Advanced multiple target tracking techniques
Nous vous rappelons que, afin de garantir l'accès de tous
les inscrits aux salles de réunion, l'inscription aux réunions est
gratuite mais obligatoire.
Inscriptions closes à cette réunion.
29 personnes membres du GdR ISIS, et 30 personnes non membres du GdR, sont inscrits à cette réunion.
Capacité de la salle : 80 personnes.
Réunion d'animation en visio-conférence
La réunion aura lieu en visioconférence. Cependant pour des raisons techniques liées au nombre de connexions simultanées, l'inscription aux réunions est gratuite, mais obligatoire.
Les identifiants de connexion sont communiquées par mail aux inscrits la veille ou le matin de la réunion.
Target tracking consists in recursively estimating the state of an object of interest from uncertain measurements. The applications are widespread, ranging from military surveillance to space situational awareness and robotics. The seminal algorithm is the Kalman filter but it fails to address complex situations such as the presence of several objects in the vicinity of the sensors. Additional challenges may arise such as undetected objects or false alarms. To deal with origin uncertainty in the measurements, a variety of solutions were proposed in the 80's which mainly proceed by running several Kalman filters in parallel and coupling them with a track-to-measurement association step. Well-known techniques are the joint probability data association filter and the multi-hypothesis tracking.
In the early 90's, a new class of algorithms was developed based on the formalism of random finite sets and point process theory. They consist in handling all the targets and all the measurements as whole by representing them as single unordered sets of random vectors. They can then be described by convenient multi-object probability density functions that encompass all the sources of uncertainty including the number of targets and measurements as well as their values.The latter makes it possible to directly generalize the Bayes filter to the multi-object case. Several tractable approximations have been proposed such as the probability density hypothesis filter or variants of the multi-target multi Bernoulli filter. Despite the potential of such algorithms, many issues have still to be addressed such as including interactions between the targets, decreasing the computational complexity or dealing with scenarios including several groups of targets.
The purpose of this special day is to review the recent advances and new challenges on multiple-target tracking but also to emphasize interesting applications. In addition to the invited speakers, researchers are welcome to contribute. The presentation proposals should be sent to the organizers before the 31th of October.
Daniel Clark, Telecom SudParis - CNRS UMR Samovar : firstname.lastname@example.org
François Desbouvries, Telecom SudParis - CNRS UMR Samovar : email@example.com
Audrey Giremus, Université de Bordeaux : firstname.lastname@example.org
Yohan Petetin, Telecom SudParis - CNRS UMR Samovar : email@example.com
8h30-9h : welcoming coffee
9h-9h45 : Multi-object filtering -- Daniel Clark--Telecom SudParis
9h45-10h30 : Opportunistic self-organization and fusion in networks of sensors -- Murat Uney -- Centre for Maritime Research and Experimentation (La Spezia Italy)
10h30-11h15 : Advanced Monte Carlo methods for multi-traget tracking -- François Septier -- University of Bretagne Sud
11h15-12h30 : Group target tracking -- Simon Godsill -- University of Cambridge
12h30-14h : lunch break
14h-14h30 : Target Classification using multi-object filtering -- Isabel Schlangen -- Fraunhofer FKIE
14h30-15h : Problématique et retour d'expérience de la fusion de données pour les tenues de situation dans le domaine des systèmes de combat naval -- Frédéric Livernet -- DGA
15h - 15h30 : EO/IR sensor for UAV detect and avoid -- Dominique Maltese -- Safran Electronics and Defense
15h30 - 16h : Flavio de Melo
Résumés des contributions
Target classification using multi-object filtering (Isabel Schlangen)
In surveillance and reconnaissance applications, the monitored scene is often cluttered with objects that are of secondary interest to the operator, e.g. static structures like buildings or landscape features or moving clutter objects like animals, wind turbines or sea waves. In most state-of-the-art approaches, the clutter signal is treated in a static manner by either manipulating the sensor directly to record less clutter signal at instantaneous time instances (i.e. by using moving target indication to suppress static objects) or by looking for an optimal threshold that only confirms signals of interest. However, the response of moving clutter objects might be very similar to that of the critical targets, therefore simple time-independent thresholding is not suitable for a reliable object classification. In this talk, a dynamic approach is presented that incorporates temporal information through filtering. The chosen approach is a modified Probability Hypothesis Density (PHD) filter that assumes two object populations with distinct modelling parameters whose number is Panjer distributed. It is shown on the example of coastal surveillance that the filter classifies and tracks both critical and clutter targets reliably and hence brings a considerable improvement with respect to situation assessment and efficiency of resource allocation.
Opportunistic self-organisation and fusion in networks of sensors (Murat Uney)
In many situation awareness and machine perception applications, the required level of accuracy can be achieved only with the use of multiple sensors and their networks. Multiple sensor exploitation can be split into network self-organisation and multi-sensor fusion (in the underlying probabilistic model) which are compounded of many problems with undesirable computational complexity and/or intractable aspects. This talk aims to give an overview of recent solutions to opportunistic self-calibration that address these issues in a "latent parameter estimation in state space models" framework. First, we will introduce separable pseudo-likelihoods as underpinning modelling entities for computationally feasible variational inference. Second, we will consider another difficult calibration problem: Synchronization in multi-static active sensing. We will introduce a state space model for fusion at an array receiver and an Empirical Bayes solution for opportunistic synchronization with geographically separated transmitters. These developments together with results from real data experiments motivate the vision of fusion networks that develop network self-awareness by exploiting the rich information content of measurements collected for situation awareness.
 Uney, M., Mulgrew, B., and Clark, D., Latent parameter estimation in distributed fusion networks using separable likelihoods, IEEE Transactions on Signal and Information Processing Over Networks, December 2018.
 Uney, M., Mulgrew, B., and Clark, D., A cooperative approach to sensor localisation in distributed fusion networks, IEEE Transactions on Signal Processing 64, pp. 1187-1199, March 2016.
 Uney, M., Copsey, K., Page, S., Mulgrew, B., Thomas, P., Enabling self-configuration of fusion networks via scalable opportunistic sensor calibration, SPIE Defence+Security 2018, Orlando, April 2018.
 Kim, K., Uney, M., Mulgrew, B., Opportunistic synchronisation of multi-static staring array radars via track-before-detect, IEEE ICASSP 2018, Calgary, April 2018.
 Kim, K., Uney, M., Mulgrew, B., Detection via simultaneous trajectory estimation and long time integration, IEEE Transactions on Aerospace and Electronic Systems, under review.