Bayesian Transfer Learning in Sensor Networks for Autonomous Urban Driving

advisor: prof. Anthony Quinn, Ph.D.
e-mail: show e-mail
type: bachelor thesis, master thesis
branch of study: MI_MM, MI_AMSM
key words: Bayesian filtering, wireless sensor networks, autonomous urban driving, Bayesian transfer learning, distributed decision
description: Wireless sensor networks are an essential infrastructure for progress with the Industry 4.0 agenda, and for smartening our urban environments. In particular, progress towards autonomous driving in smart urban transport infrastructures requires distributed knowledge representation and decision-making, via transfer learning between fixed and mobile sensors. Traditionally, these sensors process locally sensed, nonstationary time series realisations (signals) into summary statistics, either for use in local decision-making or for sharing between sensors in a distributed network. Increasingly, these sensors are Bayesian: they convert local signals into probability distributions, notably the filtering and state-predictive distributions calculated by the Kalman filter and broader classes of Bayesian filters. Bayesian transfer learning formalizes the task of - and optimizes the solution for - probabilistic knowledge transfer between Bayesian sensors. Our recently published progress in this area has shown benefits for this framework over classical moment-transfer schemes, at least in the case of correlated states. If our algorithms are to be applied successfully in networked position-velocity sensors for autonomous driving, it is important to make progress with knowledge transfer in multi-sensor networks, as well as in cases where physically distinct quantities are sensed at different nodes in the network. This ambition will inform the core aim of the proposed project. The steps involved in successful delivery of the project will be: 1) A review of the currently implemented algorithms for vehicle localization and tracking in the urban environment, critically comparing Bayesian and classical approaches; 2) Extension of the current solutions for Bayesian transfer learning in a pair of Kalman filters to multiple Kalman filter networks, focussing on (i) specified scenarios of local variable sensing, and (ii) global (coordinated) versus local (distributed) transfer; 3) Experimental validation of the multi-sensor Bayesian transfer solutions in comparison with a classical state-of-the-art; 4) An optional task will be to extend the (networked) Bayesian filter/sensor characterisation beyond the linear-Gaussian restrictions of the Kalman filter, notably to filters which track local variables via sequential Monte Carlo algorithms.
references: 1. Papež, M. and Quinn, A., “Robust Bayesian Transfer Learning between Kalman Filters”, Proc. IEEE International Workshop on Machine Learning for Signal Processing, Pittsburg, 2019; 2. Särkkä, S., “Bayesian Filtering and Smoothing”, Cambridge Univ. Press, 2013; 3. Petrovskaya, A. and Thrun, S., “Model based vehicle detection and tracking for autonomous urban driving”, Autonomous Robots, vol. 26, no. 2, pp. 123-139, 2009.
last update: 23.08.2019 11:26:50

administrator for this page: Ľubomíra Dvořáková | last update: 09/12/2011
Trojanova 13, 120 00 Praha 2, tel. 224 358 540, pevná linka 224 923 098, fax 234 358 643
Czech Technical Univeristy in Prague | Faculty of Nuclear Sciences and Physical Engineering | Department of Mathematics