MMS-SLAM

Multi-modal semantic SLAM in dynamic environments (Intel Realsense L515 as an example)

This code is modified from SSL_SLAM

Modifier: Wang Han, Nanyang Technological University, Singapore

[Update] AGV dataset is available online! (optional)

1. Solid-State Lidar Sensor Example

1.1 Scene Reconstruction in Dynamic Environments

1.2 Mapping result

1.3 Human & AGV recognition result

1.4 Performance Evaluation

1.5 Detection Result

2. Prerequisites

2.1 Ubuntu and ROS

Ubuntu 64-bit 18.04.

ROS Melodic. ROS Installation

2.2. Ceres Solver

Follow Ceres Installation.

2.3. PCL

Follow PCL Installation.

Tested with 1.8.1

2.4 OctoMap

Follow OctoMap Installation.

$ sudo apt install ros-melodic-octomap*

2.5. Trajectory visualization

For visualization purpose, this package uses hector trajectory sever, you may install the package by

sudo apt-get install ros-melodic-hector-trajectory-server

Alternatively, you may remove the hector trajectory server node if trajectory visualization is not needed

3. Build

3.1 Clone repository:

    cd ~/catkin_ws/src
    git clone https://github.com/wh200720041/mms_slam.git
    cd ..
    catkin_make
    source ~/catkin_ws/devel/setup.bash

chmod python file

roscd mms_slam
cd src
chmod +x solo_node.py

3.2 install mmdetection

create conda environment (you need to install conda first)

conda create -n solo python=3.7 -y
conda activate solo

install PyTorch and torchvision following the official instruction (find your cuda version)

conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
conda install -c conda-forge addict rospkg pycocotools

install mmdet 2.0

roscd mms_slam 
cd dependencies/mmdet
python setup.py install

it takes a while (a few minutes to install)

3.3 Download test rosbag and model

You may download our trained model and recorded data if you dont have realsense L515, and by defult the file should be under /home/username/Downloads

put model under mms_slam/config/

cp ~/Downloads/trained_model.pth ~/catkin_ws/src/MMS_SLAM/config/

unzip rosbag file under Download folder

cd ~/Downloads
unzip ~/Downloads/dynamic_warehouse.zip

3.4 Launch ROS

if you would like to create the map at the same time, you can run

    roslaunch mms_slam mms_slam_mapping.launch

if only localization is required, you may refer to run

    roslaunch mms_slam mms_slam.launch

if you would like to test instance segmentation results only , you can run

    roslaunch mms_slam mms_slam_detection.launch

if ModuleNotFoundError: No module named 'alfred', install alfrey-py from pip install

pip install alfred-py

4. Sensor Setup

If you have new Realsense L515 sensor, you may follow the below setup instructions

4.1 L515

4.2 Librealsense

Follow Librealsense Installation

4.3 Realsense_ros

Copy realsense_ros package to your catkin folder

    cd ~/catkin_ws/src
    git clone https://github.com/IntelRealSense/realsense-ros.git
    cd ..
    catkin_make

4.4 Launch ROS with live L515 camera data

In you launch file, uncomment realsense node like this

    <include file="$(find realsense2_camera)/launch/rs_camera.launch">
        <arg name="color_width" value="1280" />
        <arg name="color_height" value="720" />
        <arg name="filters" value="pointcloud" />
    </include>

and comment rosbag play like this

<!-- rosbag
    <node name="bag" pkg="rosbag" type="play" args="- -clock -r 0.4 -d 5 $(env HOME)/Downloads/dynamic_warehouse.bag" />
    <param name="/use_sim_time" value="true" />  
-->

6 Training on AGV & Human dataset

6.1

The human data are collected from COCO dataset train2017.zip(18G) and val_2017.zip(1G) The AGV data are manually collected and labelled Download(1G)

cd ~/Downloads
unzip train2017.zip
unzip val2017.zip
unzip agv_data.zip
mv ~/Downloads/train2017 ~/Downloads/train_data
mv ~/Downloads/val2017 ~/Downloads/train_data
mv ~/Downloads/train_data/agv_data/* ~/Downloads/train_data/train2017

note that it takes a while to unzip

to train a model

roscd mms_slam
cd train
python train.py train_param.py

if you have multiple gpu (say 4 gpus), you can change '1' to your GPU number The trained model is under mms_slam/train/work_dirs/xxx.pth,

7 Acknowlegement

Thanks for A-LOAM and LOAM and LOAM_NOTED and MMDetection and SOLO.

VinMing/MMS_SLAM