|Abstract:||Nonlinear and non-Gaussian navigation applications often call for using the particle filter, which handles nonlinearities but requires substantial computer resources. This problem is exacerbated by two issues: (1) a navigation application is by definition hosted on a mobile platform, often making computational resources scarce; and (2) the number of particles required to achieve a desired accuracy grows exponentially with the number of dimensions of the state vector. The combination of these two issues makes particle filters impractical for many navigation applications, especially if an IMU is used. In this paper, we present an economical alternative, which mimics the results of a particle filter using Deep Learning (DL) neural networks. A DL network is trained on a massive set of data, which is obtained using a conventional particle filter. This training is done only once, and the dataset can be prepared using powerful offline resources (e.g., cloud computing). During the training stage, we compute a priori and a posteriori sets of particles and then convert the representation of the state probability distribution to a compact form. The trained network learns to process data, whether from an IMU output or from sensor measurements. During the execution stage, the neural network processes sensor and IMU measurements. This step does not require using particles; rather, it only uses a compact DL representation. As a result, it is substantially more economical than a particle filter. Since the computer-intensive training is done offline, and the actual navigation mission would use only the execution step, this approach effectively unloads the computational burden offline, enabling a nonlinear filter to run on a mobile platform with limited computational resources. In the paper, we apply this concept to a 10-state filter, which solves for position, velocity and the attitude quaternion. We present a proof of the concept simulation results for processing IMU and ranging measurements.|
Proceedings of the 2017 International Technical Meeting of The Institute of Navigation
January 30 - 2, 2017
Hyatt Regency Monterey
|Pages:||1189 - 1195|
|Cite this article:||
Draganov, Alexandr, "FUNNEL: Filtering Using Neural Networks for Exactness and Leanness," Proceedings of the 2017 International Technical Meeting of The Institute of Navigation, Monterey, California, January 2017, pp. 1189-1195.
ION Members/Non-Members: 1 Download Credit