#### DMCA

## The costs of fusion in smart camera networks

### Citations

804 | Consensus and cooperation in networked multi-agent systems
- Olfati-Saber, Fax, et al.
- 2007
(Show Context)
Citation Context ...centralised or distributed [28]. In centralised fusion, all nodes send their local information to a fusion centre (FC) via single-hop or multi-hop communications [8]. The decentralised scheme [10] considers various FCs that collect and fuse information from nodes in their neighbourhood. The allocation of nodes to FCs can be static [10] or dynamic [18]. To support topology changes and scalability, dynamic decentralisation (or clustering) is preferred. In distributed fusion [27], each node runs an identical peer-to-peer algorithm to exchange information with other nodes. Flooding [7], consensus [21] and token passing [11] are widely used distributed fusion schemes. In this paper, we analyse the communication and computation costs of five fusion schemes, namely centralised fusion [8], flooding [7], token passing [11], average consensus [22] and dynamic clustering [18]. We employ the Extended Information Filter (EIF) [19] for all the schemes to perform target tracking in smart camera networks using decision-based fusion. Based on this analysis we discuss which scheme to use based on the communication topology and the available communication and computation resources. The software of the fu... |

530 | Randomized gossip algorithms
- Boyd, Ghosh, et al.
(Show Context)
Citation Context ...heme to smart camera networks using distributed PFs assuming that viewing nodes can communicate with each other. The scheme is suitable when cameras with overlapping FOVs are connected or routing tables are provided. Reaching consensus means that all nodes have the same value for the considered variable(s) such as the target state [29, 22]. Consensus schemes operate at two time scales: collecting measurements and performing iterations between consecutive measurement collections [23]. In each iteration, nodes exchange information with neighbours and perform fusion using the average [2], gossip [1, 3], maximum or minimum [9] approaches. Average consensus is widely used in wireless sensor networks [22] and smart camera networks [27, 26, 28]. The distributed Kalman Consensus Filter (KCF) [22] computes local estimates (xik|k) via Kalman Filters (KF). Non-linear measurement models or motion models require other filters such as EIF [16] or PF [23]. In average consensus, each node ci exchanges its posterior (y i k|k and Y i k|k) with neighbours where non-viewing nodes send either zeros or predicted posterior [14] as information. Each node ci executes a consensus step as: yi,lk|k = y i,l−1 k|k + ... |

253 | A scheme for robust distributed sensor fusion based on average consensus
- Xiao, Boyd, et al.
- 2005
(Show Context)
Citation Context ...based nodes using centralised fusion, flooding, token passing, average consensus and dynamic clustering. This section describes each scheme for target tracking in smart camera networks. In centralised fusion [8, 25], all viewing nodes send their local posteriors (yik|k and Y i k|k) to a FC for computing the global posterior (yFk|k and Y F k|k). Centralised fusion is suitable for small-scale networks as it has high communication cost near the FC. Other drawbacks of centralised fusion are the vulnerability to FC failures and limited robustness to topology changes. In flooding (or dissemination) [29, 13] all viewing nodes broadcast their local posterior (yik|k and Y i k|k) to all or to subsets of nodes (e.g. viewing nodes) in the network. Information can be distributed in a single iteration if the network is fully connected [17]. Otherwise, flooding requires multihop or multiple iterations of communications. In each iteration, each node sends its own and the previously received information to its neighbours. Eventually all participating nodes have the same set of posteriors [7, 6]. Then, each participating node performs fusion, updates its local posterior. Note that when we aim to flood infor... |

94 |
Decentralized Estimation & Control for Multisensor Systems”,
- Mutambara
- 1998
(Show Context)
Citation Context ...tatic [10] or dynamic [18]. To support topology changes and scalability, dynamic decentralisation (or clustering) is preferred. In distributed fusion [27], each node runs an identical peer-to-peer algorithm to exchange information with other nodes. Flooding [7], consensus [21] and token passing [11] are widely used distributed fusion schemes. In this paper, we analyse the communication and computation costs of five fusion schemes, namely centralised fusion [8], flooding [7], token passing [11], average consensus [22] and dynamic clustering [18]. We employ the Extended Information Filter (EIF) [19] for all the schemes to perform target tracking in smart camera networks using decision-based fusion. Based on this analysis we discuss which scheme to use based on the communication topology and the available communication and computation resources. The software of the fusion schemes is available at http://www.eecs.qmul.ac.uk/~andrea/software.htm. The paper is organised as follows. Section 2 reviews the EIF-based state estimation. Section 3 discusses the fusion schemes for target tracking and Section 4 compares their costs. Finally, Section 5 concludes the paper. 2. EXTENDED INFORMATION FILTE... |

