Object-Oriented Semantic Mapping This repository contains the accompanying code for the paper Online Object-Oriented Semantic Mapping and Map Updating by N. Dengler and T. Zaenker and F. Verdoja and M. Bennewitz accepted at EMCR, 2021. you can find the paper at https://arxiv.org/pdf/2011.06895.pdf
- download pretrained neural net
"faster_rcnn_inception_resnet_v2_atrous_oidv4"
fromhttps://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md
- place it in
tensorflow_object_detector/data/models/
- change camera and pointcloud topic in
tensorflow_object_detector/launch/object_detection.launch
- To use it tensorflow has to be installed follow step 3 of installation readme.md in the
tensorflow_object_detector
folder
- the map of the environment has to be loaded
- amcl or other localizer has to run
- build the repo in release mode:
catkin config --profile release -x _release --cmake-args -DCMAKE_BUILD_TYPE=Release && catkin build --profile release
- run tensorflow:
source ~/tensorflow/bin/activate && roslaunch tensorflow_object_detector object_detect.launch
- run mapping:
rosrun semmapping mapping _camera_info:=YOUR/CAMERA/INFO/TOPIC
Some Params ar configureable with a rqt configure server. You can change the default value in /cfg/Params.cfg and change the value with rosrun rqt_reconfigure rqt_reconfigure
.
important Params are:
- queue_size: size of Queue which is used for the object likelihood. If the que limit is reaced the oldest value will be poped for a new one
- certainty_thresh: If the obj certainty is under the threshold, the obj will be removed
- queue_thresh: Only remove obj if Queue size > queue_thresh
- view_min: obj cen which are < view_min will not be removed
- view_max: obj cen which are > view_max will not be removed
Please cite this work as:
@InProceedings{Dengler21ecmr, author = {N. Dengler and T. Zaenker and F. Verdoja and M. Bennewitz}, title = {Online Object-Oriented Semantic Mapping and Map Updating}, booktitle = {Proc.~of the European Conference on Mobile Robots (ECMR)}, year = 2021 }