Skip to Main content Skip to Navigation
Conference papers

GNSS Multipath Detection using Embedded Deep CNN on Intel Neural Compute Stick

Abstract : 1. Introduction GNSS signals are often subject to different kinds of events causing significant errors in positioning. This work explores the application of Machine Learning (ML) methods to data from GNSS receivers to detect the signal anomalies. More specifically, we propose to detect the multipath contamination using samples of the correlator output signal as a ML model dataset. We focus on multipath detection because it is still considered to be an important source of error, especially in urban environment. In this work, we investigate the GPS L1 C/A signal. ML algorithms require large amount of data to perform efficient training. Unfortunately labelled GNSS data is often available in limited number. For this reason we have developed a synthetic correlator output generator. It provides several tuning strategies for the multipath, based on doppler shift, propagation delay and phase estimation errors. The phase variable is processed through I and Q channels. At the same time, the doppler shift and the code delay are sampled in 2D grid (referred later as images). Our innovative approach consists in feeding the full correlator output images to the ML model instead of constructing handmade features such as in the case of double delta correlators [1]. This contrasts with traditional algorithms that use a limited number of correlators. Deep convolutional neural networks (CNN) have been used naturally as they have proven to be very efficient in image processing [2]. The CNN has been trained offline to detect multipath situations in the synthetic signal dataset. The trained model is then exported to a Neural Compute Stick (NCS) [3] to assess the performance of the detection in real time tracking conditions. The outline of the given work is as follows: first, we will present the design and validation procedure of the synthetic GNSS data generator and then we will proceed with tests conducted on the prediction model, offline and in real time. 2. Design of synthetic data generator Our generator models the GPS L1 C/A received signal as the sum of desired signal and an additive white Gaussian noise term. The corresponding correlator output is represented in the form of in-phase and quadrature components [4] as a function of the auto-correlation function of the PRN code, the doppler shift, the phase and the propagation delay estimation errors. In this work we take into account only a single multipath signal. Its contribution to the correlator output is considered as an additional term to the main signal. The multipath is parametrized through its difference with the main signal in code delay, doppler shift and phase. In order to validate our synthetic generator we have compared its output to the data from a real receiver. An IFEN SX3 GNSS software receiver [5] has been fed with signals generated with a Spirent GSS6560 GNSS signal generation equipment [6]. Various multipath free and multipath scenarios have produced reference images. We further propose a metric to quantify the consistency of the synthetic output to the ground truth. It is worth noting that these real data were used only to validate our generator. Indeed, for training, it would be unrealistic to produce sufficient amount of labelled physical signals with the pipeline of Spirent GSS6560 - IFEN SX3 receiver. 3. Real time multipath detection using ML 3.1. Methodology To show that our multipath detection method can be embedded into a GNSS receiver with limited computational resources (power, memory, CPU time budget), we propose to implement and assess it on a NCS dongle. In the first step, training is carried out offline with synthetic data. Then the trained model is copied and exported to the NCS dongle. Finally this model is run in real time over real signals or synthetic data for performance assessment or validation. 3.2. Prior work In [7, 8] experiments have already been conducted to detect multipath with a CNN in an offline setting. This work has shown that CNN can achieve very good performance when compared to another ML based multipath detection method. These results are confirmed for various 2D grid sizes (or image resolution), which is directly linked to the required number of correlator outputs. This work has demonstrated that the CNN was able to catch geometrical dependencies in the data. The heatmap extracted from the output of the last convolutional layer highlights the regions of importance for the CNN to discriminate multipath. 3.3. Real time performance assessment The objective now is to assess the performance of the detection method presented in [8] in a real time context. One of the challenges is to integrate the prediction model developed offline into an embedded target. The choice of the NCS dongle was motivated by its architecture dedicated to computer vision tasks. Indeed, it implements Myriad X Vision Processing Unit. The question of the most suitable embedded online CNN architecture is therefore raised. We also aim at demonstrating that the proposed prediction method can be operated in real time. Prediction will be made on unseen real data from the Spirent GSS6450-IFEN SX3 receiver pipeline. The performance of the algorithm will then be evaluated with respect to 2D grid resolution and detection time constraint budget. 4. Innovative contribution of the work - The proposed detection technique captures the entire correlation information instead of processing limited samples (ex: Early/Prompt/Late values); - A dataset generator has been developed and validated on real data. It is able to build on demand as many labelled samples as required for CNN training; - A real time embedded detection solution is evaluated using the Intel Neural Compute Stick (NCS).
Document type :
Conference papers
Complete list of metadatas

https://hal-enac.archives-ouvertes.fr/hal-02964023
Contributor : Laurence Porte <>
Submitted on : Monday, October 12, 2020 - 9:46:42 AM
Last modification on : Tuesday, October 20, 2020 - 10:32:07 AM

Identifiers

  • HAL Id : hal-02964023, version 1

Collections

Citation

Evgenii Munin, Antoine Blais, Nicolas Couellan. GNSS Multipath Detection using Embedded Deep CNN on Intel Neural Compute Stick. ION GNSS+ 2020, 33rd International Technical Meeting of the Satellite Division of the Institute of Navigation, Sep 2020, Virtual event, United States. ⟨hal-02964023⟩

Share

Metrics

Record views

12