93 | Broadcast gossip algorithms for consensus. Signal Processing,
- Aysal, Yildiz, et al.
- 2009
(Show Context)
Citation Context ...heme to smart camera networks using distributed PFs assuming that viewing nodes can communicate with each other. The scheme is suitable when cameras with overlapping FOVs are connected or routing tables are provided. Reaching consensus means that all nodes have the same value for the considered variable(s) such as the target state [29, 22]. Consensus schemes operate at two time scales: collecting measurements and performing iterations between consecutive measurement collections [23]. In each iteration, nodes exchange information with neighbours and perform fusion using the average [2], gossip [1, 3], maximum or minimum [9] approaches. Average consensus is widely used in wireless sensor networks [22] and smart camera networks [27, 26, 28]. The distributed Kalman Consensus Filter (KCF) [22] computes local estimates (xik|k) via Kalman Filters (KF). Non-linear measurement models or motion models require other filters such as EIF [16] or PF [23]. In average consensus, each node ci exchanges its posterior (y i k|k and Y i k|k) with neighbours where non-viewing nodes send either zeros or predicted posterior [14] as information. Each node ci executes a consensus step as: yi,lk|k = y i,l−1 k|k + ... |

64 |
Distributed particle filter with GMM approximation for multiple targets localization and tracking in wireless sensor network
- Sheng, Hu, et al.
- 2005
(Show Context)
Citation Context ...iors [7, 6]. Then, each participating node performs fusion, updates its local posterior. Note that when we aim to flood information only to viewing nodes, non-viewing nodes might hold less accurate results as they do not receive the posterior of all viewing nodes. In such cases, non-viewing nodes do not perform fusion to save computation. For large and sparse networks, flooding has high communication cost, high processing cost and high memory requirements [23]. This scheme is therefore suitable for sharing low amounts of information when high connectivity exists among the nodes. Token passing [24, 11, 12] is a sequential estimator in which viewing nodes form an aggregation chain (AC). Each node in the AC receives a partial posterior from the previous one, updates this posterior using its local posterior and sends the result to the next node. The process finishes when all AC nodes are visited once. The most informative node (decided based on the local posterior and the global knowledge of the network) is selected as the next node [13]. The last AC node provides the global posterior at the current time step. Then, this node initiates the AC for the next time step (often also becoming the first A... |

33 | Data Fusion Improves the Coverage of Wireless Sensor Networks",
- Xing, Tan, et al.
- 2009
(Show Context)
Citation Context ...h, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from Permissions@acm.org. ICDSC ’14, November 4-7, 2014, Venezia Mestre, Italy Copyright 2014 ACM 978-1-4503-2925-5/14/11 ...$15.00. http://dx.doi.org/10.1145/2659021.2659032. limited sensor networks require selecting the most suitable scheme to trade-off performance and resources used. Thus, quantifying the costs of resources helps in choosing the most appropriate scheme for each scenario. Fusion schemes can share raw data (e.g. measurements) or decisions (e.g. estimations) [30]. In the former case, measurements or features are fused to obtain the global estimate. In the latter case, local estimates at each node are fused to get the global estimate. Fusion can be centralised, decentralised or distributed [28]. In centralised fusion, all nodes send their local information to a fusion centre (FC) via single-hop or multi-hop communications [8]. The decentralised scheme [10] considers various FCs that collect and fuse information from nodes in their neighbourhood. The allocation of nodes to FCs can be static [10] or dynamic [18]. To support topology changes and scalabili... |

29 |
Model-based multi-sensor data fusion.
- Wen, Durrant-Whyte
- 1992
(Show Context)
Citation Context ...e-off performance and resources used. Thus, quantifying the costs of resources helps in choosing the most appropriate scheme for each scenario. Fusion schemes can share raw data (e.g. measurements) or decisions (e.g. estimations) [30]. In the former case, measurements or features are fused to obtain the global estimate. In the latter case, local estimates at each node are fused to get the global estimate. Fusion can be centralised, decentralised or distributed [28]. In centralised fusion, all nodes send their local information to a fusion centre (FC) via single-hop or multi-hop communications [8]. The decentralised scheme [10] considers various FCs that collect and fuse information from nodes in their neighbourhood. The allocation of nodes to FCs can be static [10] or dynamic [18]. To support topology changes and scalability, dynamic decentralisation (or clustering) is preferred. In distributed fusion [27], each node runs an identical peer-to-peer algorithm to exchange information with other nodes. Flooding [7], consensus [21] and token passing [11] are widely used distributed fusion schemes. In this paper, we analyse the communication and computation costs of five fusion schemes, nam... |

