Learning a Family of Detectors

Object detection and recognition are important problems in computer vision. The challenges of these problems come from the presence of noise, background clutter, large within class variations of the object class and limited training data. In addition, the computational complexity in the recognition process is also a concern in practice. In this thesis, we propose one approach to handle the problem of detecting an object class that exhibits large within-class variations, and a second approach to speed up the classification processes. In the first approach, we show that foreground-background classification (detection) and within-class classification of the foreground class (pose estimation) can be jointly solved with using a multiplicative form of two kernel functions. One kernel measures similarity for foreground-background classification. The other kernel accounts for latent factors that control within-class variation and implicitly enables feature sharing among foreground training samples. For applications where explicit parameterization of the within-class states is unavailable, a nonparametric formulation of the kernel can be constructed with a proper foreground distance/similarity measure. Detector training is accomplished via standard Support Vector Machine learning. The resulting detectors are tuned to specific variations in the foreground class. They also serve to evaluate hypotheses of the foreground state. When the image masks for foreground objects are provided in training, the detectors can also produce object segmentation. Methods for generating a representative sample set of detectors are proposed that can enable efficient detection and tracking. In addition, because individual detectors verify hypotheses of foreground state, they can also be incorporated in a tracking-by-detection frame work to recover foreground state in image sequences. To run the detectors efficiently at the online stage, an input-sensitive speedup strategy is proposed to select the most relevant detectors quickly. The proposed approach is tested on data sets of human hands, vehicles and human faces. On all data sets, the proposed approach achieves improved detection accuracy over the best competing approaches. In the second part of the thesis, we formulate a filter-and-refine scheme to speed up recognition processes. The binary outputs of the weak classifiers in a boosted detector are used to identify a small number of candidate foreground state hypotheses quickly via Hamming distance or weighted Hamming distance. The approach is evaluated in three applications: face recognition on the face recognition grand challenge version 2 data set, hand shape detection and parameter estimation on a hand data set, and vehicle detection and estimation of the view angle on a multi-pose vehicle data set. On all data sets, our approach is at least five times faster than simply evaluating all foreground state hypotheses with virtually no loss in classification accuracy.

Tracking Human Body Pose on a Learned Smooth Space

Particle filtering is a popular method used in systems for tracking human body pose in video. One key difficulty in using particle filtering is caused by the curse of dimensionality: generally a very large number of particles is required to adequately approximate the underlying pose distribution in a high-dimensional state space. Although the number of degrees of freedom in the human body is quite large, in reality, the subset of allowable configurations in state space is generally restricted by human biomechanics, and the trajectories in this allowable subspace tend to be smooth. Therefore, a framework is proposed to learn a low-dimensional representation of the high-dimensional human poses state space. This mapping can be learned using a Gaussian Process Latent Variable Model (GPLVM) framework. One important advantage of the GPLVM framework is that both the mapping to, and mapping from the embedded space are smooth; this facilitates sampling in the low-dimensional space, and samples generated in the low-dimensional embedded space are easily mapped back into the original highdimensional space. Moreover, human body poses that are similar in the original space tend to be mapped close to each other in the embedded space; this property can be exploited when sampling in the embedded space. The proposed framework is tested in tracking 2D human body pose using a Scaled Prismatic Model. Experiments on real life video sequences demonstrate the strength of the approach. In comparison with the Multiple Hypothesis Tracking and the standard Condensation algorithm, the proposed algorithm is able to maintain tracking reliably throughout the long test sequences. It also handles singularity and self occlusion robustly.

Target Detection and Tracking With Heterogeneous Sensors

We present a multimodal detection and tracking algorithm for sensors composed of a camera mounted between two microphones. Target localization is performed on color-based change detection in the video modality and on time difference of arrival (TDOA) estimation between the two microphones in the audio modality. The TDOA is computed by multiband generalized cross correlation (GCC) analysis. The estimated directions of arrival are then postprocessed using a Riccati Kalman filter. The visual and audio estimates are finally integrated, at the likelihood level, into a particle filter (PF) that uses a zero-order motion model, and a weighted probabilistic data association (WPDA) scheme. We demonstrate that the Kalman filtering (KF) improves the accuracy of the audio source localization and that the WPDA helps to enhance the tracking performance of sensor fusion in reverberant scenarios. The combination of multiband GCC, KF, and WPDA within the particle filtering framework improves the performance of the algorithm in noisy scenarios. We also show how the proposed audiovisual tracker summarizes the observed scene by generating metadata that can be transmitted to other network nodes instead of transmitting the raw images and can be used for very low bit rate communication. Moreover, the generated metadata can also be used to detect and monitor events of interest.

