Intro

In Kinect.md, the previous generations dicussed the prospects and limitations of using a Kinect camera. We attempted to use the new Kinect camera v2, which was released in 2014.

Thus, we used the libfreenect2 package to download all the appropiate files to get the raw image output on our Windows. The following link includes instructions on how to install it all properly onto a Linux OS.

https://github.com/OpenKinect/libfreenect2

Issues

We ran into a lot of issues whilst trying to install the drivers, and it took about two weeks to even get the libfreenect2 drivers to work. The driver is able to support RGB image transfer, IR and depth image transfer, and registration of RGB and depth images. Here were some essential steps in debugging, and recommendations if you have the ideal hardware set up:

• Even though it says optional, I say download OpenCL, under the "Other" option to correspond to Ubuntu 18.04+

• If your PC has a Nvidia GPU, even better, I think that's the main reason I got libfreenect to work on my laptop as I had a GPU that was powerful enough to support depth processing (which was one of the main issues)

• Be sure to install CUDA for your Nvidia GPU

Please look through this for common errors:

https://github.com/OpenKinect/libfreenect2/wiki/Troubleshooting

Although we got libfreenect2 to work and got the classifier model to locally work, we were unable to connect the two together. What this meant is that although we could use already saved PNGs that we found via a Kaggle database (that our pre-trained model used) and have the ML model process those gestures, we could not get the live, raw input of depth images from the kinect camera. We kept running into errors, especially an import error that could not read the freenect module. I think it is a solvable bug if there was time to explore it, so I also believe it should continued to be looked at.

However, also fair warning that it is difficult to mount on the campus rover, so I would just be aware of all the drawbacks with the kinect before choosing that as the primary hardware.

Database

https://www.kaggle.com/gti-upm/leapgestrecog/data

Machine Learning model

https://github.com/filipefborba/HandRecognition/blob/master/project3/project3.ipynb

• What this model predicts: Predicted Thumb Down Predicted Palm (H), Predicted L, Predicted Fist (H), Predicted Fist (V), Predicted Thumbs up, Predicted Index, Predicted OK, Predicted Palm (V), Predicted C

GITHUB REPO

https://github.com/campusrover/gesture_recognition

Intro

As a very last minute and spontaneous approach, we decided to use a Leap Motion device. Leap Motion uses an Orion SDK, two IR camerad and three infared LEDs. This is able to generate a roughly hemispherical area where the motions are tracked.

It has a smaller observation area dn higher resolution of the device that differentiates the product from using a Kinect (which is more of whole body tracking in a large space). This localized apparatus makes it easier to just look for a hand and track those movements.

s### Hand Gesture Recognition After reaching the dead-end in the previous approach and being inspired by several successful projects (on Github and other personal tech blogs), I implemented an explict-feature-driven hand recognition algorithm. It relies on background extraction to "extract" hands (giving gray scale images), and based on which compute features to recognize the number of fingers. It worked pretty well if the camera and the background are ABSOLUTELY stationary but it isn't the case in our project: as the camera is mounted on the robot and the robot keeps moving (meaning the background keeps changing). Code can be founded

References

Intro

As a very last minute and spontaneous approach, we decided to use a Leap Motion device. Leap Motion uses an Orion SDK, two IR camerad and three infared LEDs. This is able to generate a roughly hemispherical area where the motions are tracked.

It has a smaller observation area dn higher resolution of the device that differentiates the product from using a Kinect (which is more of whole body tracking in a large space). This localized apparatus makes it easier to just look for a hand and track those movements.

The set up is relatively simple and just involved downloading for the appropriate OS. In this case, Linux (x86 for a 32 bit Ubuntu system).

Steps to downloading Leap Motion and getting it started:

1. download the SDK from https://www.leapmotion.com/setup/linux; you can extract this package and you will find two DEB files that can be installed on Ubuntu.