28 |
Distributed tracking in sensor networks with limited sensing range.
- Olfati-Saber, Sandell
- 2008
(Show Context)
Citation Context ...rmation from nodes in their neighbourhood. The allocation of nodes to FCs can be static [10] or dynamic [18]. To support topology changes and scalability, dynamic decentralisation (or clustering) is preferred. In distributed fusion [27], each node runs an identical peer-to-peer algorithm to exchange information with other nodes. Flooding [7], consensus [21] and token passing [11] are widely used distributed fusion schemes. In this paper, we analyse the communication and computation costs of five fusion schemes, namely centralised fusion [8], flooding [7], token passing [11], average consensus [22] and dynamic clustering [18]. We employ the Extended Information Filter (EIF) [19] for all the schemes to perform target tracking in smart camera networks using decision-based fusion. Based on this analysis we discuss which scheme to use based on the communication topology and the available communication and computation resources. The software of the fusion schemes is available at http://www.eecs.qmul.ac.uk/~andrea/software.htm. The paper is organised as follows. Section 2 reviews the EIF-based state estimation. Section 3 discusses the fusion schemes for target tracking and Section 4 compares ... |

25 | Tracking and activity recognition through consensus in distributed camera networks.
- Song, Kamal, et al.
- 2010
(Show Context)
Citation Context ...imate. In the latter case, local estimates at each node are fused to get the global estimate. Fusion can be centralised, decentralised or distributed [28]. In centralised fusion, all nodes send their local information to a fusion centre (FC) via single-hop or multi-hop communications [8]. The decentralised scheme [10] considers various FCs that collect and fuse information from nodes in their neighbourhood. The allocation of nodes to FCs can be static [10] or dynamic [18]. To support topology changes and scalability, dynamic decentralisation (or clustering) is preferred. In distributed fusion [27], each node runs an identical peer-to-peer algorithm to exchange information with other nodes. Flooding [7], consensus [21] and token passing [11] are widely used distributed fusion schemes. In this paper, we analyse the communication and computation costs of five fusion schemes, namely centralised fusion [8], flooding [7], token passing [11], average consensus [22] and dynamic clustering [18]. We employ the Extended Information Filter (EIF) [19] for all the schemes to perform target tracking in smart camera networks using decision-based fusion. Based on this analysis we discuss which scheme t... |

23 | Distributed object tracking using a cluster-based Kalman filter in wireless camera networks.
- Medeiros, Park, et al.
- 2008
(Show Context)
Citation Context ...surements) or decisions (e.g. estimations) [30]. In the former case, measurements or features are fused to obtain the global estimate. In the latter case, local estimates at each node are fused to get the global estimate. Fusion can be centralised, decentralised or distributed [28]. In centralised fusion, all nodes send their local information to a fusion centre (FC) via single-hop or multi-hop communications [8]. The decentralised scheme [10] considers various FCs that collect and fuse information from nodes in their neighbourhood. The allocation of nodes to FCs can be static [10] or dynamic [18]. To support topology changes and scalability, dynamic decentralisation (or clustering) is preferred. In distributed fusion [27], each node runs an identical peer-to-peer algorithm to exchange information with other nodes. Flooding [7], consensus [21] and token passing [11] are widely used distributed fusion schemes. In this paper, we analyse the communication and computation costs of five fusion schemes, namely centralised fusion [8], flooding [7], token passing [11], average consensus [22] and dynamic clustering [18]. We employ the Extended Information Filter (EIF) [19] for all the schemes t... |

19 |
Distributed fusion in sensor networks,”
- Cetin, Chen, et al.
- 2006
(Show Context)
Citation Context ...es for the fusion process itself but requires additional communication. Consensus has the highest communication and computation costs but it is the only scheme that can be applied when not all viewing nodes are connected directly and routing tables are not available. Keywords Information fusion, communication cost, computation cost, target tracking, smart camera networks 1. INTRODUCTION Fusion schemes are widely used in sensor networks to improve task performance and robustness to failures [13]. These schemes define when and what information to share under specific communication architectures [4]. ResourceSandeep Katragadda is supported by the Erasmus Mundus Joint Doctorate in Interactive and Cognitive Environments, which is funded by the EACEA Agency of the European Commission under EMJD ICE FPA 2010-0012. Juan C. SanMiguel is supported by the EU Crowded Environments monitoring for Activity Understanding and Recognition (CENTAUR, FP7-PEOPLE-2012-IAPP) project under GA number 324359. Andrea Cavallaro acknowledges the support of the Artemis JU and in part by the UK Technology Strategy Board through the COPCAMS Project under Grant 332913. Permission to make digital or hard copies of all... |