Multicamera sport player tracking with Bayesian estimation of measurements

We propose a complete application capable of tracking multiple objects in an environment monitored by multiple cameras. The system has been specially developed to be applied to sport games, and it has been evaluated in a real association-football stadium. Each target is tracked using a local importance-sampling particle filter in each camera, but the final estimation is made by combining information from the other cameras using a modified unscented Kalman filter algorithm. Multicamera integration enables us to compensate for bad measurements or occlusions in some cameras thanks to the other views it offers. The final algorithm results in a more accurate system with a lower failure rate. (C) 2009 Society of Photo-Optical Instrumentation Engineers. [DOI: 10.1117/1.3114605]

Exploiting spatio-temporal constraints for robust 2D pose tracking

We present a Spatio-temporal 2D Models Framework (STMF) for 2D-Pose tracking. Space and time are discretized and a mixture of probabilistic "local models" is learnt associating 2D Shapes and 2D Stick Figures. Those spatio-temporal models generalize well for a particular viewpoint and state of the tracked action but some spatio-temporal discontinuities can appear along a sequence, as a direct consequence of the discretization. To overcome the problem, we propose to apply a Rao-Blackwellized Particle Filter (RBPF) in the 2D-Pose eigenspace, thus interpolating unseen data between view-based clusters. The fitness to the images of the predicted 2D-Poses is evaluated combining our STMF with spatio-temporal constraints. A robust, fast and smooth human motion tracker is obtained by tracking only the few most important dimensions of the state space and by refining deterministically with our STMF.

Mean field approach for tracking similar objects

In this paper, we consider the problem of tracking similar objects. We show how a mean field approach can be used to deal with interacting targets and we compare it with Markov Chain Monte Carlo (MCMC). Two mean field implementations are presented. The first one is more general and uses particle filtering. We discuss some simplifications of the base algorithm that reduce the computation time. The second one is based on suitable Gaussian approximations of probability densities that lead to a set of self-consistent equations for the means and covariances. These equations give the Kalman solution if there is no interaction. Experiments have been performed on two kinds of sequences. The first kind is composed of a single long sequence of twenty roaming ants and was previously analysed using MCMC. In this case, our mean field algorithms obtain substantially better results. The second kind corresponds to selected sequences of a football match in which the interaction avoids tracker coalescence in situations where independent trackers fail.

Robust visual tracking based on online learning sparse representation

Handling appearance variations is a very challenging problem for visual tracking. Existing methods usually solve this problem by relying on an effective appearance model with two features: (1) being capable of discriminating the tracked target from its background, (2) being robust to the target's appearance variations during tracking. Instead of integrating the two requirements into the appearance model, in this paper, we propose a tracking method that deals with these problems separately based on sparse representation in a particle filter framework. Each target candidate defined by a particle is linearly represented by the target and background templates with an additive representation error. Discriminating the target from its background is achieved by activating the target templates or the background templates in the linear system in a competitive manner. The target's appearance variations are directly modeled as the representation error. An online algorithm is used to learn the basis functions that sparsely span the representation error. The linear system is solved via ℓ1 minimization. The candidate with the smallest reconstruction error using the target templates is selected as the tracking result. We test the proposed approach using four sequences with heavy occlusions, large pose variations, drastic illumination changes and low foreground-background contrast. The proposed approach shows excellent performance in comparison with two latest state-of-the-art trackers.

Generalized Laplacian Eigenmaps for Modeling and Tracking Human Motions

This paper presents generalized Laplacian eigenmaps, a novel dimensionality reduction approach designed to address stylistic variations in time series. It generates compact and coherent continuous spaces whose geometry is data-driven. This paper also introduces graph-based particle filter, a novel methodology conceived for efficient tracking in low dimensional space derived from a spectral dimensionality reduction method. Its strengths are a propagation scheme, which facilitates the prediction in time and style, and a noise model coherent with the manifold, which prevents divergence, and increases robustness. Experiments show that a combination of both techniques achieves state-of-the-art performance for human pose tracking in underconstrained scenarios.