2. Open Terminal on the extracted location and install the DEB file using the following command (for 64-bit PCs):

   $ sudo dpkg -install Leap-*-x64.deb

   If you are installing it on a 32-bit PC, you can use the following command:

Once having Leap Motion installed, we were able to simulate it on RViz. We decided to program our own motion controls based on angular and linear parameters (looking at directional and normal vectors that leap motion senses):

This is what the Leap Motion sees (the raw info):

In the second image above, the x y and z parameters indicate where the leap motion detects a hand (pictured in the first photo)

This is how the hand gestures looked relative to the robot's motion:

Stationary

Forward

Backward

Left Rotation

Right Rotation

Conclusion

So, we got the Leap Motion to successfully work and are able to have the robot follow our two designated motion. We could have done many more if we had discovered this solution earlier. One important thing to note is that at this moment we are not able to mount the Leap Motion onto the physical robot as LeapMotion is not supported by the Raspberry Pi (amd64). If we are able to obtain an Atomic Pi, this project should be able to be furthered explored. Leap Motion is a very powerful and accurate piece of technology that was much easier to work with than the Kinect, but I advise still exploring both options.

GIT HUB REPO

https://github.com/campusrover/gesture_recognition

This still seems the best approach (if it ever worked). The Kinect camera was developed, at the first place,for human body gesture recognition and I could not think of any reason not using it but to reinvent wheels. However, since the library interfacing Kinect and ROS never worked and contains little documentation ( meaning it is very difficult to debug), I spent many sleepless nights and still wasn't able to get it working.However, for the next generation, I will still strongly recommend you give it a try.

ROS BY EXAMPLE V1 INDIGO: Chapter 10.9 contains detailed instructions on how to get the package openni_tracker working but make sure you install required drivers listed in chapter 10.3.1 before you start.

Dependencies Overview

• OpenCV

• ROS Kinetic (Python 2.7 is required)

• A Camera connected to your device

How to Run

1. Copy this package into your workspace and run catkin_make.

2. Simply Run roslaunch gesture_teleop teleop.launch. A window showing real time video from your laptop webcam will be activated. Place your hand into the region of interest (the green box) and your robot will take actions based on the number of fingers you show.

Brief Explaination towards the package

This package contains two nodes.

1. detect.py: Recognize the number of fingers from webcam and publish a topic of type String stating the number of fingers. I won't get into details of the hand-gesture recognition algorithm. Basically, it extracts the hand in the region of insteret by background substraction and compute features to recognize the number of fingers.

2. teleop.py: Subscribe to detect.py and take actions based on the number of fingers seen.

Later Plan

1. Using Kinect on mutant instead of local webcam.

2. Furthermore, use depth camera to extract hand to get better quality images

3. Incorporate Skeleton tracking into this package to better localize hands (I am using region of insterests to localize hands, which is a bit dumb).

Deep learning

The final approach I took was deep learning. I used a (Single Shot multibox Detector) model to recongnize hands. The model was trained on the Egohands Dataset, which contains 4800 hand-labeled JPEG files (720x1280px). Code can be found

Raw images from the robot’s camera are sent to another local device where the SSD model can be applied to recognize hands within those images. A filtering method was also applied to only recognize hands that are close enough to the camera.

This semester I mainly worked on the hand gesture recognition feature. Honestly I spent most of the time learning new concepts and techniques because of my ignorance in this area. I tried several approaches (most of them failed) and I will briefly review them in terms of their robustness and reliability.

As said before, we need a reliable way to extract hands. The first approach I tried was to extract hands by colors: we can wear gloves with striking colors (green, blue and etc) and desired hand regions can be obtained by doing color filtering. This approach sometimes worked fine but easily got distracted by subtle changes in lighting at most of the time. In a word, it was not robust enough. After speaking to Prof. Hong, I realized this approached had been tried by hundreds of people decades ago and will never work. Code can be found here.

Gestures

Author: Tirtho Aungon and Jade Garisch

Part 1: Kinect aka libfreenect2

Intro

In Kinect.md, the previous generations dicussed the prospects and limitations of using a Kinect camera. We attempted to use the new Kinect camera v2, which was released in 2014.

