TensorFlow implementation of Multiplicative Gaussian Particle Filter. We apply MGPF on robot localization as a proof-of-concept.

Xuan Su, Wee Sun Lee, and Zhen Zhang: Multiplicative Gaussian Particle Filter. The 23rd International Conference on Artificial Intelligence and Statistics (AISTATS 2020). https://arxiv.org/pdf/2003.00218.pdf

Pipeline for extracting Gaussians using depth image. (1): simulated laser scan, followed by edge detection; (2): rotation; (3): convolution over wall map; (4): Gaussian extraction.

Python 2.7, Tensorflow 1.5.0

Additional packages are specified in the yml file, to be installed using conda. Configurations in the yml file need to be set properly before installation.

mgpf-env.yml 

The dataset for robot localization experiments is available at:

https://drive.google.com/open?id=1v7m1B8H5N7avcFmdrDl7DtsMbPLcHamk

The folder contains only one file: wallsonly_test.tfrecords. Download and save into the data folder.

To produce experimental results in the paper, run e.g.

python evaluate_aistats.py -c ./configs/eval-tracking.conf

replacing the configuration file as appropriate.

Results used in the paper are also included in their respective folders.

Our code is written largely based on a previous framework, the particle filter network (PF-Net): https://github.com/AdaCompNUS/pfnet. Credits to Peter Karkus.

Xuan Su <suxuan (at) comp.nus.edu.sg>