Possibily the simplest example of loop closure for Visual SLAM. More information on my blog.
As I'm experimenting with alternative approaches for SLAM loop closure, I wanted a baseline that was reasonably close to state-of-the art approaches. The approach here is pretty similar to ORB-SLAM's, and uses SURF descriptors and bag of words to translate them to a global image description vector.
For testing, I've used the New College dataset published alongside FAB-MAP.
It's available for download
here. It's
ideal for loop-closure testing, since it includes manual place associations
that can be used for evaluation. The scripts/download_data.sh will
download the data files (bag of words vocabulary and images) needed to run
the code.
You can build and run the code using docker-compose and Docker. The Docker
configuration uses a Ubuntu 16.04 base image, and builds OpenCV 3 from source.
# Download the data files
./scripts/download_data.sh
# Will take ~10 minutes to download and build OpenCV 3
docker-compose build runner
# Enter the docker shell
docker-compose run runner bash
# You're now in a shell inside the Docker container, build and run the code:
./scripts/build.sh
./build/new_college ./data/brief_k10L6.voc.gz ./data
Only tested on Ubuntu 16.04 LTS with OpenCV3, gcc 5.4.0
The ground_truth_comparison.py plots and compares the loop closures from the
ground truth to the actual results from the code.