Papers by Melanie Bocquel
HAL (Le Centre pour la Communication Scientifique Directe), Sep 1, 2011
Over the last few decades multi-target tracking (MTT) has proved to be a challenging and attracti... more Over the last few decades multi-target tracking (MTT) has proved to be a challenging and attractive research topic. MTT applications span a wide variety of disciplines, including robotics, radar/sonar surveillance, computer vision and biomedical research. The primary focus of this dissertation is to develop an effective and efficient multi-target tracking algorithm dealing with an unknown and time-varying number of targets. The emerging and promising Random Finite Set (RFS) framework provides a rigorous foundation for optimal Bayes multi-target tracking. In contrast to traditional approaches, the collection of individual targets is treated as a set-valued state. The intent of this dissertation is two-fold; first to assert that the RFS framework not only is a natural, elegant and rigorous foundation, but also leads to practical, efficient and reliable algorithms for Bayesian multi-target tracking, and second to provide several novel RFS based tracking algorithms suitable for the specific Track-Before-Detect (TBD) surveillance application. One main contribution of this dissertation is a rigorous derivation and practical implementation of a novel algorithm well suited to deal with multi-target tracking problems for a given cardinality. The proposed Interacting Population-based MCMC-PF algorithm makes use of several Metropolis-Hastings samplers running in parallel, which interact through genetic variation. Another key contribution concerns the design and implementation of two novel algorithms to handle a varying number of targets. The first approach exploits Reversible Jumps. The second approach is built upon the concepts of labeled RFSs and multiple cardinality hypotheses. The performance of the proposed algorithms is also demonstrated in practical scenarios, and shown to significantly outperform conventional multi-target PF in terms of track accuracy and consistency. The final contribution seeks to exploit external information to increase the performance of the surveillance system. In multi-target scenarios, kinematic constraints from the interaction of targets with their environment or other targets can restrict target motion. Such motion constraint information is integrated by using a fixed-lag smoothing procedure, named Knowledge-Based Fixed-Lag Smoother (KB-Smoother). The proposed combination IP-MCMC-PF/KB-Smoother yields enhanced tracking.
International Conference on Information Fusion, Jul 9, 2013
ABSTRACT In this paper we address the problem of tracking multiple targets based on raw measureme... more ABSTRACT In this paper we address the problem of tracking multiple targets based on raw measurements by means of Particle filtering. Bayesian multitarget tracking, in the Random Finite Set framework, propagates the multitarget posterior density recursively in time. Sequential Monte Carlo (SMC) approximations of the optimal filter are computationally expensive and lead to high-variance estimates as the number of targets increases. We present an extension of the Interacting Population-based MCMC-PF (IP-MCMC-PF) [1]. This extension exploits reversible jumps. Incorporation of Reversible Jump MCMC (RJMCMC) [2] methods into a tracking framework gives the possibility to deal with multiple appearing and disappearing targets, and makes the statistical inference more tractable. In our case, the technique is adopted to efficiently solve the high-dimensional state estimation problem, where the estimation of the existence and positions of many targets from a sequence of noisy measurements is required. Simulation analyses demonstrate that the proposed IP-RJMCMC-PF yields higher consistency, accuracy and reliability in multitarget tracking.
In this paper, we have addressed the problem of multiple target tracking in Track-Before-Detect (... more In this paper, we have addressed the problem of multiple target tracking in Track-Before-Detect (TBD) context using ambiguous Radar data. TBD is a method which uses raw measurement data, i.e. reflected target power, to track targets. Tracking can be defined as the estimation of the state of a moving object based on measurements. These measurements are in this case assumed to be the radar echoes ambiguous in range and doppler. The estimated states are produced by means of a tracking filter. The filtering problem has been solved by using a Particle Filter (PF). Particle filtering is a signal processing methodology, which can effectively deal with nonlinear and non-Gaussian signals by a sample-based approximation of the state probability density function (pdf). A standard multitarget SIR Particle Filter is extended so that it can handle range/doppler ambiguities and eclipsing effects. Such extension is required for its use in practice and to enhance tracking accuracy. The proposed particle filter succeeds in resolving range and doppler ambiguities, detecting and tracking multiple targets in a TBD context.
