INCORPORATING BAYESIAN TRANSFER LEARNING INTO NONLINEAR FILTERS FOR MULTI-SENSOR OBJECT TRACKING

In this dissertation, we present a novel filtering algorithm that employs Bayesian transfer learning to address the challenges posed by mismatched intensity of the noise in a pair of sensors, each of which tracks an object using a nonlinear dynamic system model. In this setting, the primary sensor experiences a higher noise intensity in tracking the object than the source sensor. To improve the estimation accuracy of the primary sensor, we propose a framework that integrates Bayesian transfer learning into an unscented Kalman filter and a cubature Kalman filter. In this approach, the parameters of the predicted observations in the source sensor are transferred to the primary sensor and used as an additional prior in the filtering process. We develop a framework for incorporating Bayesian transfer learning into an unscented Kalman filter to track a nonlinear dynamic motion model in a multi-source system consisting of two or more sensors. Our objective is to extend the application of transfer learning to leverage knowledge gained from several source sensors. To achieve this aim, we generalize the transfer learning approach to online multi-source Bayesian transfer learning. The generalized framework is approximated via the unscented Kalman filter, where the predicted observation densities of one or more sources are transferred to a primary sensor experiencing higher measurement noise intensity. By leveraging knowledge from multiple sources instead of a single source, the tracking performance of the primary sensor is able to achieve a higher level of estimation accuracy. The performance gain increases proportionally with the number of source sensors. Using Bayesian transfer learning, we develop a particle filter approach for tracking a nonlinear dynamical motion model in a dual-sensor system where intensities of measurement noise for both sensors are asymmetric. The densities for Bayesian transfer learning are approximated with the sum of weighted particles to improve the tracking performance of the primary sensor, which experiences a higher noise intensity compared to the source sensor. Furthermore, increasing the number of particles shows an improvement in the performance of transfer learning applied to the particle filter with a higher rate compared to the isolated particle filter. However, the cost of increasing the number of particles is that the computational time per time step increases as the number of particles increases. Moreover, the performance gain of incorporating Bayesian transfer learning is approximately linearly proportional to the absolute difference value between the noise intensities of thesensors in the dual-sensor system. Using simulations for the dual-source system, the transfer learning approach significantly outperforms the conventional isolated unscented Kalman filter and cubature Kalman filter. The effectiveness of incorporating transfer learning into a particle filter is validated compared to an isolated particle filter and transfer learning applied to the unscented Kalman filter and the cubature Kalman filter. In the multi-source system, numerical results demonstrate the effectiveness of multi-source Bayesian transfer learning in improving tracking accuracy.

Author

• Alotaibi, Omar

Identifier

hdl:1920/15258

OAI identifier oai:identifier

oai:MARS:1920/15258