19 T.W. Alexander Drive P.O.Box 14006 Research Triangle Park, NC 27709-4006 Tel: 919.685.9350 Fax: 919.685.9360 info@samsi.info

Tracking Working Group - Meeting Summary Page

October
September
August
July
June
May
April
March
February
January
December
November
October
September

October 2009

Local Attendees: Julien Remote Attendees: Brett McClintock, Len Thomas, Chunlin, Bin, Simon

• Brett McClintock on The challenges of fitting animal movement models to data
.
• Discussion on it:
  • jump diffusion models,
  • see previous work from Simon on computer use for motion impaired users,
  • and multiple target tracking with leader.
  • See also work by Emily Fox on bayesian non-parametric for switching linear dynamic systems and application to bees' waggle.

September 2009

August 2009

Local Attendees: Julien Remote Attendees: Chunlin, Simon

• Recap of JSM 2009: attendance between 15 and 20 persons (faced 40 other parallel sessions). Great talks ! List of speakers. Slides:
  1. Dynamic Spatial Mixture Modeling and Its Application in Bayesian Tracking for Cell Fluorescent Microscopic Imaging, Chunlin Ji, Duke University; Mike West, Duke University
  2. Variational Mean Field Approach to Efficient Multitarget Tracking, Ernest Fokoue, Kettering University
  3. Source Term Estimation with Expensive Likelihoods, Nathan Green, Defense Science and Technology Laboratory
  4. Bayes Filtering for Spatio-Temporal Poisson Cluster Point Processes, Daniel E. Clark, Heriot-Wat University; Anthony Swain, Heriot-Wat University
  5. Multi-Target Tracking using MCMC-Based Particle Algorithm, Francois Septier, University of Cambridge
• Mark and Viktor are now in touch about Block Sampling strategies for Multi target tracking.
• Transition workshop: Takes place on Nov 9-10. Who will be attending/presenting ?

Simon emphasize that the aim is now to get the articles ready.

Next meeting in the week of September 7 - 12 (exact day/time to be announced).

July 2009

Remote Atendees: Simon, Francois, Ernest

• Discussion of Ernest's variational approach to multiple target tracking

Remote Attendees: Julien, Ernest

• Sorry for the late reminder: this combined with July 4th week-end allowed only two of us to participate. Next time will be earlier reminer !
• Ernest on Multiple Target Tracking: Ernest has been coding variational methods for Multitarget Tracking, and will be exposing his code and results next Monday, to get our feedback.
• Metrics for comparison: Julien and François will focus on short term on metrics targetted specially at comparison of algorithms for multitarget tracking, in order to get results quickly -- the generic framework is more far-fetched and is for the longer term.

Next meeting: will take next Monday, Monday July 13th, usual time 10am UTC-0400. Beware of WEBEX password change ! (see invitation email). Francois or Simon will be hosting the meeting.

June 2009

A couple of short meetings, no summary.

May 2009

Local Attendees: Julien
Remote Attendees: Bin, Francois, Namrata, Nathan Simon, Viktor,
Chun Lin warned that he could not make it this week.

• Count of papers/technical reports that we can expect for September:
  1. Cell tracking: Chunlin's work on Celltracking
  2. Plume tracking and STE: Nathan, Francois, and Namrata on Plume tracking. Francois has already prepared two technical reports, on Plume tracking and Contour tracking.
  3. PHD filter: Daniel confirmed after the meeting that he currently has one ongoing article about FISST smoothing (presented at the mid-program workshop), and one article under review.
  4. Comparison metrics: Julien, Simon, Francois, Viktor will have one, maybe two (one more practical benchmark, the other more methodological) on metrics to compare SMC algorithms.
  5. ABC algorithms: Mark and Gareth will have one on ABC, and maybe a second one on ABC for source-term estimation
  which makes a rough estimate between 7 and 10 articles or technical reports, which is really great !
• Revival of contour tracking for plume tracking: Namrata, Nathan, and Francois are getting in touch (after this meeting) to see how contour tracking can be applied to plume tracking, and investigate this direction. See Francois' technical report.
• Metrics for SMC algortihms, status update: Julien briefly summarized informally the current status of the metrics, especially the thoughts exchanged with Simon during his stay at SAMSI for the Adaptive Design workshop. Brief slides available.