International Conference on Information Fusion, Jul 7, 2014
This paper proposes techniques to improve the properties of Sequential Markov Chain Monte Carlo (... more This paper proposes techniques to improve the properties of Sequential Markov Chain Monte Carlo (SMCMC) methods in the context of multi-target tracking. In particular, we extend the Interacting Population-based MCMC Particle Filter (IP-MCMC-PF) with three different methods: delayed rejection, genetic algorithms, and simulated annealing. Each of these methods furnishes the IP-MCMC-PF algorithm with different theoretical guarantees which are empirically analysed in this paper. Firstly, the use of delayed rejection in the Metropolis-Hastings (MH) samplers is proposed in order to reduce the asymptotic variance of the estimate. Secondly, the crossover operator, inspired by genetic algorithms, is presented as a mechanism to increase the interaction of the MH samplers. Thus, attaining fast convergence of the time-consuming MCMC step. Thirdly, simulated annealing is introduced with the goal of increasing the robustness of the algorithm against divergence due to e.g. poor initialisations. Finally, the results from our experiments show that the proposed methods strengthen the multi-target tracker in the aforementioned aspects.
Tracking of multiple objects via particle filtering faces the difficulty of dealing effectively w... more Tracking of multiple objects via particle filtering faces the difficulty of dealing effectively with high dimensional state spaces. One efficient solution consists of integrating Markov chain Monte Carlo (MCMC) sampling at the core of the particle filter. To accomplish such integration, a few different approaches have been proposed in the literature during the last decade. In this paper, we introduce the derivation of the acceptance probability of the interacting population MCMC particle filter (IP-MCMC-PF), one of the most recent approaches to MCMC-based particle filtering. Additionally, we show that the previous expression known in the literature was incomplete.
International Conference on Information Fusion, Jul 6, 2015
This paper introduces the Langevin Monte Carlo Filter (LMCF), a particle filter with a Markov cha... more This paper introduces the Langevin Monte Carlo Filter (LMCF), a particle filter with a Markov chain Monte Carlo algorithm which draws proposals by simulating Hamiltonian dynamics. This approach is well suited to non-linear filtering problems in high dimensional state spaces where the bootstrap filter requires an impracticably large number of particles. The simulation of Hamiltonian dynamics is motivated by leveraging more model knowledge in the proposal design. In particular, the gradient of the posterior energy function is used to draw new samples with high probability of acceptance. Furthermore, the introduction of auxiliary variables (the so-called momenta) ensures that new samples do not collapse at a single mode of the posterior density. In comparison with random-walk Metropolis, the LMC algorithm has been proven more efficient as the state dimension increases. Therefore, we are able to verify through experiments that our LMCF is able to attain multi-target tracking using small number of particles when other MCMC-based particle filters relying on random-walk Metropolis require a considerably larger particle number. As a conclusion, we claim that performing little additional work for each particle (in our case, computing likelihood energy gradients) turns out to be very effective as it allows to greatly reduce the number of particles while improving tracking performance.
Multi-target tracking requires the joint estimation of the number of target trajectories and thei... more Multi-target tracking requires the joint estimation of the number of target trajectories and their states from a sequence of observations. In low signal-to-noise ratio (SNR) scenarios, the poor detection probability and large number of false observations can greatly degrade the tracking performance. In this case an approach called Track-Before-Detect (TBD) that operates on the pre-detection signal, is needed. In this paper we present a labeled random finite set solution to the multitarget TBD problem. To the best of our knowledge this is the first provably Bayes optimal approach to multi-target tracking using image data. Simulation results using realistic radar-based TBD scenarios are also presented to demonstrate the capability of the proposed approach.
