Bayesian-Based Decision-Making for Object Search and Classification

Abstract

This paper focuses on the development of real-time decision-making criteria for an autonomous vehicle whose tasks to be performed are competing under limited sensory resources. More specifically, we are interested in the search and classification of multiple static objects of unknown number and positions given a single autonomous sensor vehicle. In this case, search and classification are two competing demands since an autonomous vehicle can perform either task but not both at the same time. During a search task, once the vehicle finds an object of interest, it has to decide on whether to neglect the found object and continue searching, or stop and observe that object. If the vehicle decides to stop and classify the object, it has to decide on how much time it can afford to do so. This is a very critical decision as choosing one option over the other may mean missing other, more important objects not yet found, or missing the opportunity to satisfactorily classify a found critical object. Building on previous deterministic work by the authors, in this paper we develop Bayesian-based probabilistic and information-theoretic search versus classification decision-making criteria that result in guaranteed detection and classification of all the unknown objects in the domain. Simulation-based results are provided to study the performance of the proposed decision-making strategy.

Cost-Aware Bayesian Sequential Decision-Making for Search and Classification

“…This is very important in applications where D is large scale (i.e., too large to be optimally or suboptimally covered by a single set of static sensors). The application of limited-range sensors is consistent with previous work for dynamic coverage control of multi-sensor network with flocking and guaranteed collision avoidance [27][28][29]31], awareness-based coverage control and decision-making for search versus tracking using multiple autonomous vehicles with intermittent communications [3,33], Bayesian-based binary decision-making for search versus characterization/classfication [4,5], and underwater effective coverage [30,32].…”

“…Other motion schemes, such as gradient-based, awareness-based, and information-driven control laws ( [3][4][5][27][28][29][30][31][32][33]), can be adopted without difficulty.…”

“…This decision is especially critical when the mission domain is too large to be covered by statically distributed sensors with limited sensory ranges, or the number of unknown objects is far more than that of the available sensors [3][4][5]. To accomplish these competing tasks under limited sensory resources with minimum costs, in this paper, we develop a decision-making strategy that dynamically chooses the task to be performed based on an overall risk assessment associated with the decision.…”

“…To remedy this, the authors developed a decision-making strategy for search versus tracking within a deterministic framework in [3] and extended the result to a probabilistic framework in [4] and [5]. In [3] the model describes how "aware" the vehicle fleet is of events over a given domain.…”

“…The proposed strategy guarantees the detection of all objects of interest and the tracking of each object's state for a minimum guaranteed amount of time. In [4] and [5] a probabilistic Bayesian counterpart was developed for unknown object search versus classification treated as two competing tasks. A probabilistic framework is desirable as it takes into account sensor errors, as well as allows for future incorporation of other tasks such as object tracking, data association, data/decision fusion, sensor registration, and clutter resolution [4,5,13].…”