1 / 25

Distributed Data Association for Multi-Target Tracking in Sensor Networks

Distributed Data Association for Multi-Target Tracking in Sensor Networks. Alan S. Willsky SensorWeb MURI Review Meeting December 2, 2005. A Notional Example. Multiple sensors with one or more bearing or location measurements. Possibly additional signal features.

iona
Télécharger la présentation

Distributed Data Association for Multi-Target Tracking in Sensor Networks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Distributed Data Association for Multi-Target Tracking in Sensor Networks Alan S. Willsky SensorWeb MURI Review Meeting December 2, 2005

  2. A Notional Example Multiple sensors with one or more bearing or location measurements Possibly additional signal features Challenge: Scalable and resource-aware algorithms for data association and tracking/estimation under network constraints

  3. Reminder of where we were last year • Inference algorithms for graphical models • Exploitation of embedded tractable subgraphs • Truly optimal algorithms through “tree-reweighting” • Formulation of approaches to data association as problems in inference for graphical models • Automatic construction of sensor-target-region graphical models • Communication-sensitive message passing protocol • Experimental results demonstrating the power of this method and a performance-versus-comms-load threshold effect • Demonstration of scalability • New approach to inference for continuous variables: Non-parametric Belief Propagation (NBP)

  4. What we’ve done lately • Important new approaches to high-performance inference for graphical models • Recursive Cavity Modeling for near-optimal inference • Walk-sum analysis of message-passing algorithms • Neither will be discussed today • Substantial extension of our graphical model-based methods for multi-target tracking • A new approach to dynamic hypothesis management for multitarget tracking • New graphical model structure • Use of NBP to manage hypotheses • Natural accommodation of late-arriving data

  5. Distributed data association using advanced methods for graphical models • Why graphical models? • Natural match to sensor networks • A variety of parallel message-passing algorithms that provide basis for comms-constrained solutions and tradeoff analysis • How do data association problems map to graphical model inference problems? • Several approaches, which highlight key issues • Careful “hybrid” organization of hypotheses • Different approaches for “organized” networks and networks requiring some degree of organization

  6. Data association in an organized network

  7. Graphical models for organized networks - I • Sensor-oriented modeling • Nodes correspond only to sensors – GOOD! • Node variables correspond to associations of measurements for each individual sensor • Trouble if > 2 sensors cover same region – BAD! • Leads to complex, messaging among many sensors • Target-oriented modeling • Nodes correspond to targets – BAD! • Has straightforward, graphically simple messaging structure – GOOD!

  8. Graphical models for organized networks - II • Hybrid models • Start from sensor-based model • Introduce target nodes only for targets seen by more than two sensors • Leads to models with pairwise cliques only • Mapping of inference computations and messages to physical nodes and comms • “Maximally explicit” • Makes crystal clear an important issue related to target handoff (which node takes responsibility for which target) • Easy to construct automatically and efficiently

  9. Region-based representation when self-organization needs to be accomplished • Elementary variables are the numbers of objects in each of a set of disjoint subregions covering the surveillance region • Each subregion is that surveilled by a specific subset of sensors • For simplicity, assume sensors are very simple “proximity” indicators • Local signal processing provides likelihood function for the number of targets present within the range of that sensor • Effects of individual false alarms and missed detections are captured in these local likelihood functions

  10. Algorithmic Complexity • In fully asynchronous, comms-unconstrained implementation • Complexity per sensor per iteration is constant • This is the key to scalability • The value of that constant depends on sensor/target densities: • Number of targets seen by each sensor • Number of targets seen by more than two sensors • Number of sensors which all overlap the same subregion

  11. Communications-sensitive message-passing • Objective: • Provide each node with computationally simple (and completely local) mechanism to decide if sending a message is worth it • Need to adapt the algorithm in a simple way so that each node has a mechanism for updating its beliefs when it doesn’t receive a full set of messages • Simple rule: • Don’t send a message if the K-L divergence from the previous message falls below a threshold • If a node doesn’t receive a message, use the last one sent (which requires a bit of memory: to save the last one sent)

  12. Experiments to assess tradeoff of comms vs performance • 25 sensors • ~40-75 targets (2-4 seen by each sensor) • Results: • Sharp transitions in tradeoff as a function of message tolerance threshold • Provides rational basis for setting threshold • When using suboptimal inference algorithms (e.g., loopy BP): • Stopping messaging can improve performance! • Dynamics of messaging provides scenario-dependent adaptivity automatically

  13. Typical example for organized network data association

  14. Illustrating comms-sensitive message-passing dynamics Self-organization with region-based representation Organized network data association

  15. “Partially-Organized” Network • Include sensor, target, and region nodes • Sensors linked to regions (which measurements go with which region?) • Targets are linked to regions (in which region is each target?)

  16. N-Scan Tracking and Data Association • Widely used in multiple hypothesis tracking • Defer making hypothesis decisions about target tracks until additional information is available • Usual “track-based” approach has exponential complexity

  17. An Equivalent Graphical Representation • Add nodes that allow us to separate target dynamics from discrete data associations • Perform explicit data association within each frame (using evidence from other frames) • Stitch across time through temporal dynamics

  18. Hybrid Message Passing - I • Discrete BP in the discrete subgraph for each scan • NBP for the continuous tracking • Distribution for each target at each time is a sum of many terms (corresponding to all of the possible measurement associations) • However, one key is that NBP samples these distributions

  19. Hybrid Message Passing - II • Stitching discrete and continuous nodes • Discrete-to-continuous node messages • Mixtures weighted by current discrete-node probabilities for alternate association hypotheses • Continuous-to-discrete node messages • Updated evidence from current continuous-node track distribution for alternate association hypotheses

  20. Hybrid Message Passing - III • The big win: • We still maintain explicit data association at each frame • However, we avoid the computationally disastrous problem of enumerating all data associations across time • Instead, we let NBP-based sampling provide statistically significant evidence for enhanced data association through the use of multiple scans

  21. An Example: 20 targets, 25 sensors • Each sensor has ~ 5 targets within its coverage region (and sensors are also subject to false alarms and missed detections)

  22. A small example

  23. The Way Ahead - I • Further savings over standard MHT: • Hypothesis pruning • Use K-L divergence to decide when there is no more evidence to enhance particular data associations • Hypothesis merging • Use K-L divergence to decide when two or more alternate data association hypotheses lead to nearly identical target track distributions

  24. The Way Ahead - II • More detailed look at comms-sensitive message-passing • Depends on where non-sensor node computations are done • Messages internal to a single processor should be censored only for the control of hypothesis explosion (which might have a different threshold) • This begs the question of how one assigns inference responsibilities to nodes • AND to the question of how these are handed off as tracks move • There’s a comms cost of handoff!

  25. The Way Ahead - III • Enhancing performance through the use of our emerging, new methods for inference in graphical models • Recursive cavity modeling • Involve propagating information out radially from initiating nodes (and then back toward these nodes) • Walk-sum-based algorithms • Takes greater advantage of local memory at each processing node

More Related