Next meeting: will take place in two weeks, Monday May 18th, usual time 10am UTC-0400. Beware of WEBEX password change ! (see invitation email). On the menu, status update from all subgroups. Julien won't be attending (giving a seminar at this time this day), but will find another postdoc to start the meeting.

April 2009

Local Attendees: Chunlin, Julien
Remote Attendees: Daniel, Ernest, Mark, Simon

• JSM articles: We made a quick tour of the presentations. No mandatory article to be published, those who want can. In any case, speakers please send a short bibliography to Julien, who will be chairing the session.
  
• Aims until the end of the program: We aim at pulling together around three articles to be submitted before September. Though any work done after SAMSI's program is of course entirely welcome, end of September is an important deadline for SAMSI's final report of activity.
• Revival of contour tracking: Simon will see with Nathan and Namrata how to revive the contour tracking project.

We are now aiming to cristalize the work that has been done so far into publishable shape.

Next meeting will be exceptionally held on Tuesday May 5th (instead of the usual Monday), same time (10am), same phone number. We will then count exactly how many papers are ongoing and with who.

Local Attendees: Julien, Gentry
Remote Attendees: Chunlin

A brief discussion of current research and the upcoming workshop.

Local Attendees: Julien, Chunlin, Gentry
Remote Attendees: Bin Liu, Ernest, Namrata

A brief discussion of current research and the upcoming workshop.

March 2009

Local Attendees: Julien, Chunlin, Gentry
Remote Attendees: Bin Liu, Ernest, Simon, Nathan, Namrata

A good discussion of several topics today. Simon will miss the next two weeks of meetings due to travel. The workshop April 15-17 is comming up shortly, so be aware of those dates.

• Chunlin Continued discussion of his slides from last week and possible new applications for use in a paper.
• Julien: Further discussion of model selection metrics and some clarification by Simon as to the problem of interest.
• Nathan and Namrata: Discussion of their future work and some discussion between Gentry and Nathan as to the emulation issues and methodology.

Local Attendees: Chunlin, Julien
Remote Attendees: Francois. Simon joined at the very end.

Few attendees today unfortunately, due to mess up on timezones. Chunlin's and Julien's slides are online to compensate.

• Chunlin on Dynamic Spatial Mixture Modelling for Multiple Extended Target Tracking. Slides.
  Note that the results exposed here do not rely on any prefatory segmentation step.
• Julien on Refined metrics on SMC algorithms: first thought. Slides
  Any feedback much welcome, specially from Francois and Simon, as this is very starting work.

Next Monday, be careful, Europe might move to summer time, double check the meeting times and the time differences.

Local Attendees: Ernest, Julien
Remote Attendees: Bin, Chunlin, Francois

• Francois on Plume Tracking
  Francois showed primary results for the Plume Tracking model, multiple Gaussian releases with LIDAR simulated data with thresholded values. Dataset and code will be posted on the website.
• Francois on Evaluation of Particle Methods for Multitarget Tracking
  Simon had an idea of using, as a metric, the complete likelihood of the ground truth wrt the regularized particle filtering estimate.
• Collegial discussion about the metrics
  Discussion with Julien and Ernest on extension to variational methods, and on splitting this metric (which measures the divergence between the ground truth and the approximation of the filtering) in two parts:
  1. divergence bw the ground truth and the filtering distribution, which is related to the model chosen (TBD or else),
  2. divergence bw particle approximation and filtering distribution, which depends on the variant of SMC chosen.
  Note that the "ground truth" has to be integrated, typically by MC, to have a metric which is not conditional on a single dataset.

Next Monday, Chunlin will present his results, as Simon will be there, and Ernest will present some possible links to clustering metrics. Julien is in charge on investigating a possible Kullback Leibler Divergence (KLD) approach to the metric.

Monday, March 2nd, 10:00-10:30 AM EDT.

Remote Attendees: Julien, Simon,Chumlin Ji, Gentry White, Bin Liu, Caglar, Namrata
Bin discussed his areas of interest and what projects he may like to work on. Chunlin mentioned he was progressing on the PHD filter for multi target tracking. Simon mentioned a possible set of field data and possible date change for the closing workshop.

