Spatio-temporal pattern detection using dynamic Bayesian networks

Denis, Nicholas J.; Jones, Eric K.

doi:10.1109/cdc.2003.1272263

Cited by 7 publications

(5 citation statements)

References 1 publication

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In contrast, conventional HMM approach could be sensitive to variations in initial conditions especially if the resulting model is initially off-phase with the current behavior. In Dynamic Bayesian Networks [7], the state machines are handcoded and the probabilities are estimated with ExpectationMaximization algorithms that are computationally much slower than either SDF or CSSR.…”

Section: A Pfsm Construction Algorithmsmentioning

confidence: 99%

Mathematical Foundations of Sensor Network Design Based On Linguistic Informatics

Chattopadhyay

Wen

Phoha

et al. 2010

Proceedings of the 2010 American Control Conference

View full text Add to dashboard Cite

We propose a design approach for sensor networks based on formal linguistic representations of information. The approach exploits the concepts of space-time neighborhoods for dynamic sensor grid formation in the vicinity of an event, and symbolization and nonlinear filtering to formulate rigorous mathematical methods that capture the causal dynamics of distributed fusion processes. We formulate the Fundamental Equation of Linguistic Sensing relating physical design parameters to those in the Information Space, and lays the framework for design and operation of sensor networks that dynamically cluster sensing, processing and communications resources in space-time neighborhoods of emergent hotspots for efficient event tracking. I. I & MA sensor network operates on an infrastructure of sensing, computation, and communication, through which it perceives the evolution of physical dynamic processes in its environment. Sensors require physical interaction with the sensed phenomena and are subject to a number of noise factors. To get reliable performance from less reliable individual sensors, collaborative intelligent inference in the vicinity of a stimulus is necessary to circumvent limitations of sensor data, communications, power, and equipment faults. Basic tradeoffs exist between energy, information, and their time critical effect on operations. Tradeoffs of architectural design parameters like number of nodes, node placement, routing, clustering density, and resource constraints for in-situ fusion of spatial-temporal information studied in recent years [11], [12], [6], typically use lassical continuous domain signal processing and pattern classification, and are often based on worst case context-blind design principles.We envision a fundamentally new approach to sensor network operations. Instead of specifying parameters for worst-case design, we postulate designing these systems by dynamically organizing a scalable set of diverse sensing and computational resources-that interact to best support fusion needs in operational environments. The central idea is to affect dynamic space-time sensor clustering in the vicinity of the stimulus, specifically keeping in mind that sensors in the vicinity of a stimulus may need to generate more data than the communication network can effectively handle.Using quantization (referred to as symbolization in the sequel) for low level autonomous aggregation of data, formal linguistic representations of sensed data and patterns are generated, which lead to reliable and efficient handling of information. Thus the proposed approach can effectively relate the overall design problem to pattern complexity, sensor resolution, and other relevant effects.A key contribution of this paper is the formulation of the Fundamental Equation of Linguistic Sensing which relates the parameters of the physical design space to those in the abstract information space, and consequently allows the formulation of a general design methodology based on linguistic sensing.The rest of the paper...

show abstract

Section: A Pfsm Construction Algorithmsmentioning

confidence: 99%

Mathematical Foundations of Sensor Network Design Based On Linguistic Informatics

Chattopadhyay

Wen

Phoha

et al. 2010

Proceedings of the 2010 American Control Conference

View full text Add to dashboard Cite

show abstract

“…Moreover, random variables are continuous (state), discrete (type), and time dependent. The graphical models represent an interesting formalism in object-of-interest detection (OID) and have already been used in similar topics [26][27][28][29]. Graphical models are traditionally used to represent dependency relations between a set of N random variables.…”

Section: Object-of-interest Detectionmentioning

confidence: 99%

Situation Assessment: An End-to-End Process for the Detection of Objects of Interest

Pollard

Rombaut

Pannetier

2013

IEEE Trans. Aerosp. Electron. Syst.

View full text Add to dashboard Cite

“…It is important to mention here that conventional HMM-based approaches could be sensitive to variations in initial conditions especially if the resulting model is initially off-phase with the current behavior. Also, in Dynamic Bayesian Networks [16], the state machines are hand-coded and the probabilities are estimated with Expectation-Maximization algorithms that are computationally much slower than either S DF or CS S R [9]. .…”

Section: A Data-driven Pfsa Construction Algorithmsmentioning

confidence: 99%

Optimal path-planning under finite memory obstacle dynamics based on probabilistic finite state automata models

Chattopadhyay

Ray

2009

2009 American Control Conference

View full text Add to dashboard Cite

The ν ⋆ -planning algorithm is generalized to handle finite memory obstacle dynamics. A sufficiently long observation sequence of obstacle dynamics is algorithmically compressed via Symbolic Dynamic Filtering to obtain a probabilistic finite state model which is subsequently integrated with the navigation automaton to generate an overall model reflecting both navigation constraints and obstacle dynamics. A ν ⋆ -based solution then yields a deterministic plan that maximizes the difference of the probabilities of reaching the goal and of hitting an obstacle. The approach is validated by simulated solution of dynamic mazes. I & MRecently, a novel path planning algorithm ν ⋆ was reported that models the navigation problem in the framework of Probabilistic Finite State Machines and computes robust optimal plans via optimization of the PFSA from a strictly control-theoretic viewpoint. In this paper, we present a significant improvement; ν ⋆ is generalized to handle finite memory obstacle dynamics. A sufficiently long observation sequence of obstacle movement in terms of an evolving obstacle map is symbolically compressed to obtain a finite state probabilistic model of obstacle dynamics which is then used to compute optimal plans in dynamic cluttered environments. It is important to note that the problem considered in this paper is different from the ones that modify plans on-the-fly (e.g. D ⋆ ), to incorporate newly learnt information about obstacle locations; in this paper we propose an approach for optimally incorpoating apriori knowledge of the expected dynamical evolution of obstacles. The key advantages are: 1) Pre-processing is cheap: The cellular decomposition required by ν ⋆ is simple and computationally cheap. The cells are mapped to PFSA states which are defined to have identical connectivity via symbolic inter-state transitions. 2) Fundamentally different from search: ν ⋆ optimizes the resultant PFSA via a iterative sequence of combinatorial operations which elementwise maximizes the language measure vector [1] [2]. 3) Computational efficiency: The time complexity of each iteration step can be shown to be linear in problem size implying significant numerical advantage over search-based methods for highdimensional problems. 4) Global monotonicity: The solution iterations are globally monotonic. The final waypoint sequence is generated essentially by following the measure gradient which is maximized at the goal. The measure gradient is reminiscent of potential field methods [3]. However, ν ⋆ automatically generates the measure gradient; no potential function is necessary. Furthermore, the potential function based planners often get trapped in local minimum which can be shown to be a mathematical impossibility for ν ⋆ . The paper is organized in five sections including the present one. Section 2 briefly explains the language-theoretic models considered in this paper, reviews the language-measure-theoretic optimal control †

show abstract

Spatio-temporal pattern detection using dynamic Bayesian networks

Cited by 7 publications

References 1 publication

Mathematical Foundations of Sensor Network Design Based On Linguistic Informatics

Mathematical Foundations of Sensor Network Design Based On Linguistic Informatics

Situation Assessment: An End-to-End Process for the Detection of Objects of Interest

Optimal path-planning under finite memory obstacle dynamics based on probabilistic finite state automata models

Contact Info

Product

Resources

About