[Documentation] [TitleIndex] [WordIndex

Introduction

This project and wiki page is the product of a semester long independent study conducted for the Civil and Mechanical Engineering Department. Additional assistance and expertise was received from the Robotics Research Center within the Department of Electrical Engineering and Computer Science at The United States Military Academy at West Point.

Purpose

The purpose of this project was to develop a system that allows a user to control an autonomous vehicle through the use of hand and arm signals, similar to how military personnel ground guide vehicles during motor pool operations. Currently the Army is beginning to integrate more unmanned and autonomously capable assets into the fighting force. However, in order to effectively interface with these systems, the operator must complete extensive training on the equipment in order to be considered qualified a qualified operator. The goal for this project was to assess the possibility of giving autonomous/unmanned vehicles the capability of being controlled via some means that does not require any additional training on behalf of the individual, or the units that have access to these assets.

Assessment of Success

The success and viability of this system was determined based on ease of use, possibility for future expansion, cost of implementation, and the feasibility for implementation within the military. Each of these categories was considered when determining the success of the system. However, a large focus was placed on assessing the feasibility of implementation for military and combat uses without the need for additional training.

Development

Initial Approach

The initial idea behind this project was to make use of a motion capture camera, that has the ability to track a user's movements, as a means of converting certain poses or gestures into digital input into a RoS controllable robot. The first camera that was to be used in this project was made by Orbbec. This camera was initially selected because it has high quality sensors, and it was thought that an open source platform would make developing a project from scratch easier than having to overcome the inevitable software conflicts involved with using a sensor with preloaded software, like the Kinect sensor made by Microsoft.

As far as the methodology was concerned for how the user's poses would be recognized and converted into input for the robot, there were a few ideas I was considering. The first idea came from a user named TanvirParhar who published a RoS Wiki page, and package called neuro_gesture_kinect. The concept of this package was to use a method of "training" the camera to recognize certain gestures by giving the sensor a series of the same gesture. This series would then be stored, and at runtime the input being received from the user would then be compared against a previously compiled log of gestures. If the user input matched one of the gestures stored in the log then a move command would be sent to the robot

The second approach was to use the reference frames that were generated by the camera, and fixed to each of the user's joints, and compare the change in these reference frames to the fixed coordinate axes on the camera as a means of determining the user's joint/body position in space relative to the camera. Based on the determined body position, if the input matched the orientation of the reference frame for a given pose, that was predefined in the program, then a move command would be sent to the robot.

Revision 1.0 - The Camera

After experimenting with the Orbbec sensor for a few weeks, and trying to integrate the camera into the RoS framework it was determined that the open source community did not have an effective means of tracking a user's body movements or seamlessly integrating with the RoS environment. At this point, the decision was made to use the Kinect v2.0 sensor from Microsoft as it natively comes with the ability to track a user's joint position. The downside of this camera is that it is not open source, so the opportunities for expanding or fine tuning the sensor for this specific project were greatly limited. To track a user within RoS, using the Kinect sensor, a package called openni_tracker was used. The documentation for this package is fairly straightforward, and does not require any dependencies. This package is nice because once the user is calibrated by the camera, the generated reference frames can be visualized using rviz making the development process much easier since a lot of guess work was taken out of the equation.

Revision 1.1 - Pose Recognition

The initial approach for pose recognition was to use the RoS package called neuro_gesture_kinect created by TanvirParhar. After downloading the various dependencies for this package and struggling with the limited documentation detailing how to implement the package I determined that it was taking too much time to get the program to a useable level and began thinking of other means to detect specific gestures.

I first started with figuring out how the Kinect sensor was determining the position and orientation of the user's joints in space. The conclusion that I came to was that the camera sets itself as a fixed reference frame, and everything else in the world that moves is fixed spatially based on the movement relative to the Kinect. When the openni_tracker package prompts the user to calibrate using the Psi pose it is using this preprogrammed pose as a reference point that the sensor is able to recognize, and fix the positons of the user's joints in space. Once the calibration is complete, any movement by the user are published to the tf node where the displacement and rotation of each of the joints, relative to the sensor, can be seen. Using this information in conjunction with visualizing the movement of each of the joint fixed coordinate systems in rviz led me to the idea of finding a fixed joint position value in the /tf node (AKA a gesture), and using a series of conditional statements to determine if the user's body position fell within the predefined limits for the values previously determined for a specific pose.

