Color Nearest Neighboor Segmenter

Author: Felix Jonathan (fjonath1@jhu.edu)

This rosnode will perform a color based segmentation for objects with flat color. The color training is performed using k-means clustering of each object color category from their color point clouds. Then the model is used for perfoming nearest neighboor search on input color point cloud.

Requirements

This package requires these dependencies:

• PCL 1.7.2+
• OpenCV 2.4.x
• ROS Indigo
• object_on_table_segmenter

This package has been tested on 14.04 using PCL 1.7.2, OpenCV 2.4.8, and ROS Indigo.

Usage

Run this node using roslaunch with:

roslaunch color_nn_segmenter segmenter.launch

Roslaunch Parameters

The default ros parameter use this following layout for training and saving data

$(catkin source directory)/color_nn_segmenter
├── include
├── launch
├── models
│   ├── model.dat
│   ├── model.pcd
│   └── ...
├── src
└── training_data
    ├── color_category1
    │   ├── file1.pcd
    │   ├── file2.pcd
    │   └── ...
    ├── color_category2
    │   ├── file1.pcd
    │   ├── file2.pcd
    │   └── ...
    └── ...

These are the required parameters for loading the model:

• load_existing_model: Loads the color model saved specified in the model path. If set to false, model training will be done before color segmentation is performed;
• model_directory: Directory that contains the model data;
• model_name: The name of the model to be loaded/saved.

These are the required parameters for saving the model:

• save_new_model
• training_data_directory: Location of the training data that contains color category folders
• kmeans_point_per_model: The number of kmeans points per color category
• save_directory: Target directory for saving the new model

To specify the point cloud input/output topic, these parameters can be set:

• cloud_input_topic: Name of the point cloud topic to be segmented
• segmented_cloud_topic: Name of the topic to publish segmented cloud that contains PointXYZL
• visualized_cloud_topic: Name of the topic to publish segmented cloud in rviz-compatible point cloud

This program also include a way to perform segmentation on a volume above a convex hull/table. This convex hull is generated by searching largest region around a tf frame based on object_on_table_segmenter.