February 2009

January 2009

• Ernest Fokue
• François Septier
• Nathan Green
• Chunlin Ji
• Daniel Clark

Local Attendees: Gentry, Ernest
Remote Attendees: Julien, Mark, Viktor

The meeting was sparsely attended, but some discussion of Viktor's work occured and he presented a pdf summarizing some of his current work.
Some discussion of the JSM meeting was brought up, abstracts are due sometime in February so any organization needs to take place soon if there is to be a tracking session at JSM.

December 2008

Local Attendees: Gentry
Remote Attendees: Julien, Nathan, Viktor, Simon, François

A summary of sub-group progress:

• Source Term Estimation/Plume Tracking
  Nathan reported on the STE/PT sub-group's progress. He has been working on the model for STE and tuning various parameters etc. He is also approximately 2/3 of the way to completion of a paper. The PT model is still in process and there may be new LIDAR data available soon.
  
• The Neuron Activity Tracking had no report to make, they are on break until after Christmas.
• It was decided to meet in three weeks January 16 at the regular time 10:30 Eastern Time.

Local Attendees: Gentry and Chunlin
Remote Attendees: Julien, Nathan, Viktor, Joshua, Sze, Simon, François, Mark

A summary of sub-group progress:

• Multiple Target Tracking
  Franç:ois outlined the subgroup's status.
  There are now two scenarios to be used for evaluating the various algorithims
  Scenario1
  Scenario2
• Source Term Estimation/Plume Tracking
  Nathan has supplied a decription of the likeliohoods involved here
  There was some discussion of a more challenging or interesting plume model to use for evaluation in lieu of the simple Gaussian puff model
  Future work in the near term includes a Bayesian aproach to STE and an attempt to implement the simple models by Christmas. R code should be available soon.
• The Neuron Activity Tracking had no report to make.
• It was decided to meet in two weeks December 16 at the regular time 10:30 Eastern Time.

November 2008

Local Attendees: Julien and Gentry
Remote Attendees: Simon, Joshua, Nathan, Steve, Daniel Clark and Viktor

A summary of sub-group proposals:

Joshua presented the additional slides from his previous presentation concerning future research; see the sldes marked "Next Steps" .
In summary the areas he addressed are

• Assigning priors to parameters
• On-line parameter estimation
• The extension to continuous time
• The implementation of a spatial filter
• Improvements in sampling efficeincies
• Population Sampling
• Image segemntation

Nathan presented several slides and a short proposal covering STE/Plume Tracking.
In terms of the STE problem the areas for exploration include

• SMC samplers
• ABC
• Adaptive proposal dsitributions
• Trans-dimensional SMC
• Trials Data

• A comaprison of methodologies including the Kalman Filter, and Particle Filter as baselines
• Contour tracking
• ABC
• Trials Data

Simon presented proposals for the Multi-Target Tracking sub-group in lieu of François

1. A Survey Paper
   • Comparing Algorithms using François's models
   • SMC methods
   • MCMC particles algorithm
   • PHD Filter
   Specific methods of comparison to be addressed are
   • Evaluation Methods
   • Label Switching
   • Convergence Properties of MCMC particles algorithm
2. Invesitgate new methods to deal with joint inferences of large scale systems
   • Population SMC/MCMC methods
   • Proposals based on genetic operators
   • Hybrid algorithms
   • State partitions
   • Fixed-lag smoothing
   • Adaptive proposals

Chunlin was unavailable to give a report of activities but the PHD filter code is running for his example.

Tuesday, November 4th, 10:30-11:30 AM EDT.

Local Attendees: Julien, Chunlin and Gentry
Remote Attendees: Simon, Joshua, Mark, and Viktor

Some general updates:

Nathan has progressed with his toy filtering R code for STE

Chunlin will see with Daniel Clark whether the R code for PHD filter can be made available.

This week we discussed narrowing the scope of activities to four areas:

• Cell Tracking
• Multiple Target Tracking
• Neuron Activity Tracking
• Source Term Estimation.

We appointed Provisonal heads of these areas as:

• Chunlin: Cell Tracking
• François: Multiple Targets
• Josh: Neurons
• Nathan: STE.