Characterisation of gas and particle emissions from wildfires

Os incêndios florestais são uma importante fonte de emissão de compostos gasosos e de aerossóis. Em Portugal, onde a maioria dos incêndios ocorre no norte e centro do país, os incêndios destroem todos os anos milhares de hectares, com importantes perdas em termos económicos, de vidas humanas e qualidade ambiental. As emissões podem alterar consideravelmente a química da atmosfera, degradar a qualidade do ar e alterar o clima. Contudo, a informação sobre as caraterísticas das emissões dos incêndios florestais nos países do Mediterrâneo é limitada. Tanto a nível nacional como internacional, existe um interesse crescente na elaboração de inventários de emissões e de regulamentos sobre as emissões de carbono para a atmosfera. Do ponto de vista atmosférico da monitorização atmosférica, os incêndios são considerados um desafio, dada a sua variabilidade temporal e espacial, sendo de esperar um aumento da sua frequência, dimensão e severidade, e também porque as estimativas de emissões dependem das caraterísticas dos biocombustíveis e da fase de combustão. O objetivo deste estudo foi quantificar e caraterizar as emissões de gases e aerossóis de alguns dos mais representativos incêndios florestais que ocorreram no centro de Portugal nos verões de 2009 e de 2010. Efetuou-se a colheita de amostras de gases e de duas frações de partículas (PM2.5 e PM2.5-10) nas plumas de fumo em sacos Tedlar e em filtros de quartzo acoplados a um amostrador de elevado volume, respetivamente. Os hidrocarbonetos totais (THC) e óxidos de carbono (CO e CO2) nas amostras gasosas foram analisados em instrumentos automáticos de ionização de chama e detetores não dispersivos de infravermelhos, respetivamente. Para algumas amostras, foram também quantificados alguns compostos de carbonilo após reamostragem do gás dos sacos Tedlar em cartuchos de sílica gel revestidos com 2,4-dinitrofenilhidrazina (DNPH), seguida de análise por cromatografia líquida de alta resolução. Nas partículas, analisou-se o carbono orgânico e elementar (técnica termo-óptica), iões solúveis em água (cromatografia iónica) e elementos (espectrometria de massa com plasma acoplado por indução ou análise instrumental por ativação com neutrões). A especiação orgânica foi obtida por cromatografia gasosa acoplada a espectrometria de massa após extração com recurso a vários solventes e separação dos extratos orgânicos em diversas classes de diferentes polaridades através do fracionamento com sílica gel. Os fatores de emissão do CO e do CO2 situaram-se nas gamas 52-482 e 822-1690 g kg-1 (base seca), mostrando, respetivamente, correlação negativa e positiva com a eficiência de combustão. Os fatores de emissão dos THC apresentaram valores mais elevados durante a fase de combustão latente sem chama, oscilando entre 0.33 e 334 g kg-1 (base seca). O composto orgânico volátil oxigenado mais abundante foi o acetaldeído com fatores de emissão que variaram desde 1.0 até 3.2 g kg-1 (base seca), seguido pelo formaldeído e o propionaldeído. Observou-se que as emissões destes compostos são promovidas durante a fase de combustão latente sem chama. Os fatores de emissão de PM2.5 e PM10 registaram valores entre 0.50-68 e 0.86-72 g kg-1 (base seca), respetivamente. A emissão de partículas finas e grosseiras é também promovida em condições de combustão lenta. As PM2.5 representaram cerca de 90% da massa de partículas PM10. A fração carbonosa das partículas amostradas em qualquer dos incêndios foi claramente dominada pelo carbono orgânico. Foi obtida uma ampla gama de rácios entre o carbono orgânico e o carbono elementar, dependendo das condições de combustão. Contudo, todos os rácios refletiram uma maior proporção de carbono orgânico em relação ao carbono elementar, típica das emissões de queima de biomassa. Os iões solúveis em água obtidos nas partículas da pluma de fumo contribuíram com valores até 3.9% da massa de partículas PM2.5 e 2.8% da massa de partículas de PM2.5-10. O potássio contribuiu com valores até 15 g mg-1 PM2.5 e 22 g mg-1 PM2.5-10, embora em massa absoluta estivesse maioritariamente presente nas partículas finas. Os rácios entre potássio e carbono elementar e entre potássio e carbono orgânico obtidos nas partículas da pluma de fumo enquadram-se na gama de valores relatados na literatura para emissões de queima de biomassa. Os elementos detetados nas amostras representaram, em média, valores até 1.2% e 12% da massa de PM2.5 e PM2.5-10, respetivamente. Partículas resultantes de uma combustão mais completa (valores elevados de CO2 e baixos de CO) foram caraterizadas por um elevado teor de constituintes inorgânicos e um menor conteúdo de matéria orgânica. Observou-se que a matéria orgânica particulada é composta principalmente por componentes fenólicos e produtos derivados, séries de compostos homólogos (alcanos, alcenos, ácidos alcanóicos e alcanóis), açúcares, biomarcadores esteróides e terpenóides, e hidrocarbonetos aromáticos policíclicos. O reteno, um biomarcador das emissões da queima de coníferas, foi o hidrocarboneto aromático dominante nas amostras das plumas de fumo amostradas durante a campanha que decorreu em 2009, devido ao predomínio de amostras colhidas em incêndios em florestas de pinheiros. O principal açúcar anidro, e sempre um dos compostos mais abundantes, foi o levoglucosano. O rácio levoglucosano/OC obtido nas partículas das plumas de fumo, em média, registaram valores desde 5.8 a 23 mg g-1 OC. Os rácios levoglucosano/manosano e levoglucosano/(manosano+galactosano) revelaram o predomínio de amostras provenientes da queima de coníferas. Tendo em conta que a estimativa das emissões dos incêndios florestais requer um conhecimento de fatores de emissão apropriados para cada biocombustível, a base de dados abrangente obtida neste estudo é potencialmente útil para atualizar os inventários de emissões. Tem vindo a ser observado que a fase de combustão latente sem chama, a qual pode ocorrer simultaneamente com a fase de chama e durar várias horas ou dias, pode contribuir para uma quantidade considerável de poluentes atmosféricos, pelo que os fatores de emissão correspondentes devem ser considerados no cálculo das emissões globais de incêndios florestais. Devido à falta de informação detalhada sobre perfis químicos de emissão, a base de dados obtida neste estudo pode também ser útil para a aplicação de modelos no recetor no sul da Europa.

