 |
    |
Fast and Adaptive Deep Fusion Learning for Detecting Visual Objects
Nikolaos Doulamis1 and Anastasios Doulamis2
1National Technical University of Athens, 9 Heroon Polytechnious Str. 15773, Zografou, Athens, Greece
ndoulam@cs.ntua.gr
2Technical University of Crete, University Campus, Chania, Greece
adoulam@cs.ntua.gr
Abstract. Currently, object tracking/detection is based on a “shallow learning” paradigm; they locally process features to build an object model and then they apply adaptive methodologies to estimate model parameters. However, such an approach presents the drawback of losing the “whole picture information” required to maintain a stable tracking for long time and high visual changes. To overcome these obstacles, we need a “deep” information fusion framework. Deep learning is a new emerging research area that simulates the efficiency and robustness by which the humans’ brain represents information; it deeply propagates data into complex hierarchies. However, implementing a deep fusion learning paradigm in a machine presents research challenges mainly due to the highly non-linear structures involved and the “curse of dimensionality”. Another difficulty which is critical in computer vision applications is that learning should be self adapted to guarantee stable object detection over long time spans. In this paper, we propose a novel fast (in real-time) and adaptive information fusion strategy that exploits the deep learning paradigm. The proposed framework integrates optimization strategies able to update in real-time the non-linear model parameters according in a way to trust, as much as possible, the current changes of the environment, while providing a minimal degradation of the previous gained experience.
LNCS 7585, p. 345 ff.
lncs@springer.com
© Springer-Verlag Berlin Heidelberg 2012