IEEE Transactions on Signal Processing, Jun 1, 2014
ABSTRACT In this work, we are interested in the improvements attainable when multiscan processing... more ABSTRACT In this work, we are interested in the improvements attainable when multiscan processing of external knowledge is performed over a moving time window. We propose a novel algorithm that enforces the state constraints by using a Fixed-Lag Smoothing procedure within the prediction step of the Bayesian recursion. For proving the improvements, we utilize differential entropy as a measure of uncertainty and show that the approach guarantees a lower or equal posterior differential entropy than classical single-step constrained filtering. Simulation results using examples for single-target tracking are presented to verify that a Sequential Monte Carlo implementation of the proposed algorithm guarantees an improved tracking accuracy.
IEEE Journal of Selected Topics in Signal Processing, Jun 1, 2013
ABSTRACT Exploitation of external knowledge through constrained filtering guarantees improved per... more ABSTRACT Exploitation of external knowledge through constrained filtering guarantees improved performance. In this paper we show how multiscan processing of such information further enhances the track accuracy. This can be achieved using a Fixed-Lag Smoothing procedure, and a proof of improvement is given in terms of entropy reduction. Such multiscan algorithm, i.e., named KB-Smoother (“Fixed-lag smoothing for Bayes optimal exploitation of external knowledge,” F. Papi , Proc. 15th Int. Conf. Inf. Fusion, 2012) can be implemented by means of a SIR-PF. In practice, the SIR-PF suffers from depletion problems, which are further amplified by the Smoothing technique. Sequential MCMC methods represent an efficient alternative to the standard SIR-PF approach. Furthermore, by borrowing techniques from genetic algorithms, a fully parallelizable multitarget tracker can be defined. Such approach, i.e., named Interacting Population (IP)-MCMC-PF, was first introduced in “Multitarget tracking with interacting population-based MCMC-PF” (M Bocquel , Proc. 15th Int. Conf. Inf. Fusion, 2012). In this paper, we propose and analyze a combination of the KB-Smoother along with the IP-MCMC-PF. As will be shown, the combination of the two methods yields an improved track accuracy while mitigating the loss of particles diversity. Simulation analyses for single and multitarget tracking scenarios confirm the benefits of the proposed approach.
arXiv (Cornell University), May 1, 2017
Track-before-detect (TBD) is a powerful approach that consists in providing the tracker with sens... more Track-before-detect (TBD) is a powerful approach that consists in providing the tracker with sensor measurements directly without pre-detection. Due to the measurement model non-linearities, online state estimation in TBD is most commonly solved via particle filtering. Existing particle filters for TBD do not incorporate measurement information in their proposal distribution. The Langevin Monte Carlo (LMC) is a sampling method whose proposal is able to exploit all available knowledge of the posterior (that is, both prior and measurement information). This letter synthesizes recent advances in LMCbased filtering to describe the Riemann-Langevin particle filter and introduces its novel application to TBD. The benefits of our approach are illustrated in a challenging low-noise scenario. Index Terms-particle filter, Langevin Monte Carlo, trackbefore-detect (TBD).
International Conference on Information Fusion, Aug 5, 2016
Track fusion is the problem of combining tracks based on different sensor observations. In the se... more Track fusion is the problem of combining tracks based on different sensor observations. In the sequential Monte Carlo framework, track fusion is solved by either imposing linear or Gaussian assumptions, or relying on kernel density estimation (KDE). In this paper, we introduce a novel track fusion algorithm suited to the hierarchical multi-sensor architecture. The algorithm can be incorporated in the particle filtering framework without restricting the densities by imposing assumptions, or requiring the non-trivial selection of additional parameters, as e.g., is needed in KDE. Furthermore, the proposed method is equivalent to the optimal centralised fusion architecture, in which all sensor measurements are communicated to the fusion node. Numerical results show that the newly proposed method outperforms the existing methods either by reducing estimation errors or by reducing the computation time significantly.
