Laser Calibration with People Detection

Overview

This ROS package combines multiple laser scans from RPLIDAR A3M1 lasers into a calibrated point cloud. This package will then track the ankles of moving people in the point cloud and predict their paths with linear regression.

Keywords: 2-dimensional laser, RPLIDAR A3M1, point cloud, matrix transformations, calibration, linear regression

Authors

• Tomohide Bessho - UNR REU student - https://github.com/tzxb018
• Andrew Palmer - UNR Grad Mentor - https://github.com/ahpalmerUNR

Install

• Clone rplidar_ros into your kinetic catkin source directory: https://github.com/Slamtec/rplidar_ros
• Install sckikit-learn: https://scikit-learn.org/stable/install.html

  pip install -U scikit-learn
  

  • If you are using Python 2.7, you will need to install scikit-learn 0.20 instead of the latest version
• Run the following commands to install the package

    cd catkin_workspace/src
    git clone https://github.com/tzxb018/laser_calibration_with_people_detection/
    cd ../
    catkin_make
  

Usage

• Use the combined launch file to run the package

  roslaunch lc start.launch
  

• If this launch file does not work (normally happens when playing back a rosbag file), you will need to run each script sepeartely. Do each command in a sepearte terminal.

  roscore
  rosrun rviz rviz
  rosbag play -l <rosbag.file> (If you are running a rosbag file)
  rosrun lc laser_tf.py
  rosrun lc detection_target.py <len. of leg 1><len. of leg 2><len. of leg 3><radius of target>
  rosrun lc tf_publisher.py
  rosrun lc people_detection.py
  

• If this script gives a "tf2.ExtrapolationException: Lookup would require extrapolation into the past." error, rerun the script until the error resolves (only occurs when playing an old rosbag file).

Important Notes

• To search for the triangular calibration target, you will need to change the arguments for the detection_target.py script.
  • If you are using the launch file, please go into the launch file and change the arguments in the following format:
    • length of shortest leg, length of longer leg, length of hypotnuse, length of radius of targets (all in meters)
    • If you are manually running each script, add the same agruments to the script command.
    • Ex: $ rosrun lc detection_target.py .57 .89 1.05 .07
• The laser subscriber topics (explained after), were all namespaced into the following three names: hog, mouse, snake.
  • If you are trying to use four lasers instead of three, go into the tf_publisher.py script and uncomment lines 156, 179-181, 218-223, and 283-293.
  • Also, go into people_detection.py script and uncomment lines 395, 404-405, 419, 488-504.
• To launch the lasers, download and catkin_make this package onto each Raspberry Pi with the RPLIDAR A3M1 laser connected to it. For each Pi, run a different launch file. Please make sure that the Pis run lc_hog.launch, lc_mouse.launch, lc_snake.launch, and lc_duck.launch in that following order.

Subscribed Topics

• /hog/scan0
  • the parent laser for the trasnformations
• /mouse/scan0
  • one of the children lasers
• /snake/scan0
  • another one of the children lasers
• /duck/scan0
  • if you are running 4 lasers, this would be another child laser

Published Topics

• /ankles
  • displays the ankle markers and the prediction markers on rviz
• /combined
  • displays the overlap of all the lasers being run as a point cloud
• /combined_movement
  • displays the overlap of all the lasers that have detected movement as a point cloud
• /possible/triangles
  • displays the triangular calibration target in rviz
• /tf
  • manifests the tf between the different lasers

Under the Hood

Files for running

• src/laser_tf.py - calculates the tfs between two different lasers
• src/detection_target.py - finds and calibrates to the calibration target by searching for cirlces and finding the distances between each circle to search for the triangular shape of the target. Requires the lengths of the three legs of the target in ascending order and the radius of the circles of the target as arguments in the terminal
• src/tf_publisher.py - publishes the calculated tfs and applies them to the laser scans. The publisher publishes the combined scans as a point cloud.
• src/people_detection.py - searches for ankles in the combined point cloud, marks and tracks them, and uses linear regression to predict their paths.

Acknowledgement

We used some code from the Palmer's pr2_intent_rec package: https://github.com/ahpalmerUNR/pr2_intent_rec. The circle detection algorithm is modified from this package to better suit the laser scans we use in this package.