Earth-Mover's Distance as a Tracking Regularizer (bibtex)

by A. Charles, N. Bertrand, J. Lee and C.J. Rozell

Abstract:

Tracking time-varying signals is an important part of many engineering systems. Recently, signal processing techniques have been developed to improve tracking performance when the signal of interest is known a-priori to be sparse. Leveraging sparsity, however, depends heavily on gridding the space, treating the signal as a collection of active or inactive pixels in an image, rather than traditional methods which track the continuous spatial coordinates. Using the dynamics constraint in this setting is challenging, as a model which approximately predicts target location may result in seemingly large errors, as measured by the lp-norm typically used in such algorithms. To take advantage of approximate spatial priors without introducing unnecessary penalties, we present a tracking algorithm using the earth-mover’s distance (EMD) as an alternate dynamics regularization term. We note that while requiring a higher computational burden, the EMD can more effectively utilize target location prediction when the space is gridded.

Reference:

Earth-Mover's Distance as a Tracking RegularizerA. Charles, N. Bertrand, J. Lee and C.J. Rozell. In IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP), December 2017. Submitted.

Bibtex Entry:

@InProceedings{charles.17c, author = {Charles, A. and Bertrand, N. and Lee, J. and Rozell, C.J.}, title = {{Earth-Mover's Distance} as a Tracking Regularizer}, booktitle = {{IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP)}}, year = 2017, month = dec, address = {Cura\c{c}ao, Dutch Antilles}, abstract = {Tracking time-varying signals is an important part of many engineering systems. Recently, signal processing techniques have been developed to improve tracking performance when the signal of interest is known a-priori to be sparse. Leveraging sparsity, however, depends heavily on gridding the space, treating the signal as a collection of active or inactive pixels in an image, rather than traditional methods which track the continuous spatial coordinates. Using the dynamics constraint in this setting is challenging, as a model which approximately predicts target location may result in seemingly large errors, as measured by the lp-norm typically used in such algorithms. To take advantage of approximate spatial priors without introducing unnecessary penalties, we present a tracking algorithm using the earth-mover’s distance (EMD) as an alternate dynamics regularization term. We note that while requiring a higher computational burden, the EMD can more effectively utilize target location prediction when the space is gridded.}, note = {Submitted.} }

Powered by bibtexbrowser