Aerial Vehicle Localization Using a Downward-Facing Depth Camera for Precision Agriculture

At TIH, IIT Bombay, I evaluated visual-inertial odometry (VIO) systems for UAV pose estimation in precision agriculture. VIO combines visual information from a camera with high-rate motion measurements from an IMU, enabling pose estimation in GNSS-denied environments.

Agricultural environments present particular challenges for VIO due to changing illumination, highly self-similar crop textures, unstructured and dynamic objects, and irregular terrain that can produce more aggressive camera motion than typical indoor or urban environments.

I simulated agricultural environments in Gazebo and evaluated a feedback- based visual-inertial system (FVIS) for estimating the position and orientation of an Iris quadcopter equipped with an Intel RealSense depth camera.

The implementation was evaluated in three simulated environments:

1. Manhattan world with rich and diverse visual features
2. Structured farm with a regular grid arrangement
3. Unstructured farm with a randomly generated arrangement

The experiments compared the drone's true flight path and velocity profile with the trajectory and speed estimates generated by the VIO-based localization system at flight heights of 3.5 m, 10 m, and 20 m.

<td>
  <figure>
    <img src="mh_vel_plot.png"
         alt="Manhattan world velocity plots">
    <figcaption>
      Manhattan world: velocity plots at different heights
    </figcaption>
  </figure>
</td>

<td>
  <figure>
    <img src="fg_vel_plot.png"
         alt="Farm grid world velocity plots">
    <figcaption>
      Farm grid world: velocity plots at different heights
    </figcaption>
  </figure>
</td>

<td>
  <figure>
    <img src="fr_vel_plot.png"
         alt="Random farm world velocity plots">
    <figcaption>
      Farm random world: velocity plots at different heights
    </figcaption>
  </figure>
</td>