My initial idea when using the joint transformations was to determine if a pose was being made by the user based on the rotation of the joints relative to the sensor. For this approach I examined the change in rotation for my right elbow. The challenge I ran into with this approach is that the tf node uses a quaternion based reference frame rather than roll, pitch, and yaw (RPY) to define the rotation of a joint in space. The reason for this is to prevent a phenomena referred to as gimbal lock. "Gimbal lock is the loss of one degree of freedom in a three-dimensional, three-gimbal mechanism that occurs when the axes of two of the three gimbals are driven into a parallel configuration, "locking" the system into rotation in a degenerate two-dimensional space (Gimbal Lock, Wikipedia)." In order to overcome this I converted from quaternion to RPY to make more sense of the output.

After a week of testing this approach it was found that the rotational reference frame of a given joint was different every time the program was restarted. As a result of this, there was no way to accurately determine whether or not the user was in fact issuing a gesture to the sensor. For this reason using the rotational positon of a specific joint in space would not have yielded the fine-tuned control that was required of the autonomous vehicle during motor pool operations. The next approach was to use the translation of a joint in space relative to the sensor. For this method the position of a joint in space is resolved based on its displacement from the calibration "Psi pose". Since the "Psi pose" is used as the "zero" for all other joint positions relative to the Kinect, the translational reference frame remains constant between program instances and allows for predictable and accurate joint position recognition

Revision 1.2 - Dynamic Tracking

The design aspect for this project was to develop the ability to dynamically track the user with the sensor mounted to the vehicle. However, after a few weeks of research it was determined that I was limited by the capabilities of the sensor that I was working with. Since the Microsoft Kinect sensor defines its position in space as the origin, and it requires the user to calibrate themselves to the camera before it is able to locate and track their body movements. When the sensor is mounted on a moving vehicle the Kinect is no longer able to determine its position in space, and therefore it cannot maintain its lock on the user. In order to overcome this problem I propose two possible alternatives. The first is to use a different sensor that is capable of dynamically updating its own position in space to redefine its origin, and allow for a continual lock on the user's body orientation as both the vehicle and the operator move through space. The second option is to set up an array of stationary cameras in the motor pool that cross-talk with each other to produce a 360 degree web in order to allow user tracking within a specified area.

Implementation

Required Hardware

Dependencies

It is expected that all of the listed dependencies should function correctly prior to beginning this walkthrough. Follow the installation tutorials on the linked RoS wiki pages for installation/implementation assistance.

Walkthrough

It is highly recommended to complete the first eleven lessons of the beginner intro to RoS tutorial on the RoS wiki before following this walkthrough. The reason behind this is that, in the essence of time, it is assumed that the reader has a very basic understanding of how RoS works, how to interact with RoS, and how each of the different components (ex. publisher and listener) interface with each other to send and receive data.

The following instructions for setting up the Kinect sensor come from TanvirParhar who outlined how he installed all of the drivers for the Kinect sensor to make it work with OpenNI.

Setting up the Kinect Sensor

"First of all download the suitable drivers from here. Start by intalling the suitable OpenNI driver, preferably the latest one. The OpenNI drivers are needed for the functioning of packages like openni_launch and software like RTABMap, or libraries like PCL, etc.

Once the tar file is downloaded, go to the directory where it is downloaded and uncompress it, by using:

$  tar -zxvf filename.tar.gz

or simply double-click it to open it with Archive Manager. Then install the driver by:

$  sudo ./install.sh

Next Download the NITE middleware, form the same place. Again uncompress it and install it.

$  tar -zxvf filename.tar.gz
$  sudo ./install.sh

Now, it's time to install the Openni_launch for ROS. Simply type:

$  sudo apt-get install ros-<rosdistro>-openni-launch

At this point, to check everything is installed, just connect the Kinect and type:

$  roslaunch openni_launch openni.launch

If the launch file runs, then all the drivers are working fine."

Downloading the Autonomy Script

For this portion, my gesture recognition script can be found on my GitHub page. The script's name is "gesture.py". The other two scripts on the page are: "gesture_comment.py" and "quat_to_rpy.py". The "gesture_comment.py" script contains the same code as "gesture.py", but it is fully commented to describe how everything works. The "quat_to_rpy.py" script simply converts any outputs that are in quaternion to the RPY frame. However, for the purposes of this walkthrough this script will not be used.