We are also looking into establishing e-mail lists for these subgroups as well

Next week we would also like to hear proposals for a course of activities from each of the subgroups outlining the scope of their activities and a general idea of their possible deliverables over the course of the program.

October 2008

Local Attendees: Nathan, Julien, Chunlin and Gentry
Remote Attendees: Simon, Joshua, Mark,
This week we discussed possible future work including trying to narrow the scope of activities picking specific projects to pursue.

Nathan's talk

Nathan reviewed what areas have been explored so far and covered the four main areas for further research, his slides are here

Chunlin showed some of the results he has obtained using the PHD Filter for cell tracking his presentation is here

Tuesday, October 14th, 9:10-10:30 AM EDT.

Locally present: Gareth, Julien, Nathan. Arnaud Doucet joined for the ABC part.
Remote: François, Mark.
Simon warned last week that he could not be join us today.
Apologies: Due to a technical hick-up, though the meeting was announced on the website, the reminder email was not sent ahead of the meeting. We apologize to those of you who consequently missed it.

• Slides of his presentation,
• Paper detailing the model, given ahead of the meeting,
• MATLAB source code.

Nathans's talk

Nathan presented the Approximate Bayesian Computation (ABC) algorithm, based on [14] and [15]. His slides will soon be available. Gareth and Mark are currently working on a review article of the wide range of variants of ABC intertwinned with SMC.

Next week agenda

So as to allow Gentry (the webmaster) to be present, the meeting time will now be 10:30am (EDT) on Tuesdays.
Any lecturer willing to talk at next workgroup, for example to present a general paper, is more than welcome to e-mail Nathan, Gentry, or Julien.
As Simon mentionned last week, it is also now time to chose our interests, so as to split into subgroups. Please think of which project you want to be involved in.

Locally present: Julien
Remote: Joshua, Kevin, Mark, Simon, Viktor
Nathan warned in advance that he could not be here.

Joshua's talk

Joshua presented his application in neuron activity tracking. Slides and videos will be available soon on this website, along with a list of references and some exemplar datasets.
This problem is part of his PhD thesis, and any help will be appreciated on that. He will be able to come to SAMSI in spring.

Next week agenda

On next meeting:

• François will present his code and the documentation about multi-target tracking. He sent these to Gentry these morning, hence they should be available on the website much before the meeting so that everyone can have a look at them.
• Nathan will present ABC algorithm and tell us more on LIDAR.

Simon outlines that, now that we have summarised all of our problems of interest, it is time to work ahead and see exactly who wish to work on each topic, bringing some algorithms, developing, etc. For now on, the meetings will be kept weekly, so next meeting is same day, same hour, next week (Tuesday October 14th, 9:00-10:00am EDT (UTC-0400).

September 2008

Locally present: Chunlin, Gareth, Julien, Nathan.
Remote: Amadou, Lyudmila, Simon -- Viktor and Ernest joined during the meeting.

Before the meeting

Lyudmila Mihaylova furnished the reference [11] Gning et al. (2008) on group object motion estimation.

Multiple Object Tracking

Lyudmila gives a quick overview and summar of [11] about random graph and multiple object tracking. She can send videos to be put on the website, and also points to the second article [12] Gning et al. (2008b). She would be interested in convergence results, such has having upper bounds when the noise belongs to a particular class of function, similar to what Dan Crisan presented at SMC Opening Workshop.

Gareth suggests possible relation between multiple object tracking and online estimation of correlation matrices in multivariate econometrics (he will send references), specially trying to find sparsity of the covariance matrix structure (which is well known to be a very hard problem even in the static case.)

Simons agrees: the fundamental problem is tracking high-dimensional objects, which also relates to other articles posted on the website.

Cell tracking

Presentation by Chunlin, slides available, along with these four movies: first, second, third, and fourth, which did not make it through Webex. Initial localisation and segmentation of the cells is known, we now want to track them. Dynamic lineage analysis. $(x, y)$ are the locations and $v$ is the velocity or whatever information we are interested in about the cell, such as its size.

Correlation in the motion, due to the lineage -- Simon outlines difference between correlation due to lineage and correlation due to other mechanisms. Currently working under the hypothesis a random motion, though biologists could maybe give informations about a pattern in the motion.