International Conference on Information Fusion, Jul 9, 2012
Particle Filters (PFs) nowadays represent the state of art in nonlinear filtering. In particular,... more Particle Filters (PFs) nowadays represent the state of art in nonlinear filtering. In particular, their high flexibility makes PFs particularly suited for Bayes optimal exploitation of possibly available external knowledge. In this paper we propose a new method for optimal processing of external knowledge that can be formalized in terms of hard constraints on the system dynamics. In particular, we are interested in the tracking performance improvements attainable when forward processing of external knowledge is performed over a moving window at every time step. That is, the one step ahead prediction of each particle is obtained through a Fixed-Lag Smoothing procedure, which uses Pseudo-Measurements to evaluate the level of adherence between each particle trajectory and the knowledge over multiple scans. A proof of improvements is presented by utilizing differential entropy [1] as a measure of uncertainty. That is, we show that the differential entropy of the posterior PDF targeted by the proposed approach is always lower or equal to the differential entropy of the posterior PDF usually targeted in constrained filtering. Thus, for a sufficiently large number of particles, a PF implementation of the proposed Knowledge-Based Fixed-Lag Smoother can only improve the track accuracy upon classical algorithms for constrained filtering. Preliminary simulations show that the proposed approach guarantees substantial improvements when compared to the Standard SISR-PF and to the Pseudo-Measurements PF.
Dans cet article, nous nous interessons a la detection et au pistage d'une cible en contexte ... more Dans cet article, nous nous interessons a la detection et au pistage d'une cible en contexte Track-Before-Detect (TBD). Nous proposons ici un filtre particulaire efficace, fonde sur le choix d'une loi instrumentale pertinente motivee par des considerations de detection radar et permettant un gain significatif par rapport aux lois classiquement utilisees dans la litterature, notamment en terme de rapidite de convergence du filtre pour la detection. Nous determinons egalement un nombre minimal de particules requis pour garantir des performances de detection interessantes.
In this paper we address the problem of tracking multiple targets based on raw measurements by me... more In this paper we address the problem of tracking multiple targets based on raw measurements by means of Particle filtering. Bayesian multitarget tracking, in the Random Finite Set framework, propagates the multitarget posterior density recursively in time. Sequential Monte Carlo (SMC) approximations of the optimal filter are computationally expensive and lead to high-variance estimates as the number of targets increases. We propose a novel, efficient and reliable Labeled RFS based tracking algorithms suitable for, among others, the TBD surveillance application. This algorithm uses the Interacting Population based MCMC-PF (IP-MCMC-PF), first introduced in [6], as the core engine of a Multiple Cardinality Hypotheses Tracker (MCHT), where each cardinality is treated independently. The proposed multi-target filter is built upon the concept of labeled Random Finite Set (RFS) [40], [41], and formally incorporates the propagation and estimation of track labels within the RFS filtering framework. Simulation analyses demonstrate that the proposed Multiple Cardinality Hypotheses Particle Filter (MCHPF) yields higher consistency, accuracy and reliability in multitarget tracking.
Proceedings of the 16th International Conference on Information Fusion, 2013
In this paper we address the problem of tracking multiple targets based on raw measurements by me... more In this paper we address the problem of tracking multiple targets based on raw measurements by means of Particle filtering. Bayesian multitarget tracking, in the Random Finite Set framework, propagates the multitarget posterior density recursively in time. Sequential Monte Carlo (SMC) approximations of the optimal filter are computationally expensive and lead to high-variance estimates as the number of targets increases. We present an extension of the Interacting Population-based MCMC-PF (IP-MCMC-PF) [1]. This extension exploits reversible jumps. Incorporation of Reversible Jump MCMC (RJMCMC) [2] methods into a tracking framework gives the possibility to deal with multiple appearing and disappearing targets, and makes the statistical inference more tractable. In our case, the technique is adopted to efficiently solve the high-dimensional state estimation problem, where the estimation of the existence and positions of many targets from a sequence of noisy measurements is required. Si...
