Visual Odometry using Classical Computer Vision

RBE 549: Computer Vision - Worcester Polytechnic Institute, Fall 2021

Team Megatron (Team M)

Members: Aniket Patil, Chinmay Todankar, Nihal Navale, Prathamesh Bhamare

---

What is Visual Odometry?

Visual odometry (VO) is the process of determining the position and orientation (ego-motion) of a robot/agent by analyzing images taken from a monocular or stereo camera system attached to the robot/agent. Visual Odometry operates by estimating the pose of the robot/agent by analyzing the changes that motion induces on the images of its onboard cameras.

Our Implementation:

Requirements:

1. Ubuntu with VSCode for best implementation (Open this cloned repository as a folder in VSCode)
2. OpenCV (C++): Steps for Installation
3. GNUplot: sudo apt install gnuplot (Optional: Only to generate plots from the .dat files)

IMPORTANT: You will also need the KITTI Dataset of grayscale sequences. Download

Once downloaded, extract the Sequences folder into the Dataset folder such that the structure is like:

• dataset
  • poses
  • sequences

How to run the code?

1. If running from VSCode, open the folder Visual-Odometry in the VSCode navigator. Then open the main.cpp file in the code editor and press Ctrl + Shift + B to build the code. Then run the following command in the terminal below:

./src/build/main 00

Replace 00 with sequence number to run other sequences

2. If running from terminal, go to the parent folder of this repo, that is, Visual-Odometry and enter the command:

g++ -g src/main.cpp -o src/build/main `pkg-config --cflags --libs opencv4`

Our Results (Trajectory images and error plots):