Skip to content

EvgeniiMunin/gnss-multipath-detector

Repository files navigation

gnss-multipath-detector

The model generates the synthetic multipath anomaly in GPS L1 C/A signals and provides the CNN model to predict the presence of multipath in real and synthetic signal.

Installation

To get started with pretrained multipath detection model based on MobileNet please pull the Docker image from DockerHub repository

docker pull evgeniimunin/mp-app-mod:latest

Then run the container as follows on the port 5000. This runs the server and makes it listen

docker run -it -p 5000:5000 mp-app-mod

Download two .csv files for I,Q channels and move them to your current directory. To make the prediction of probability of multipath open another terminal window and run the following command

curl -X POST -F matrixi=@imgi.csv -F matrixq=@imgq.csv http://localhost:5000/predict

Deployment

The multipath detection app is also deployed on the GCP Kubernetes cluster and is available publicly by the API. To make the prediction run the following command

curl -X POST -F matrixi=@imgi.csv -F matrixq=@imgq.csv http://35.226.109.248:5000/predict

Hardware/ Software

  • GPU: 1xTesla K80
  • Keras

Synthetic GNSS data generation and preparation

To generate the GNSS correlator output data you first need to give the execution permission to the run_acquisition_sampling.sh script

chmod +x ./run_acquisition_sampling.sh

Before generating the correlator output generate the noise using the following:

cd corr_noise_generator/
octave run_realiz_fake_noise.m <NB_SAMPLES> <WRITE_TO_CSV (0 or 1)>

Then execute the run_data_sampler.py script to generate the correlator outputs in the form of .csv files with the given carrier-to-noise ratio:

cd ..
python run_data_sampler.py --mp=<MULTIPATH (0 or 1)> --nb_samples=<NB_SAMPLES> --cn0=<CARRIER-TO-NOISE RATIO in dBHz> --discr=<DISCR LEVEL>

DL Model

  • MobileNet
  • DenseNet121

The architectures were chosen as they are both suitable for mobile and embedded based CV applications where there are constraints of RAM and memory. The both architectures are produced by Google.

Inference

To run the inference run the script test_sx3_model.py giving the path for the input image converted into .csv format.

Inference on Visual Processign Unit (VPU) Movidius Neural Compute Stick 2 (NCS2)

The script for running the inference on the NCS2 is located in the notebook convert_keras_ncs2_mpcnn.ipynb. Here the trained Keras model is converted into Tensorflow computation graph and then into IR model compatible with NCS2.

About

Model to predict the multipath anomaly in GPS L1 C/A signals

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •