search

gisbi-kim / FAST_LIO_SLAM

LiDAR SLAM = FAST-LIO + Scan Context
494 stars 74 forks source link

How to tune parameters for indoor environment? #13

Closed yanliang-wang closed 3 years ago

yanliang-wang commented 3 years ago

Hello, @gisbi-kim this work is very cool.

But it does not work well in my indoor test. My sensors are Velodyne16 and Xsens Mti-300. I adjust some parameters(reduce sc_dist_thres to 0.2, reduce sc_max_radius to 20) , but it still does not work. image

Can you give me some advice about how to tune parameters in my indoor test? I'll appreciate it. Thanks in advance!

gisbi-kim commented 3 years ago

Hi, @yanliang-wang thanks for using our project. I think the parameter is good for the indoor env.

but the current version of SC-PGO in this repository does not support the reverse-loop-closing. (Actually Scan Context supports reverse loop detections, but only for when 360 deg horizontal fov is available. In common cases, a robot loses it rear or some horizontal fovs so I just ommited the reverse loop closings using scan context)

For the reverse loop detections and closings, I've been using radius-search-based search. the radius-search-based loop closing was not implemented in this SC-PGO yet, but in SC-LIO-SAM or SC-LeGO-LOAM - please use them.

yanliang-wang commented 3 years ago

Thanks, that makes much sense. I'll have a try. Thank you for this work again.

yanliang-wang commented 3 years ago

Hi, @gisbi-kim. Thanks for your reply again. That helped a lot. I tried radius-search-based loop closure detection and that worked very well for reverse-loop-closing.

By the way, scan context is also a nice work. Thanks for your share.

marshal-shi commented 2 years ago

Hi @yanliang-wang Great question. @gisbi-kim Thanks for your great work.

May I ask there is any sample radius search implementation for SC-PGO here? both SC-LIO-SAM and SC-LeGO-LOAM are deeply integrated radius search

Thanks.

gisbi-kim commented 2 years ago

@marshal-shi I think the radius-search implementation is straightforward, but I may have enough time to get it after Dec, 2021.

or this is a trick, you can use the LT-SLAM (originally designed for multi-session (two-session currently)) by setting the central and query session as equal. Then, the both SC search and radius search were implemented there. please see https://github.com/gisbi-kim/lt-mapper/tree/main/ltslam

marshal-shi commented 2 years ago

Hi @gisbi-kim Thanks for your reply. I will check the repo which you shared.

I have implemented a version of RS yesterday. Still on testing. But looking forward to your version which will be great.

Thanks.

gisbi-kim commented 2 years ago

@marshal-shi happy to hear that. I always thank you guys for having an interest in my project and makes it better. thanks.

marshal-shi commented 2 years ago

Super great job @gisbi-kim , well done. Thanks for your contribution. Looking forward to your new projects.

Also, hopefully I could contribute some as well after catching up.

yanliang-wang commented 2 years ago

Hi, @gisbi-kim @marshal-shi , I released my implementation of the radius-search-based loop closure in the repository, in which the pose and map in the iterated extended Kalman filter of FAST-LIO will be updated according to the pose graph optimization. This is different from the implementation of FAST_LIO_SLAM and I think this will improve its performance.

You can have a try if you are interested in it.

marshal-shi commented 2 years ago

@yanliang-wang well done. Just checked the video. The result looks promising. Will check code later. Well done.

gisbi-kim commented 2 years ago

@yanliang-wang cool!!! thanks! I'll link your repo within the readme.

yanliang-wang commented 2 years ago

@yanliang-wang cool!!! thanks! I'll link your repo within the readme.

Wow, it's my pleasure. Thanks for your interest in it.