DUNE RCO – Deep learning for multi-technology fusion in industrial indoor asset localization and tracking

Idea and relation to VEDLIoT

Foreseeing an increase in the complexity of indoor positioning, DUNE addresses deep learning aided positioning  fusion  mechanisms  at  different  stages  of  the  localization  process  with  the  goal  to  deliver cost-effective and optimal indoor localization performance. Our models learn from diversified localization inputs and technologies to derive a real-time estimate of the asset’s location. DUNE relies on far-edge, edge and cloud distributed computing capabilities to address real time application needs, functionally splitting the fusion methods to also meet the time, performance, and scalability demands of common industrial applications.

DUNE will contribute to the VEDLIoT ecosystem extension with mainly 3 aspects:

  1. Sensor integration at  the  far-edge  system,  incorporating  diversified  and  industrial  positioning technologies as inputs to the technological framework.
  2. Deep Learning and Exact, Algorithmic and Geometric Models will be incorporated at the Model Zoos for third parties to build positioning systems relying on the framework.
  3. Demonstration that DL enabled edge architectures may become enablers for a next generation of indoor positioning technologies, thus extending the number of application use cases of the VEDLIoT project.

Objectives

  1. Consolidate an end-to-end multi-technology indoor positioning framework.
  2. Explore and evaluate the use of DL as a tool to improve existing art in positioning methods.
  3. Benchmark the performance of DL-based Direction-Finding methods against state-of-the-art methods.
  4. Contribute to the advance on DL-based positioning methods via the publication of open datasets and indoor positioning DL models in a public model Zoo.
  5. The publication of the obtained results in the target dissemination venues, aligning to the VEDLIoT dissemination strategy.
  6. Studying possible IPR protection approaches for the developed methods.
  7. Participation in the dissemination and exploitation activities of the VEDLIoT project, including our participation in the 4YFN at the MWC in Barcelona (2023).

Approach

The radio signals from the tags or cell phones attached to the object/person to be tracked are received by the antenna array of the locator. These raw samples, referred to as IQ samples, need to be transformed into the estimation of the angles that define the signal direction. In a perfect world, this transformation is a geometric process that depends on the inter-antenna spacing (distance)  and the radio frequency (wavelength)  and thus known phase  variation for  a given distance  traveled  by  the  wave.  However,  RF  environments  are  subject  to  noise  and  non-linear irregularities  due  to  interference  and  effects  such  as  fading  or  multipath.  When  dealing  with  these effects,  deep  learning  methods  may  become  a  valuable  tool  to  derive  accurate  position  estimations: They provide the ability to deal with non-linearity, versatile estimation approaches, and great classification capabilities (e.g., to distinguish between direct or reflected signals, thus acting as adaptive multipath filters). In our approach, the u.RECS node will be connected to the Direction finding system and use different NN models from those available in the model Zoo. We will benchmark these models against well-known methods such as classic beamforming or MUSIC algorithms.

Considering a deployment in which several far-edge devices control different antenna arrays or other positioning  technologies  (e.g.,  RFID  tags),  the  edge  architecture  (see  Figure  3)  will  introduce  a subsequent  filtering  step  to  exploit  multi-technology,  time,  and  spatial  domain  characteristics  of  the signal. An edge device, for example the t.RECS, will run different data fusion methods. Some of them will be based on projecting the inherent error of the technology as ellipses surrounding the estimated position and applying a computer vision algorithm to achieve a majority voting scheme, we refer to this method as confidence filtering. This approach can be further refined by using deep learning algorithms, trained to process the virtual images and obtain the points in the plane that better estimate the position of  the  objects  under  tracking  as  observed  by  the  different  locators.

Expected Impact

DUNE will contribute to the consolidation of indoor positioning technologies exploiting the capabilities of next generation CPS systems and distributed/federated architectures and building on the strong opportunity that deep learning can bring when enabled at the far-edge and edge of the infrastructure.

Further info/links