0
$\begingroup$

Rosanswers logo

I represent a team of engineers from Lancaster University. We are attempting to combine 2D LIDAR data (preferably using HectorSLAM) and RGB-D camera data (as done by Technische Universität Darmstadt https://youtu.be/olGZv05RLHI) for an autonomous UAV mapping application. We are using AND RPLIDAR A2 scanner and a Realsense Depth Camera D415. The ROS distro is Kinetic. How could we achieve this and could it be performed using ROS on an Nvidia Jetson Nano? Can we run two SLAM algorithms concurrently (e.g HectorSLAM and ORB-SLAM), or do we need to combine the sensor data before applying SLAM? Is there any open-source code available to achieve this?

Many Thanks.


Originally posted by SamH on ROS Answers with karma: 27 on 2020-02-03

Post score: 2

$\endgroup$

1 Answer 1

0
$\begingroup$

Rosanswers logo

There are many ways to approach this problem. I'll outline the simplest one, but your comment about:

to combine the sensor data before applying SLAM

will probably give you a better result. Or approaching this from a tightly-coupled approach, if that's terminology you're familiar with.

The only way to my knowledge that's fully-open source and relatively plug and play is as follows:

  1. Use the 2D laser scanner to build a map. This can be done with Hector like you mention, but also slam toolbox, karto, or gmapping.

  2. Look at Octomap. Use the positioning provided by the slam algorithm and odometry to project your points of your depth sensor into the global frame provided by the 2D slam

  3. Rejoice!

Obvious asterisks:

  • While this is a popular method for junior developers and folks that don't want to actually create a SLAM solution, there are clear downsides.

  • If you're working with a 2D laser scanner, then you're throwing out a ton of data that could be used to build a better map and position yourself, and only just using that 3D information to build the global model. Calling this 3D SLAM is a bit of a misnomer, but again, its a popular method.

  • To increase fidelity, you may need to continuously post the graph to octomap to update the positioning of individual measurements if they shift around. This is necessary for loop closure and reduction of residual error operations. Hector doesn't do loop closures, so that may not be an issue you have the ability to resolve if you're married to Hector.


Originally posted by stevemacenski with karma: 8272 on 2020-02-03

This answer was ACCEPTED on the original site

Post score: 2


Original comments

Comment by SamH on 2020-02-03:
Thank you for your fast and comprehensive response. This gives us everything we need to get started.

Can I clarify what you mean by "Use the positioning provided by the slam algorithm and odometry to project your points of your depth sensor into the global frame provided by the 2D slam".

How do you specifically suggest we combine the Odometry data with the Octomap and then combine this with the global frame provided by the 2D SLAM? Is this just a case of finding appropriate open-source code on Github? Thanks again for your help!

Comment by stevemacenski on 2020-02-03:
Awesome, please mark the answer as correct to get it off the unanswered questions queue.

Basically Hector / SLAM will give you a pose estimation in TF. You should use octomap to take in sensor readings in their current frame (camera_frame, or something) and transform into the global frame (map, or something) to insert into the octomap occupancy grid.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.