[Documentation] [TitleIndex] [WordIndex

Only released in EOL distros:  

Overview

This package provides several utilities for environment perception and modelling from RGB-D sensor data, e.g. Kinect device.

System architecture

The package consists of three major components:

Depth map segmentation

The depth map segmentation node's main purpose is to divide Kinect depth image into several regions of interest. It is supposed to be the preprocessing task for 3D environment creation but also for the other possible detectors (object detection etc.). For depth map segmentation itself, several algorithms are possible to use:

Depth based segmentation

The depth based segmentation method is the fastest segmentation method due to the nature of segmentation – simple depth difference of neighboring pixels. Even if this is not proper plane detector, the usability as a preprocessing unit it recommended due to very high efficiency. The scene is divided to multiple depth differing segments, which pose a very good input for plane detectors.

Normal based segmentation

Normal based segmentation method uses similar principle as depth based method with the difference that this algorithm uses normal difference for gradient image creation.

Depth and normal fused segmentation

Having the outputs from depth based and normal based segmentation methods, the later fusion can be applied to produce more accurate and stable final results.

Plane prediction segmentation

Plane prediction method emerges from the original need for plane detector pre-processing algorithm. It applies the same method as segmentation above – gradient image followed by watershed segmentation method, however the process of gradient map computation is different. The gradient map computation has two outputs – edge strengths gradient image and number of changes gradient image .

Tiled RANSAC segmentation

Last of implemented methods approaches to the plane segmentation from a different way. Planes are searched using RANSAC search. Tiled RANSAC has proven to be fast enough to be used in real-time systems because of small random sample search. On the other hand, the possibility of filling regions outside the tile borders ensures that found planes are marked on the whole depth image.

Plane Fitting

Plane fitting component constructs iteratively the knowledge of surrounding scene by creating plane segment information. Plane detection itself is constructed by analysing Hough Space created by iteratively adding Hough transform information from several frames.

Altough the Hough transform main disadvantage is high memory requirements, our method has overcame these drawbacks by implementation of hierarchic adaptive structure, which keeps memory requirements as low as possible while speeding up the computation.

The output of this node is a list of detected planes sent in a message. Each plane is defined in two ways - plane equation and bounding rectangle (in 3D) for visualisation purposes.

ROS API

Nodes

Node Name

Published Topics

Description

but_segmenter

/but_env_model/seg_region_image;/but_env_model/seg_deviation_image

Node implementing depth image segmenter. Publishes region image (image with region index information) and sometimes deviation image (image with computed std. deviations of detected planes). Please see below when these topics are published.

but_plane_detector

/but_env_model/plane_array

Node implements described plane detector. It publishes a plane array message, which is represented by the list of found planes.

Services

Service Name

Input

Output

Description

/but_env_model/clear_planes

-

Message about the result

Called service clears the accumulator of planes and Hough Space, which is constructed iteratively from all arriving point clouds. After calling this service the node is reset and continues detecting planes from scratch.

Published topics

Topic Name

Message

Description

/but_env_model/seg_region_image

sensor_msgs::Image

A 16bit unsigned short image of segmented indices. Indices start with index 1, if index is less than 1, it signifies a border or not segmented region. It is published by every segmentation method.

/but_env_model/seg_deviation_image

sensor_msgs::Image

A region image with computed depth std. deviations from plane approximated through every region. It is NOT published by Depth segmentation method (Normals are not computed due to speed reasons).

/but_env_model/plane_array

srs_env_model::InsertPlanes

Plane array of srs_env_model_msgs::PlaneDes with each plane information. Each plane has INSERT/MODIFY/DELETE flag, which specifies action.

Installation

All components are in srs git in srs_env_model_percp package and can be compiled with ROS standard tool rosmake

 rosmake srs_env_model_percp

command

Configuration

Optional parameters for each node. Attention, since this module is flagged as under construction, several parameters are not possible to set up yet.

Depth map segmentation

List of but_segmenter's optional parameters:

Plane Fitting

List of but_plane+detector's optional parameters:

Dependencies

Usage

A few examples of starting described nodes:

rosrun srs_env_model_percp but_segmenter

rosrun srs_env_model_percp but_segmenter -type depth

rosrun srs_env_model_percp but_segmenter -maxdepth 4500

rosrun srs_env_model_percp but_plane_detector -input pcl

rosrun srs_env_model_percp but_plane_detector -input pcl -target /world


2022-05-28 13:03