17 |
P.M.: Distributed Particle Filtering in Agent Networks: A Survey, Classification, and Comparison.
- Hlinka, Hlawatsch, et al.
- 2013
(Show Context)
Citation Context ...otiation. Negotiation helps limiting the number of participating cameras and reduces the required resources for the fusion process itself but requires additional communication. Consensus has the highest communication and computation costs but it is the only scheme that can be applied when not all viewing nodes are connected directly and routing tables are not available. Keywords Information fusion, communication cost, computation cost, target tracking, smart camera networks 1. INTRODUCTION Fusion schemes are widely used in sensor networks to improve task performance and robustness to failures [13]. These schemes define when and what information to share under specific communication architectures [4]. ResourceSandeep Katragadda is supported by the Erasmus Mundus Joint Doctorate in Interactive and Cognitive Environments, which is funded by the EACEA Agency of the European Commission under EMJD ICE FPA 2010-0012. Juan C. SanMiguel is supported by the EU Crowded Environments monitoring for Activity Understanding and Recognition (CENTAUR, FP7-PEOPLE-2012-IAPP) project under GA number 324359. Andrea Cavallaro acknowledges the support of the Artemis JU and in part by the UK Technology Strateg... |

16 |
Giannakis, “Set-membership constrained particle filter: Distributed adaptation for sensor networks
- Farahmand, Roumeliotis, et al.
- 2011
(Show Context)
Citation Context ...ks using distributed PFs assuming that viewing nodes can communicate with each other. The scheme is suitable when cameras with overlapping FOVs are connected or routing tables are provided. Reaching consensus means that all nodes have the same value for the considered variable(s) such as the target state [29, 22]. Consensus schemes operate at two time scales: collecting measurements and performing iterations between consecutive measurement collections [23]. In each iteration, nodes exchange information with neighbours and perform fusion using the average [2], gossip [1, 3], maximum or minimum [9] approaches. Average consensus is widely used in wireless sensor networks [22] and smart camera networks [27, 26, 28]. The distributed Kalman Consensus Filter (KCF) [22] computes local estimates (xik|k) via Kalman Filters (KF). Non-linear measurement models or motion models require other filters such as EIF [16] or PF [23]. In average consensus, each node ci exchanges its posterior (y i k|k and Y i k|k) with neighbours where non-viewing nodes send either zeros or predicted posterior [14] as information. Each node ci executes a consensus step as: yi,lk|k = y i,l−1 k|k + wij ∑ j∈CNi (yj,l−1k|k −... |

14 | Collaborative sensing in a distributed PTZ camera network.
- Ding, Song, et al.
- 2012
(Show Context)
Citation Context ...used in sensor networks and the existing approaches in smart camera networks, respectively; whereas Table 3 compares the communication and computation costs for each fusion scheme. Table 2: Decentralised (DEC) and Distributed (DIS) tracking techniques for smart camera networks. Key. SC: Static Clustering. DC: Dynamic Clustering. TP: Token Passing. CO: Consensus. KF: Kalman Filter. EKF: Extended Kalman Filter. PF: Particle Filter. IF: Information Filter. Reference Fusion type Fusion scheme Filter Data Decision DEC DIS SC DC TP CO [10] X X KF [18] X X EKF [31] X X KF [20] X X PF [27, 26] X X KF [5] X X EKF [14, 15] X X IF [16] X X EIF 4. QUANTIFYING THE COSTS OF FUSION We consider a wireless smart camera network of eight cameras (Figure 1(a)) that have overlapping FOVs and are single-hop neighbours. We use the communication graph shown in Figure 1(b) where error-free communications and no false measurements are assumed. We consider the 50 trajectories shown in Figure 1(c). Each target is tracked using the five approaches presented in Section 3. The target follows the motion model given by [18]: xk = xk−1 + vx,k−1δk + axδ 2 k/2 yk−1 + vy,k−1δk + ayδ 2 k/2 vx,k−1 + axδk vy,k−1 + ayδ... |

12 |
Distributed and decentralized multicamera tracking.
- Taj, Cavallaro
- 2011
(Show Context)
Citation Context .../14/11 ...$15.00. http://dx.doi.org/10.1145/2659021.2659032. limited sensor networks require selecting the most suitable scheme to trade-off performance and resources used. Thus, quantifying the costs of resources helps in choosing the most appropriate scheme for each scenario. Fusion schemes can share raw data (e.g. measurements) or decisions (e.g. estimations) [30]. In the former case, measurements or features are fused to obtain the global estimate. In the latter case, local estimates at each node are fused to get the global estimate. Fusion can be centralised, decentralised or distributed [28]. In centralised fusion, all nodes send their local information to a fusion centre (FC) via single-hop or multi-hop communications [8]. The decentralised scheme [10] considers various FCs that collect and fuse information from nodes in their neighbourhood. The allocation of nodes to FCs can be static [10] or dynamic [18]. To support topology changes and scalability, dynamic decentralisation (or clustering) is preferred. In distributed fusion [27], each node runs an identical peer-to-peer algorithm to exchange information with other nodes. Flooding [7], consensus [21] and token passing [11] are... |