Once the "gesture.py" script has been successfully installed, ensure it is saved in the Source (/src) directory that within your catkin workspace (/catkin_ws). For my instance, my /catkin_ws directory can be found in the /home directory. To navigate to this file location and save the python file in the /src folder, your command line path should look somewhat similar to this:

$  cd home/catkin_ws/src/

Once the "gesture.py" script file is saved in this directory it can now communicate with all of the other nodes within RoS. It is imperative that this script is in this file location in the catkin workspace source folder or else the python file will NOT be able to interface with the Kinect sensor or the Turtle Bot.

Putting Everything Together

The first step in bringing all of these pieces together in order to control the Turtle Bot is to start roscore, to do this in the terminal type:

$  roscore

Once roscore initializes successfully, open a new terminal window and navigate to the location of the "gesture.py" script, and run the script.

$  cd home/catkin_ws/src/
$  ./gesture.py

Once the script is up and running you will not see any output from the python file for the time being until a user is being tracked by openni_tracker. Before we can start tracking a user, we first have to initialize and start the Turtle Bot. To do this, physically turn on your Turtle Bot and plug it in to the computer. Then open yet another terminal window and type:

$  roslaunch turtlebot_bringup minimal.launch

You should hear an audible noise come from the Turtle Bot once it has been initialized. Now we are ready to start the body tracking service to begin controlling the robot. To do this type:

$  rosrun openni_tracker openni_tracker

At this point position yourself in front of the Kinect sensor, and wait until the program identifies a new user. Once a new user has been identified, assume the Psi Pose to calibrate the Kinect sensor to your body. Hold this pose until the program has successfully calibrated itself.

The body tracking will NOT work until you have successfully calibrated the Kinect sensor. Also note that the python script will NOT work if the user is not defined as USER 1. If a different user number is assigned type CTRL + C in the terminal and enter the previous command again until User 1 is defined as the current user.

How to Control the Robot

Once the calibration has completed place your arms down at your side.

NOTE: The robot can only handle one gesture at a time, and will not perform combined movements. The first gesture that is recognized is the movement that will take place

Vizualizing the Refrence Frames

$  rosrun rviz rviz

$  roscore
$  rosrun openni_tracker openni_tracker

Once you are being tracked by the sensor go in to rviz (be careful to move out of view of the sensor) and add a /tf window. This allows you to specifically track the outputs of the transformation node (/tf). Then set the openni_depth_frame as the fixed frame. This defines the sensor as the origin in space so that the program knows where to generate the coordinate axes. At this point you should see an array of reference frames that are labeled according to their body part (the "X" axis is red, the "Y" axis is green, and the "Z" axis is blue).

Another option is to output the data of the /tf node to show the numerical values for the position and rotation of each of the reference frames. To do this type in a new terminal window:

$  rostopic echo /tf

Conclusion

Accomplishments

The primary objective of this independent study was to develop a system that allows a user to control an autonomous vehicle through the use of hand and arm signals. Throughout the course of this project I was able to develop the ability to control a Turtle Bot running the Ros operating system by using motion capture technology. However, the ability to mount the sensor on the robot and dynamically track the user was not achieved in this project.

Future of the Project

Upon completion of this project, an analysis was done to determine the practicality of continuing development. In order to determine if this independent study was worthwhile the user needed to be able to maintain consistent and predictable control over the robot. The learning curve to physically control the robot also needed to be short and straightforward in order to limit the amount of time and money required to implement this system in an operational environment. After controlling the robot myself and becoming very familiar with its capabilities it is was determined that maintaining consistent and predictable control was not an issue. Additionally, while at West Point's Projects Day I had the opportunity to teach a group of middle schoolers how to control the robot. Within, five minutes of my instruction all of the kids were easily maneuvering the robot without any additional help. This served as a proof of concept for the ease of use of the system.

The program could primarily be improved to dynamically track a user, rather than requiring the user to remain in a fixed area in front of a stationary camera. This would provide better freedom of maneuver for the system as well less constraints as to where this system could be implemented. This program could be expanded by testing its functionality on multiple different kinds of autonomous vehicles, with different steering configurations to determine if there are any other further limitations to the system. Overall, there is room for expansion and further development within this area of study. However, I believe that this project could serve a potential easy solution to a overlooked and underappreciated problem that will only become more prevalent as our military progresses technologically.


2024-03-23 12:42