The framework for visual place recognition in changing enviroments. Matches two sequence of images of arbitrary trajectory overlap.
Switch branches/tags
Nothing to show
Clone or download
Latest commit 1539332 May 23, 2018
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
apps fixed issues with loading features and directories Nov 24, 2017
doc added motivational image Nov 8, 2017
examples added files Nov 6, 2017
gtest added files Nov 6, 2017
src feature files can have arbitrary names Nov 14, 2017
test fixed issues with loading features and directories Nov 24, 2017
.clang-format added files Nov 6, 2017
.gitignore added files Nov 6, 2017
.gitlab-ci.yml added files Nov 6, 2017
CMakeLists.txt added files Nov 6, 2017
LICENSE.txt added files Nov 6, 2017
readme.md Update readme.md May 23, 2018

readme.md

Visual place recognition using image sequences with Relocalization

What does this code do?

Performs matching of two image sequences of arbitrary shaped trajectories, e.g. the query trajectory may deviate from reference one. This work is an extension of our work on image sequence matching under changing visual appearances. The previous approach was dealing well with trajectories that roughly follows the same route. You can find the code by following this link online place recognition.

In this framework, the system is able to deal with trajectories of arbitrary shapes:

  • by performing relocalization step, whenever the robot is considered lost, e.g. the query trajectory deviates from the reference one.
  • by taking into account the information about the similar places within reference trajectory to ensure the multiple localization hypothesis.

image sequences from different routes and weather conditions

In Figure the images from the blue trajectory are matched against the images of orange trajectory. Blue trajectory (query) follows partially the orange trajectory (reference) as well as deviating from it. These kinds of challenges in matching images sequences are targeted in this framework. (Trajectories are plotted over Google Maps for visualization purposes. The GPS coordinates are not used for place recognition.)

For more theoretical details, please refer to our paper Relocalization under Substantial Appearance Changes using Hashing.

Installation guide

Prerequisites

  • Yaml-cpp (requires BOOST till now): sudo apt-get install libyaml-cpp-dev libboost-all-dev
  • OpenCV: sudo apt-get install libopencv-dev
  • Qt5: sudo apt-get install qt5-default
  • (optional) Doxygen (generate documentation): sudo apt-get install doxygen

Build

The code is tested on Ubuntu 16.04 and 18.04. To build the project, run the following commands from the main directory:

  • mkdir build
  • cd build
  • cmake ..
  • make -j4

Additionally, you should be able to generate documentation as follows:

  • cd doc
  • doxygen online_place_recognition.conf

To access the documentation run firefox html/index.html.

If you want to build on Ubuntu 14.04, you may run into issues with yaml-cpp.

What do I need to run this code?

  • Precomputed image descriptors (features) or cost matrix
  • Configuration file
  • Precomputed hash table
  • (optional) Estimated similar places within the reference sequence

For the examples of how to run the code please see RUN EXAMPLES.

Feature based matching

In this mode, the program operates using precomputed image descriptors. To run the matching procedure you need to provide the feature files. An example of how to run the code can be found feature based matching example.

For details on used feature descriptors please refer to feature description.

Note: In this mode, individual features will be loaded and matched on demand. In order to be able to deal with dramatic visual changes, we typically operate with high-dimensional features and the matching procedure can take quite a long time--depending on the size and the complexity of the sequences.

Cost matrix based matching

For this mode, we require the cost matrix between two sequences to be given/precomputed. To compute the matching matrix, please see the following estimating of a cost matrix.

An example on how to run the matching procedure in this mode can be found cost matrix matching example.

Note: This method may be used if you have rather small sequences (up to 1000 images). For bigger sequences, you may run into memory issues since the programs has to store quite a big matrix.

How to run this code on own dataset?

1. Extract features

In our code, we provide the way to use features from OverFeat and VGG-16.

To use overfeat: Go to OverFeat Github, follow the installation instructions and extract features using the binary as in the example: bin/linux_64/overfeat -f -l samples/pliers.jpg. Store the features from individual images in the separate files.

To use VGG-16 : Go to Keras VGG and extract features with VGG16. Store the features from individual images in the separate files.

2. Hash reference features

Hash the features of reference sequence. Use hash app for obtaining the '.txt' file with a hash table.

3. (Optional) Estimate similar places within reference sequence

To improve the localization performance, you may include the notion about the similar places in the reference trajectory. Use estimate similar places app.

4. Adapt the config file

Change the config.yaml file with your new input files.

Now, you ready to run the code! Try feature based matching!

😬 Didn't work? Leave an issue.

Adapting the code for custom features

This framework can be adapted to matching features of the different type --- your own features. To use the graph matching strategy with your own features see the following description.

Parameters

The code relies on couple of parameters. Take a look at parameters desciption to get the best code experience.

Related publication

Please cite the related publication, if you use the code:

@inproceedings{vysotska2017irosws,
	title     = {Relocalization under Substantial Appearance Changes using Hashing},
	author    = {O. Vysotska and C. Stachniss},
	booktitle = {9th Workshop on Planning, Perception and Navigation for Intelligent Vehicles at the IEEE } # IROS,
	year      = {2017},
}

Troubleshooting

In case the code is not working for you or you experience some code related problems, please consider opening an issue.