11 | Architecture for cluster-based automated surveillance network for detecting and tracking multiple persons.
- Goshorn, Goshorn, et al.
- 2007
(Show Context)
Citation Context ...s used. Thus, quantifying the costs of resources helps in choosing the most appropriate scheme for each scenario. Fusion schemes can share raw data (e.g. measurements) or decisions (e.g. estimations) [30]. In the former case, measurements or features are fused to obtain the global estimate. In the latter case, local estimates at each node are fused to get the global estimate. Fusion can be centralised, decentralised or distributed [28]. In centralised fusion, all nodes send their local information to a fusion centre (FC) via single-hop or multi-hop communications [8]. The decentralised scheme [10] considers various FCs that collect and fuse information from nodes in their neighbourhood. The allocation of nodes to FCs can be static [10] or dynamic [18]. To support topology changes and scalability, dynamic decentralisation (or clustering) is preferred. In distributed fusion [27], each node runs an identical peer-to-peer algorithm to exchange information with other nodes. Flooding [7], consensus [21] and token passing [11] are widely used distributed fusion schemes. In this paper, we analyse the communication and computation costs of five fusion schemes, namely centralised fusion [8], flo... |

10 | Distributed nonlinear estimation for robot localization using weighted consensus.
- Simonetto, Keviczky, et al.
- 2010
(Show Context)
Citation Context ...|k = ∑ i∈Cv k Yik|k, (4) where [ygk|k Y g k|k] is the global posterior and C v k is the set of all viewing nodes (cameras observing the same target) at time k and Nvk = |Cvk |. The global state estimate and corresponding error covariance can be calculated using: xgk|k = Y g k|k −1ygk|k, P g k|k = Y g k|k −1. (5) 3. FUSION SCHEMES We perform decision fusion to combine the posteriors of EIF-based nodes using centralised fusion, flooding, token passing, average consensus and dynamic clustering. This section describes each scheme for target tracking in smart camera networks. In centralised fusion [8, 25], all viewing nodes send their local posteriors (yik|k and Y i k|k) to a FC for computing the global posterior (yFk|k and Y F k|k). Centralised fusion is suitable for small-scale networks as it has high communication cost near the FC. Other drawbacks of centralised fusion are the vulnerability to FC failures and limited robustness to topology changes. In flooding (or dissemination) [29, 13] all viewing nodes broadcast their local posterior (yik|k and Y i k|k) to all or to subsets of nodes (e.g. viewing nodes) in the network. Information can be distributed in a single iteration if the network i... |

6 |
Time-space-sequential distributed particle filtering with low-rate communications.
- Hlinka, Djuric, et al.
- 2009
(Show Context)
Citation Context ...ted [28]. In centralised fusion, all nodes send their local information to a fusion centre (FC) via single-hop or multi-hop communications [8]. The decentralised scheme [10] considers various FCs that collect and fuse information from nodes in their neighbourhood. The allocation of nodes to FCs can be static [10] or dynamic [18]. To support topology changes and scalability, dynamic decentralisation (or clustering) is preferred. In distributed fusion [27], each node runs an identical peer-to-peer algorithm to exchange information with other nodes. Flooding [7], consensus [21] and token passing [11] are widely used distributed fusion schemes. In this paper, we analyse the communication and computation costs of five fusion schemes, namely centralised fusion [8], flooding [7], token passing [11], average consensus [22] and dynamic clustering [18]. We employ the Extended Information Filter (EIF) [19] for all the schemes to perform target tracking in smart camera networks using decision-based fusion. Based on this analysis we discuss which scheme to use based on the communication topology and the available communication and computation resources. The software of the fusion schemes is availab... |