2016 19th International Conference on Information Fusion (FUSION), 2016
Track fusion is the problem of combining tracks based on different sensor observations. In the se... more Track fusion is the problem of combining tracks based on different sensor observations. In the sequential Monte Carlo framework, track fusion is solved by either imposing linear or Gaussian assumptions, or relying on kernel density estimation (KDE). In this paper, we introduce a novel track fusion algorithm suited to the hierarchical multi-sensor architecture. The algorithm can be incorporated in the particle filtering framework without restricting the densities by imposing assumptions, or requiring the non-trivial selection of additional parameters, as e.g., is needed in KDE. Furthermore, the proposed method is equivalent to the optimal centralised fusion architecture, in which all sensor measurements are communicated to the fusion node. Numerical results show that the newly proposed method outperforms the existing methods either by reducing estimation errors or by reducing the computation time significantly.
Particle Filters (PFs) nowadays represent the state of art in nonlinear filtering. In particular,... more Particle Filters (PFs) nowadays represent the state of art in nonlinear filtering. In particular, their high flexibility makes PFs particularly suited for Bayes optimal exploitation of possibly available external knowledge. In this paper we propose a new method for optimal processing of external knowledge that can be formalized in terms of hard constraints on the system dynamics. In particular, we are interested in the tracking performance improvements attainable when forward processing of external knowledge is performed over a moving window at every time step. That is, the one step ahead prediction of each particle is obtained through a Fixed-Lag Smoothing procedure, which uses Pseudo-Measurements to evaluate the level of adherence between each particle trajectory and the knowledge over multiple scans. A proof of improvements is presented by utilizing differential entropy as a measure of uncertainty. That is, we show that the differential entropy of the posterior PDF targeted by th...
This paper introduces the Langevin Monte Carlo Filter (LMCF), a particle filter with a Markov cha... more This paper introduces the Langevin Monte Carlo Filter (LMCF), a particle filter with a Markov chain Monte Carlo algorithm which draws proposals by simulating Hamiltonian dynamics. This approach is well suited to non-linear filtering problems in high dimensional state spaces where the bootstrap filter requires an impracticably large number of particles. The simulation of Hamiltonian dynamics is motivated by leveraging more model knowledge in the proposal design. In particular, the gradient of the posterior energy function is used to draw new samples with high probability of acceptance. Furthermore, the introduction of auxiliary variables (the so-called momenta) ensures that new samples do not collapse at a single mode of the posterior density. In comparison with random-walk Metropolis, the LMC algorithm has been proven more efficient as the state dimension increases. Therefore, we are able to verify through experiments that our LMCF is able to attain multi-target tracking using small...
This paper proposes techniques to improve the properties of Sequential Markov Chain Monte Carlo (... more This paper proposes techniques to improve the properties of Sequential Markov Chain Monte Carlo (SMCMC) methods in the context of multi-target tracking. In particular, we extend the Interacting Population-based MCMC Particle Filter (IP-MCMC-PF) with three different methods: delayed rejection, genetic algorithms, and simulated annealing. Each of these methods furnishes the IP-MCMC-PF algorithm with different theoretical guarantees which are empirically analysed in this paper. Firstly, the use of delayed rejection in the Metropolis-Hastings (MH) samplers is proposed in order to reduce the asymptotic variance of the estimate. Secondly, the crossover operator, inspired by genetic algorithms, is presented as a mechanism to increase the interaction of the MH samplers. Thus, attaining fast convergence of the time-consuming MCMC step. Thirdly, simulated annealing is introduced with the goal of increasing the robustness of the algorithm against divergence due to e.g. poor initialisations. Fi...
Uploads
Papers by Melanie Bocquel