing

The online testing mode requires a physical robot mounted with an RGB-D camera. Requirements: Kobuki base/other turtlebot base, RGBD camera, laser scan (optional) (RGBD camera depth stream can be converted to laser scan, but usually has lower range and accuracy).

Start robot base:

roslaunch auto initialize.launch

And then run the slam and yolo_detector nodes as described in the offline test.

Notes Before usage, check that no packages publish tf transformations, i.e., 'publish_tf' flag in launch files are set to false. Only the rosbag play and robot description launch files should publish tf's.

Troubleshooting to Install on Ubuntu 18.04 and ROS Melodic

• Please have a look at INSTALL_MELODIC.md to adjust the installation.

Citation & Contact

If you find this code useful for your research, please consider citing the following papers:

    @article{semanticMapVerlab20,
	  Title          = {Extending Maps with Semantic and Contextual Object Information for Robot Navigation: a Learning-Based Framework using Visual and Depth Cues},
	  Author         = {Renato Martins and Dhiego Bersan and Mario F. M. Campos and Erickson R. Nascimento},
	  Booktitle      = {Journal of Intelligent and Robotic Systems},
	  Year           = {2020}
	}

    @inproceedings{semanticMapVerlab18,
	  Title          = {Semantic Map Augmentation for Robot Navigation: A Learning Approach based on Visual and Depth Data},
	  Author         = {Dhiego Bersan and Renato Martins and Mario F. M. Campos and Erickson R. Nascimento},
	  Booktitle      = {IEEE Latin American Robotics Symposium, LARS},
	  Year           = {2018}
	}