6 | Information consensus for distributed multi-target tracking.
- Kamal, Farrell, et al.
- 2013
(Show Context)
Citation Context ...or networks and the existing approaches in smart camera networks, respectively; whereas Table 3 compares the communication and computation costs for each fusion scheme. Table 2: Decentralised (DEC) and Distributed (DIS) tracking techniques for smart camera networks. Key. SC: Static Clustering. DC: Dynamic Clustering. TP: Token Passing. CO: Consensus. KF: Kalman Filter. EKF: Extended Kalman Filter. PF: Particle Filter. IF: Information Filter. Reference Fusion type Fusion scheme Filter Data Decision DEC DIS SC DC TP CO [10] X X KF [18] X X EKF [31] X X KF [20] X X PF [27, 26] X X KF [5] X X EKF [14, 15] X X IF [16] X X EIF 4. QUANTIFYING THE COSTS OF FUSION We consider a wireless smart camera network of eight cameras (Figure 1(a)) that have overlapping FOVs and are single-hop neighbours. We use the communication graph shown in Figure 1(b) where error-free communications and no false measurements are assumed. We consider the 50 trajectories shown in Figure 1(c). Each target is tracked using the five approaches presented in Section 3. The target follows the motion model given by [18]: xk = xk−1 + vx,k−1δk + axδ 2 k/2 yk−1 + vy,k−1δk + ayδ 2 k/2 vx,k−1 + axδk vy,k−1 + ayδk δk + , ... |

5 | Information weighted consensus filters and their application in distributed camera networks.
- Kamal, Farrell, et al.
- 2013
(Show Context)
Citation Context ...ange information with neighbours and perform fusion using the average [2], gossip [1, 3], maximum or minimum [9] approaches. Average consensus is widely used in wireless sensor networks [22] and smart camera networks [27, 26, 28]. The distributed Kalman Consensus Filter (KCF) [22] computes local estimates (xik|k) via Kalman Filters (KF). Non-linear measurement models or motion models require other filters such as EIF [16] or PF [23]. In average consensus, each node ci exchanges its posterior (y i k|k and Y i k|k) with neighbours where non-viewing nodes send either zeros or predicted posterior [14] as information. Each node ci executes a consensus step as: yi,lk|k = y i,l−1 k|k + wij ∑ j∈CNi (yj,l−1k|k − y i,l−1 k|k ), (6) where yi,lk|k is the consensus achieved after the l th iteration and CNi is the neighbourhood of ci. The same process is applied to Yik|k. The values wij can be set to guarantee the convergence to the average of the initial estimates of all nodes after L iterations [21]. The speed of convergence to the posterior average depends on the number of nodes. By multiplying the average with the total number of nodes in the network, Nc, the sum (global posterior) can be calcul... |

4 |
Distributed camera networks.
- Song, Ding, et al.
- 2011
(Show Context)
Citation Context ... when cameras with overlapping FOVs are connected or routing tables are provided. Reaching consensus means that all nodes have the same value for the considered variable(s) such as the target state [29, 22]. Consensus schemes operate at two time scales: collecting measurements and performing iterations between consecutive measurement collections [23]. In each iteration, nodes exchange information with neighbours and perform fusion using the average [2], gossip [1, 3], maximum or minimum [9] approaches. Average consensus is widely used in wireless sensor networks [22] and smart camera networks [27, 26, 28]. The distributed Kalman Consensus Filter (KCF) [22] computes local estimates (xik|k) via Kalman Filters (KF). Non-linear measurement models or motion models require other filters such as EIF [16] or PF [23]. In average consensus, each node ci exchanges its posterior (y i k|k and Y i k|k) with neighbours where non-viewing nodes send either zeros or predicted posterior [14] as information. Each node ci executes a consensus step as: yi,lk|k = y i,l−1 k|k + wij ∑ j∈CNi (yj,l−1k|k − y i,l−1 k|k ), (6) where yi,lk|k is the consensus achieved after the l th iteration and CNi is the neighbourhood of ... |

3 |
Non-centralized target tracking with mobile agents.
- Djuric, Beaudeau, et al.
- 2011
(Show Context)
Citation Context ...are the vulnerability to FC failures and limited robustness to topology changes. In flooding (or dissemination) [29, 13] all viewing nodes broadcast their local posterior (yik|k and Y i k|k) to all or to subsets of nodes (e.g. viewing nodes) in the network. Information can be distributed in a single iteration if the network is fully connected [17]. Otherwise, flooding requires multihop or multiple iterations of communications. In each iteration, each node sends its own and the previously received information to its neighbours. Eventually all participating nodes have the same set of posteriors [7, 6]. Then, each participating node performs fusion, updates its local posterior. Note that when we aim to flood information only to viewing nodes, non-viewing nodes might hold less accurate results as they do not receive the posterior of all viewing nodes. In such cases, non-viewing nodes do not perform fusion to save computation. For large and sparse networks, flooding has high communication cost, high processing cost and high memory requirements [23]. This scheme is therefore suitable for sharing low amounts of information when high connectivity exists among the nodes. Token passing [24, 11, 12... |

