Multi-View Monocular Visual Odometry

Monocular Visual Odometry Project using Conventional Multi-view Geometry with OpenCV and Python

---

pyrealsense Jetson Nano Setup

Reference
: Official Intel Realsense Python Wrapper (https://github.com/IntelRealSense/librealsense/tree/master/wrappers/python)
: Fix by import (IntelRealSense/librealsense#7540 (comment))

Since Jetson Nano is ARM-based device, pyrealsense2 cannot be installed using pip. As a result, python wrapper of Intel Realsense SDK needs to be built from the source.

1. Ensure apt-get is up to date

• sudo apt-get update && sudo apt-get upgrade

2. Instally Python and its development files via apt-get (Python2 or Python3)

• For Python2 : sudo apt-get install python python-dev
• For Python3 : sudo apt-get install python3 python3-dev

3. Check python installation path

• For Python2 : which python (Copy the path)
• For Python3 : which python3 (Copy the path)

4. Download librealsense & Move to librealsense directory

• git clone https://github.com/IntelRealSense/librealsense.git
• cd librealsense

5. Prepare build

• mkdir build
• cd build

6. Run top level CMake command with the following flags

• cmake ../ -DBUILD_PYTHON_BINDINGS:bool=true -DPYTHON_EXECUTABLE=[full path of which python/which python3]

Python3 ex : cmake ../ -DBUILD_PYTHON_BINDINGS:bool=true -DPYTHON_EXECUTABLE=/usr/bin/python3 -DBUILD_PYTHON_BINDINGS : Decide whether to build python wrapper
-DBUILD_EXECUTABLE : Specify installation directory

7. Make & Install

• make -j1
• sudo make install

make -j1 : Use one core to build the library make -j4 : Use four core to build the library

8. At python script, append /usr/local/lib/python3.6 (Python 3.6) or /usr/local/lib/python2.7 (Python 2.7) in sys.path and import pyrealsense2.pyrealsense2 as rs

import sys

sys.path.insert(0, '/usr/local/lib/python3.6')
print(sys.path)

import pyrealsense2.pyrealsense2 as rs

pipe = rs.pipeline()
pipeline = rs.pipeline()
config = rs.config()
config.enable_stream(rs.stream.color, 640, 480, r$

pipeline.start(config)

For an unknown reason, pyrealsense2 is installed at /usr/local/lib/python3.6 for python3 and /usr/local/lib/python2.7 for python2. This does not change even if -DBUILD_EXECUTABLE is changed.

As a result, it is recommend to append /usr/local/lib/python3.6 for python3 and /usr/local/lib/python2.7 for python2 as an easy solution.

Updating PYTHONPATH environment variable (export PYTHONPATH=$PYTHONPATH:/usr/local/lib) is an alternative method for importing pyrealsense2. However, in Jetson Nano, this does not seem to work well with this method.

---

pytorch & torchvision Jetson Nano Setup

Reference : Installing PyTorch, torchvision, spaCy, torchtext on Jetson Nanon [ARM] (https://gist.github.com/heavyinfo/8d0ef5a3768c5d8d22e164863670330a) : Jetson Zoo PyTorch (https://elinux.org/Jetson_Zoo#PyTorch_.28Caffe2.29)

Since Jetson Nano is ARM-based device, pytorch and torchvision cannot be installed using pip. As a result, pytorch and torchvision have to be built from the source.

1. pytorch setup

• Visit Jetson Zoo (https://elinux.org/Jetson_Zoo)
• Select PyTorch that you want and download it
• Follow the instruction and run PyTorch wheel using pip (pip3 install torch-1.X.X-cp36-cp36m-linux_aarch64.whl)

2. torchvision setup

• Install pre-requisites (sudo apt-get install libjpeg-dev zlib1g-dev)
• Clone torchvision of the version you want (git clone --branch v0.X.0 https://github.com/pytorch/vision torchvision)
• Move to torchvision directory
• Run setup.py (sudo python setup.py install)