laskama / mcelindoorloc Goto Github PK

This repository holds the implementation of the neural network model for multi-task indoor localization (building/floor/position) estimation in a single forward pass, which was proposed in:

[1] M. Laska and J. Blankenbach, "Multi-Task Neural Network for Position Estimation in Large-Scale Indoor Environments," in IEEE Access, vol. 10, pp. 26024-26032, 2022, doi: 10.1109/ACCESS.2022.3156579.

Pedestrian localization within large-scale multi-building/multi-floor indoor environments remains a challenging task. Fingerprinting-based approaches are particularly suited for such large-scale deployments due to their low requirements of hardware installments. Recently, the fingerprinting problem has been addressed by deep learning. Existing models are mostly task specific by providing floor classification or position estimation within a small area. A strategy to support localization within large-scale environments is to sequentially apply hierarchical models. This has several drawbacks including missing scalability and increased deployment complexity on smartphones. We propose a unifying approach based on training a single neural network that classifies the building/floor and predicts the position in a single forward-pass of the network. Our model classifies a grid cell and performs within grid cell regression, which solves the performance degradation of applying regression within large areas. To reduce the error in case of misclassified grid cells, we propose a novel technique called multi cell encoding learning (multi-CEL), where a model simultaneously learns several redundant position representations within an overlapping grid cell encoding. For three public WLAN fingerprinting datasets, we demonstrate that multi-CEL surpasses existing state-of-the-art multi-task learning neural networks and even outperforms regression neural networks explicitly trained for 2D-positioning by up to 17%.

The network simultaneously classifies a grid cell and perform within grid cell regression to obtain a final position estimate. This neural network architecture for this is as follows:

To prevent large errors in case of failures of the grid cell classification head, we introduced the multi-cell encoding learning (m-CEL) technique. The network is supplied with several redundant encodings by letting the encoding grid slightly overlap.

Clone the repository and install the required packages as listed in the requirements.txt file. Installation via any virtual environment such as virtualenv or conda are is strongly recommended.

The repository allows for training all models on various datasets. It provides access to the following datasets:

• giaIndoorLoc [2] (available DOI:10.5281/zenodo.6801310)
• Tampere dataset [3] (available here)
• UJI dataset [4] (available here)
• UTS dataset [5] (available here)

The downloaded datasets have to placed in the corresponding datasets/{dataset}/ folders. Please inspect the provided example config files inside the config folder for details on how to select the downloaded datasets for model training.

To train the mCEL network, a config.yml file has to be set up. Example files are contained in the config folder. Any parameter not specified in the config.yml will be taken from the default values given in config/default_params.py. Afterwards the pipelines specified within the config file can be executed via

python pipeline.py -c path/to/config.yml

The predictions of the models can be visualized and displayed along with the ground truth position for each fingerprint of the test partition of the dataset. Just add the flag --visualize_predictions when executing a pipeline as

python pipeline.py -c path/to/config.yml --visualize_predictions

For each prediction a plotting window such as
will open. Once the window is closed, the predictions for the next fingerprint of the test dataset will be shown and so on. Visualization is only supported for the datasets that include a floorplan, which are the giaIndoorLoc and the Tampere dataset. The floorplan images of the Tampere dataset have been extracted from the original article [3] that was published under Creative Commons Attribution License and are already included in the repository.

In order to reproduce the results from the VI-SLAM2tag paper [2], the pre-trained model weights have to be downloaded DOI:10.5281/zenodo.6801310, place the exp folder within the root of the directory and execute the script run_exp.sh. You can also set the pretrained: False in the default_params.py file and train the model from scratch. Note that this might result in slightly different results as the training is dependent on the hardware and the tensorflow version. The pretrained models have been obtained via tensorflow==2.8.0 trained on a Macbook Pro with M1Pro chip (CPU-only).

To reproduce the results of the original paper run the experiments within the config/{dataset} for each dataset. Note that this only provides the results of proposed model and not those of the reference SIMO architecture proposed by Kim et al. [6], since this would bloat the repository. However, open-source implementations of their architecture can be found in their github repository, which we have used to reproduce the results. Further, this repository does not cover the 2D-CNN-based backbone implementation for the same reasons. Note that the results might slightly differ from those reported in the paper depending on the hardware the model was trained on. The experiments reported in the paper have been obtained by training on an NVIDIA GTX 1050 Ti and Tensorflow version 2.3.

The repository has been designed to also provide easy integration with other datasets and models. You have to provide the following:

• Dataset connector: which extracts the data and stores it in the specified format of the repository (please see source file documentation for details)
• Data provider: transforms the raw data of a dataset connector to the format required by the model
• Model: has to implement setup and evaluation function and might overwrite the fit function.

Please study the documented source code for details.

[1] M. Laska und J. Blankenbach, „Multi-Task Neural Network for Position Estimation in Large-Scale Indoor Environments“, IEEE Access, Bd. 10, S. 26024–26032, 2022, doi: 10.1109/ACCESS.2022.3156579.

[2] M. Laska et al., „VI-SLAM2tag: Low-Effort Labeled Dataset Collection for Fingerprinting-Based Indoor Localization“, to appear at IPIN 22.

[3] E. Lohan et al., „Wi-Fi Crowdsourced Fingerprinting Dataset for Indoor Positioning“, Data, Bd. 2, Nr. 4, S. 32, Okt. 2017, doi: 10.3390/data2040032.

[4] J. Torres-Sospedra et al., „UJIIndoorLoc: A new multi-building and multi-floor database for WLAN fingerprint-based indoor localization problems“, in 2014 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Busan, South Korea, Okt. 2014, S. 261–270. doi: 10.1109/IPIN.2014.7275492.

[5] X. Song et al., „A Novel Convolutional Neural Network Based Indoor Localization Framework With WiFi Fingerprinting“, IEEE Access, Bd. 7, S. 110698–110709, Aug. 2019, doi: 10.1109/access.2019.2933921.

[6] K. S. Kim, „Hybrid Building/Floor Classification and Location Coordinates Regression Using A Single-Input and Multi-Output Deep Neural Network for Large-Scale Indoor Localization Based on Wi-Fi Fingerprinting“, in 2018 Sixth International Symposium on Computing and Networking Workshops (CANDARW), Takayama, Nov. 2018, S. 196–201. doi: 10.1109/CANDARW.2018.00045.