3 | Distributed target tracking under realistic network conditions.
- Nastasi, Cavallaro
- 2011
(Show Context)
Citation Context ...terior using its local posterior and sends the result to the next node. The process finishes when all AC nodes are visited once. The most informative node (decided based on the local posterior and the global knowledge of the network) is selected as the next node [13]. The last AC node provides the global posterior at the current time step. Then, this node initiates the AC for the next time step (often also becoming the first AC node). The sequential estimation and the transmission of high dimensional estimations such as Particle Filter (PF) posteriors cause latency [13]. Nastasi and Cavallaro [20] applied such a fusion scheme to smart camera networks using distributed PFs assuming that viewing nodes can communicate with each other. The scheme is suitable when cameras with overlapping FOVs are connected or routing tables are provided. Reaching consensus means that all nodes have the same value for the considered variable(s) such as the target state [29, 22]. Consensus schemes operate at two time scales: collecting measurements and performing iterations between consecutive measurement collections [23]. In each iteration, nodes exchange information with neighbours and perform fusion using... |

3 | Belief consensus algorithms for fast distributed target tracking in wireless sensor networks.
- Savic, Wymeersch, et al.
- 2014
(Show Context)
Citation Context ...node sends its own and the previously received information to its neighbours. Eventually all participating nodes have the same set of posteriors [7, 6]. Then, each participating node performs fusion, updates its local posterior. Note that when we aim to flood information only to viewing nodes, non-viewing nodes might hold less accurate results as they do not receive the posterior of all viewing nodes. In such cases, non-viewing nodes do not perform fusion to save computation. For large and sparse networks, flooding has high communication cost, high processing cost and high memory requirements [23]. This scheme is therefore suitable for sharing low amounts of information when high connectivity exists among the nodes. Token passing [24, 11, 12] is a sequential estimator in which viewing nodes form an aggregation chain (AC). Each node in the AC receives a partial posterior from the previous one, updates this posterior using its local posterior and sends the result to the next node. The process finishes when all AC nodes are visited once. The most informative node (decided based on the local posterior and the global knowledge of the network) is selected as the next node [13]. The last AC n... |

3 | Cluster-based distributed face tracking in camera networks.
- Yoder, Medeiros, et al.
- 2010
(Show Context)
Citation Context ...le 1 and 2 summarise the fusion schemes used in sensor networks and the existing approaches in smart camera networks, respectively; whereas Table 3 compares the communication and computation costs for each fusion scheme. Table 2: Decentralised (DEC) and Distributed (DIS) tracking techniques for smart camera networks. Key. SC: Static Clustering. DC: Dynamic Clustering. TP: Token Passing. CO: Consensus. KF: Kalman Filter. EKF: Extended Kalman Filter. PF: Particle Filter. IF: Information Filter. Reference Fusion type Fusion scheme Filter Data Decision DEC DIS SC DC TP CO [10] X X KF [18] X X EKF [31] X X KF [20] X X PF [27, 26] X X KF [5] X X EKF [14, 15] X X IF [16] X X EIF 4. QUANTIFYING THE COSTS OF FUSION We consider a wireless smart camera network of eight cameras (Figure 1(a)) that have overlapping FOVs and are single-hop neighbours. We use the communication graph shown in Figure 1(b) where error-free communications and no false measurements are assumed. We consider the 50 trajectories shown in Figure 1(c). Each target is tracked using the five approaches presented in Section 3. The target follows the motion model given by [18]: xk = xk−1 + vx,k−1δk + axδ 2 k/2 yk−1 + vy,k−1δk... |

2 |
Kullback-leibler average, consensus on probability densities, and distributed state estimation with guaranteed stability.
- Battistelli, Chisci
- 2014
(Show Context)
Citation Context ... a fusion scheme to smart camera networks using distributed PFs assuming that viewing nodes can communicate with each other. The scheme is suitable when cameras with overlapping FOVs are connected or routing tables are provided. Reaching consensus means that all nodes have the same value for the considered variable(s) such as the target state [29, 22]. Consensus schemes operate at two time scales: collecting measurements and performing iterations between consecutive measurement collections [23]. In each iteration, nodes exchange information with neighbours and perform fusion using the average [2], gossip [1, 3], maximum or minimum [9] approaches. Average consensus is widely used in wireless sensor networks [22] and smart camera networks [27, 26, 28]. The distributed Kalman Consensus Filter (KCF) [22] computes local estimates (xik|k) via Kalman Filters (KF). Non-linear measurement models or motion models require other filters such as EIF [16] or PF [23]. In average consensus, each node ci exchanges its posterior (y i k|k and Y i k|k) with neighbours where non-viewing nodes send either zeros or predicted posterior [14] as information. Each node ci executes a consensus step as: yi,lk|k =... |

