MedlarTea / MPF_GRR_SLT

A width-based Monocular Person Following system with Global descriptor, Ridge Regression online target classifier and Short-Long-Term sampling strategy (MPF_GRR_SLT)

MPF_GRR_SLT

TODO

Update our new system of "Robot Person Following Under Partial Occlusion"

Dependencies

Our computer settings:

• Ubuntu 18.04
• Melodic
• GTX 2060/ GTX 1650

1. Create a conda environment and install pytorch

conda create -n mono_following python=3.8
conda activate mono_following
# This is based on your GPU settings, other settings should be careful
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch

2. Install python related packages:

pip install -r requirements.txt
git clone https://github.com/eric-wieser/ros_numpy
cd ros_numpy
python setup.py install

3. Install cpp related packages:

• OpenCV==3.4
• Eigen==3.0+

Download pre-trained weights

• Download yolox-s and yolox-m, then make director mono_tracking/scripts/detector_2d/weights and put the checkpoints to it.

Our tf tree

base_link->camera_link->camera_optical_link

How to use

Tracking

# launch width-based monocular people tracking
# If running with rosbag, use_sim_time:=true; if the image topic is compressed, sim:=true
roslaunch mono_tracking all_mono_tracking.launch sim:=true use_sim_time:=true
# If running in real robot, use_sim_time:=false;
roslaunch mono_tracking all_mono_tracking.launch sim:=false use_sim_time:=false

• Input: /camera/color/image_raw
• Output: mono_tracking/msg/TrackArray.msg

Target Identification

# launch our GRR_SLT_MPF person following, use_sim_time:=true for rosbag
roslaunch mono_followng mono_following.launch use_sim_time:=true
# launch our GRR_SLT_MPF person following, use_sim_time:=false for robot running
roslaunch mono_following mono_following.launch use_sim_time:=false

• Input: mono_tracking/msg/TrackArray.msg
• Output: mono_following/msg/Target.msg

Controlling and Following

# launch control
roslaunch mono_control mono_controlling.launch

• Input: mono_following/msg/Target.msg; /bluetooth_teleop/joy
• Output: /cmd_vel

Target identification Test

TODO

• Improve simplicity of the code
• Release evaluation results

Citation

Acknowledge

• YOLOX_deepsort_tracker
• monocular_person_following