Thus, we used the libfreenect2 package to download all the appropiate files to get the raw image output on our Windows. The following link includes instructions on how to install it all properly onto a Linux OS.

https://github.com/OpenKinect/libfreenect2

Issues

We ran into a lot of issues whilst trying to install the drivers, and it took about two weeks to even get the libfreenect2 drivers to work. The driver is able to support RGB image transfer, IR and depth image transfer, and registration of RGB and depth images. Here were some essential steps in debugging, and recommendations if you have the ideal hardware set up:

• Even though it says optional, I say download OpenCL, under the "Other" option to correspond to Ubuntu 18.04+

• If your PC has a Nvidia GPU, even better, I think that's the main reason I got libfreenect to work on my laptop as I had a GPU that was powerful enough to support depth processing (which was one of the main issues)

• Be sure to install CUDA for your Nvidia GPU

Please look through this for common errors:

https://github.com/OpenKinect/libfreenect2/wiki/Troubleshooting

Although we got libfreenect2 to work and got the classifier model to locally work, we were unable to connect the two together. What this meant is that although we could use already saved PNGs that we found via a Kaggle database (that our pre-trained model used) and have the ML model process those gestures, we could not get the live, raw input of depth images from the kinect camera. We kept running into errors, especially an import error that could not read the freenect module. I think it is a solvable bug if there was time to explore it, so I also believe it should continued to be looked at.

However, also fair warning that it is difficult to mount on the campus rover, so I would just be aware of all the drawbacks with the kinect before choosing that as the primary hardware.

Database

https://www.kaggle.com/gti-upm/leapgestrecog/data

Machine Learning model

https://github.com/filipefborba/HandRecognition/blob/master/project3/project3.ipynb

• What this model predicts: Predicted Thumb Down Predicted Palm (H), Predicted L, Predicted Fist (H), Predicted Fist (V), Predicted Thumbs up, Predicted Index, Predicted OK, Predicted Palm (V), Predicted C

Part 2: Leap Motion: Alternative approach (semi-successful one)

Intro

As a very last minute and spontaneous approach, we decided to use a Leap Motion device. Leap Motion uses an Orion SDK, two IR camerad and three infared LEDs. This is able to generate a roughly hemispherical area where the motions are tracked.

It has a smaller observation area dn higher resolution of the device that differentiates the product from using a Kinect (which is more of whole body tracking in a large space). This localized apparatus makes it easier to just look for a hand and track those movements.

The set up is relatively simple and just involved downloading for the appropriate OS. In this case, Linux (x86 for a 32 bit Ubuntu system).

Steps to downloading Leap Motion and getting it started:

1. download the SDK from https://www.leapmotion.com/setup/linux; you can extract this package and you will find two DEB files that can be installed on Ubuntu.

2. Open Terminal on the extracted location and install the DEB file using the following command (for 64-bit PCs):

   $ sudo dpkg -install Leap-*-x64.deb

   If you are installing it on a 32-bit PC, you can use the following command:

Once having Leap Motion installed, we were able to simulate it on RViz. We decided to program our own motion controls based on angular and linear parameters (looking at directional and normal vectors that leap motion senses):

This is what the Leap Motion sees (the raw info):

In the second image above, the x y and z parameters indicate where the leap motion detects a hand (pictured in the first photo)

This is how the hand gestures looked relative to the robot's motion:

Stationary

Forward

Backward

Left Rotation

Right Rotation

Conclusion

So, we got the Leap Motion to successfully work and are able to have the robot follow our two designated motion. We could have done many more if we had discovered this solution earlier. One important thing to note is that at this moment we are not able to mount the Leap Motion onto the physical robot as LeapMotion is not supported by the Raspberry Pi (amd64). If we are able to obtain an Atomic Pi, this project should be able to be furthered explored. Leap Motion is a very powerful and accurate piece of technology that was much easier to work with than the Kinect, but I advise still exploring both options.