The aim is to rebuild the lineage tree, and then possibly track the florescence intensity (which is linked to the expression of a gene). Links to [13] CellTracer software, developed by Wang, You and West.

Pictures from Lidar and Conclusion

Nathan's presents a few slides, specially one with Lidar image for plume tracking, which displays the typically strong background, which is very difficult to modelize.

Meeting time stays on Tuesday at 9:00 EDT (UTC-0400). For next time, we can look at the references higlighted until now, about the problems, to ask any technical questions or suggest interactions. Joshua will also be presenting some of his work.

From now on, the main webmaster Gentry White has arrived to SAMSI, Julien stays as a backup.

Locally present: Nathan, Gareth, Julien, Wang.
Remote: Ernest, Simon, Mark, Joshua, Kevin, Steve, Viktor, Namrata.

Before the meeting

Before the meeting, Nathan Green pointed the references

• [5] Green et al. (2008) and
• [6] Robins and Thomas (2005)

about source term estimation, as well as

• the link to [7] CellTracker website (and specially its publications section),
• [8] Vaswani (2008) presenting PF-MT in general and
• [9] Vaswani et al. (2008) on deforming contour tracking.

Nathan Green and Chunlin Yi slides, and discussion

Slides available for download.

Quick summary of [5]. Background information. LIDAR likelihood, plume tracking Likelihood function, range, see [5]. Articles by Namrata [9] on contour tracking might be useful.

Number of agents $a$ is quite few, up to 30. Source Term Estimation uses simulated data for now on. Nathan did not yet got hold on the real data, but here are lots of it. He will get simulated data from the most elaborate dispersion model, and we can use that as a start. He also set up a basic dispersion model in R, with multiple releases (screen shots from it). His R code is available for the simple Gaussian spuff model. The function $C(x,y,t)$ is the concentration at a given location and time, $Q$ is initial release mass, and $\sigma_x$, $\sigma_y$ dispersion terms - a term is missing, figuring the time $t$, this will be corrected. Gareth suggests exact simulation of the dispersion model via Restrospectvie sampling. The first approach with the plume tracking will be to try Kalman Filter-type (gaussian assumption) tracking.

Gareth will also send some references from Kerrie Mengersen on aerosol tracking.

Slides from Chunlin Yi, presented by Nathan as Chunlin could not make it today. Challenges in Cell Tracking. Currently working on random finite set, PHD filter, for multi-target tracking. Work from Manchester university. Videos, with dramatic birth and merging of objects. Links to the videos are available in the slides.

Multiple tracking refers to [3] for multiple targets with birth and death, work in progress. Once finished, Matlab codes will be posted by Francois and Simon so that we all can play with it and figure precisely the methods.

Steve points to the reference [10] Johannesson et al. (2004) for STE and plume-tracking.

For action

Actions for next week:

• Nathan will post R code and some dispersion papers, as well as get hold of concentration data from DSTL, and try to track the plume using this data. Update: code available.
• Francois and Simon are still working on Matlab and Latex for multiple target tracking, should be able to post it next week so everybody can look and try techiques.
• For the biological problem, wait for Chunlin next week and see what would be useful for him.
• Everyone goes on the webpage, familiarizes with the references, and try some of the code on it -- when available.

Simon underlines that we are in very worthwile areas, where methodological advances are really possible, and asks if anyone here has other suggestions of problems. Joshua will prepare a few slides and references about one his problems for next time.

Schedule issues: trying to find a slot which suits everybody.

Next meeting time: Tuesday October 30th, 9:00-10:30 AM EDT.

Nathan Green and François Septier: global description of the three applications of interest.

• Slides of their presentation.

• [1] Cappe et al. (2007),
• [2] Doucet et al. (2000),
• [3] Pang et al. (2008), and
• [4] Pitt and Shephard (1999).

Description of a lot of models of interest, including plume tracking, study of fluorescence-based video of cells (based on existing very efficient imaging software). Mention of possible use for random sets. Slides by Simon Godsill.

Working Group Home Page

Entire site © 2002-2008, Statistical and Applied Mathematical Sciences Institute. All Rights Reserved.

Last modified