Active Planning, Sensing, and Recognition Using a Resource-Constrained Discriminant POMDP

In this paper, we address the problem of object class recognition via observations from actively selected views/modalities/features under limited resource budgets. A Partially Observable Markov Decision Process (POMDP) is employed to find optimal sensing and recognition actions with the goal of long-term classification accuracy. Heterogeneous resource constraints -- such as motion, number of measurements and bandwidth -- are explicitly modeled in the state variable, and a prohibitively high penalty is used to prevent the violation of any resource constraint. To improve recognition performance, we further incorporate discriminative classification models with POMDP, and customize the reward function and observation model correspondingly. The proposed model is validated on several data sets for multi-view, multi-modal vehicle classification and multi-view face recognition, and demonstrates improvement in both recognition and resource management over greedy methods and previous POMDP formulations.

1. Introduction

In real-time object recognition applications, it is often preferred to sequentially obtain the most informative sensory data in order to reduce the current recognition uncertainty. Such data acquisition scheme, generally called active sensing [14], is useful especially when a huge amount of data are available from various sensors and modalities while we do not have the luxury to capture and process all of them.