2 | Time-space-sequential algorithms for distributed Bayesian state estimation in serial sensor networks.
- Hlinka, Hlawatsch
- 2009
(Show Context)
Citation Context ...iors [7, 6]. Then, each participating node performs fusion, updates its local posterior. Note that when we aim to flood information only to viewing nodes, non-viewing nodes might hold less accurate results as they do not receive the posterior of all viewing nodes. In such cases, non-viewing nodes do not perform fusion to save computation. For large and sparse networks, flooding has high communication cost, high processing cost and high memory requirements [23]. This scheme is therefore suitable for sharing low amounts of information when high connectivity exists among the nodes. Token passing [24, 11, 12] is a sequential estimator in which viewing nodes form an aggregation chain (AC). Each node in the AC receives a partial posterior from the previous one, updates this posterior using its local posterior and sends the result to the next node. The process finishes when all AC nodes are visited once. The most informative node (decided based on the local posterior and the global knowledge of the network) is selected as the next node [13]. The last AC node provides the global posterior at the current time step. Then, this node initiates the AC for the next time step (often also becoming the first A... |

2 | Consensus protocols for distributed tracking in wireless camera networks.
- Katragadda, SanMiguel, et al.
- 2014
(Show Context)
Citation Context ...2]. Consensus schemes operate at two time scales: collecting measurements and performing iterations between consecutive measurement collections [23]. In each iteration, nodes exchange information with neighbours and perform fusion using the average [2], gossip [1, 3], maximum or minimum [9] approaches. Average consensus is widely used in wireless sensor networks [22] and smart camera networks [27, 26, 28]. The distributed Kalman Consensus Filter (KCF) [22] computes local estimates (xik|k) via Kalman Filters (KF). Non-linear measurement models or motion models require other filters such as EIF [16] or PF [23]. In average consensus, each node ci exchanges its posterior (y i k|k and Y i k|k) with neighbours where non-viewing nodes send either zeros or predicted posterior [14] as information. Each node ci executes a consensus step as: yi,lk|k = y i,l−1 k|k + wij ∑ j∈CNi (yj,l−1k|k − y i,l−1 k|k ), (6) where yi,lk|k is the consensus achieved after the l th iteration and CNi is the neighbourhood of ci. The same process is applied to Yik|k. The values wij can be set to guarantee the convergence to the average of the initial estimates of all nodes after L iterations [21]. The speed of converge... |

1 |
Non-centralized target tracking in networks of directional sensors.
- Djuric, Geng
- 2011
(Show Context)
Citation Context ...centralised, decentralised or distributed [28]. In centralised fusion, all nodes send their local information to a fusion centre (FC) via single-hop or multi-hop communications [8]. The decentralised scheme [10] considers various FCs that collect and fuse information from nodes in their neighbourhood. The allocation of nodes to FCs can be static [10] or dynamic [18]. To support topology changes and scalability, dynamic decentralisation (or clustering) is preferred. In distributed fusion [27], each node runs an identical peer-to-peer algorithm to exchange information with other nodes. Flooding [7], consensus [21] and token passing [11] are widely used distributed fusion schemes. In this paper, we analyse the communication and computation costs of five fusion schemes, namely centralised fusion [8], flooding [7], token passing [11], average consensus [22] and dynamic clustering [18]. We employ the Extended Information Filter (EIF) [19] for all the schemes to perform target tracking in smart camera networks using decision-based fusion. Based on this analysis we discuss which scheme to use based on the communication topology and the available communication and computation resources. The so... |

1 |
Distributed consensus filtering for discrete-time nonlinear systems with non-gaussian noise.
- Li, Jia
- 2012
(Show Context)
Citation Context ...s send their local posteriors (yik|k and Y i k|k) to a FC for computing the global posterior (yFk|k and Y F k|k). Centralised fusion is suitable for small-scale networks as it has high communication cost near the FC. Other drawbacks of centralised fusion are the vulnerability to FC failures and limited robustness to topology changes. In flooding (or dissemination) [29, 13] all viewing nodes broadcast their local posterior (yik|k and Y i k|k) to all or to subsets of nodes (e.g. viewing nodes) in the network. Information can be distributed in a single iteration if the network is fully connected [17]. Otherwise, flooding requires multihop or multiple iterations of communications. In each iteration, each node sends its own and the previously received information to its neighbours. Eventually all participating nodes have the same set of posteriors [7, 6]. Then, each participating node performs fusion, updates its local posterior. Note that when we aim to flood information only to viewing nodes, non-viewing nodes might hold less accurate results as they do not receive the posterior of all viewing nodes. In such cases, non-viewing nodes do not perform fusion to save computation. For large and... |