Row Detection for Agriculture Robot using RIU-Net

This repo contains python codes that detects traversible row for agriculture robots
using UNet. This work largely follows the paper
RIU-Net: Embarrassingly simple semantic segmenta-tion of 3D LiDAR point cloud
Note that the very same pipeline could be used for generic object detection, while this
repo only concerns about detecting a single row without row switching.

The purpose of this work is to enable autonomous, visual-based in-row navigation for
agriculture robots. Specifcially, to address the occassions where GPS are not reliable
or not accurate enough. The training data were collected from a vineyard field using
Velodyne's VLP-16, which was mounted on a mobile agriculture robot. The dataset is not
shared in this repo because the author does not own them.

Run with Visualization

• single model

   python3 inference.py --m ./weight/unet.pth --show "all"
  

• multiple model

   python3 inference.py --m ./weight/*.pth --show "all"
  

Input of Unet

This work was implemented with 150 training images (augmented to 1050). The model takes
2-channel inputs of range images (64x512):

• relative distance of a point to the center of LIDAR
• intensity capture by LIDAR

Note that in this work a limited horizontal filed of
view [-180, 180] is considered.

Example output

first row : output from network
second row: deprojection into topview (sensor frame)

Tricks to Improve Performance

• The data captured by VLP-16 originally corresponds to 16x512 image, which is too small
  for convulution layers. Thus each beam was repeated 4 times without extrapolation to
  make the image 64x512.

• Futhermore, since the image has a larger width then width, the maxpool kernel is adjusted
  to (2,4) after first layer.

• To improve prediction and get a clear-cut boundary, weighted mask similiar to that of oringal
  UNet paper was developed and adopted to penalize loss more on pixels near boundaries.

  

Limitations

The network was mostly trained data facing forward, with vehicle standing in the center.
Thanks to the property of spherical coordination, a rotated pointcloud is simply equiva-
lent to tranlated image which should pose no difference to convolutional netwroks. It
is observed, though, at entrance and exit of a row, the model does deterioate in terms
of performance. Such limitation may be able to be overcomed with more training data.