Formation Control Driven by Cooperative Object Tracking

In this paper we introduce a formation control loop that maximizes the performance of the cooperative perception of a tracked target by a team of mobile robots, while maintaining the team in formation, with a dynamically adjustable geometry which is a function of the quality of the target perception by the team. In the formation control loop, the controller module is a distributed non-linear model predictive controller and the estimator module fuses local estimates of the target state, obtained by a particle filter at each robot. The two modules and their integration are described in detail, including a real-time database associated to a wireless communication protocol that facilitates the exchange of state data while reducing collisions among team members. Simulation and real robot results for indoor and outdoor teams of different robots are presented. The results highlight how our method successfully enables a team of homogeneous robots to minimize the total uncertainty of the tracked target cooperative estimate while complying with performance criteria such as keeping a pre-set distance between the teammates and the target, avoiding collisions with teammates and/or surrounding obstacles.

Characterizing Dynamic Optimization Benchmarks for the Comparison of Multi-Modal Tracking Algorithms

Population-based metaheuristics, such as particle swarm optimization (PSO), have been employed to solve many real-world optimization problems. Although it is of- ten sufficient to find a single solution to these problems, there does exist those cases where identifying multiple, diverse solutions can be beneficial or even required. Some of these problems are further complicated by a change in their objective function over time. This type of optimization is referred to as dynamic, multi-modal optimization. Algorithms which exploit multiple optima in a search space are identified as niching algorithms. Although numerous dynamic, niching algorithms have been developed, their performance is often measured solely on their ability to find a single, global optimum. Furthermore, the comparisons often use synthetic benchmarks whose landscape characteristics are generally limited and unknown. This thesis provides a landscape analysis of the dynamic benchmark functions commonly developed for multi-modal optimization. The benchmark analysis results reveal that the mechanisms responsible for dynamism in the current dynamic bench- marks do not significantly affect landscape features, thus suggesting a lack of representation for problems whose landscape features vary over time. This analysis is used in a comparison of current niching algorithms to identify the effects that specific landscape features have on niching performance. Two performance metrics are proposed to measure both the scalability and accuracy of the niching algorithms. The algorithm comparison results demonstrate the algorithms best suited for a variety of dynamic environments. This comparison also examines each of the algorithms in terms of their niching behaviours and analyzing the range and trade-off between scalability and accuracy when tuning the algorithms respective parameters. These results contribute to the understanding of current niching techniques as well as the problem features that ultimately dictate their success.

Towards the automatic reconstruction of dendritic trees using particle filters

The 3D reconstruction of a Golgi-stained dendritic tree from a serial stack of images captured with a transmitted light bright-field microscope is investigated. Modifications to the bootstrap filter are discussed such that the tree structure may be estimated recursively as a series of connected segments. The tracking performance of the bootstrap particle filter is compared against Differential Evolution, an evolutionary global optimisation method, both in terms of robustness and accuracy. It is found that the particle filtering approach is significantly more robust and accurate for the data considered.

The magnetized steel and scintillator calorimeters of the MINOS experiment

The Main Injector Neutrino Oscillation Search (MINOS) experiment uses an accelerator-produced neutrino beam to perform precision measurements of the neutrino oscillation parameters in the ""atmospheric neutrino"" sector associated with muon neutrino disappearance. This long-baseline experiment measures neutrino interactions in Fermilab`s NuMI neutrino beam with a near detector at Fermilab and again 735 km downstream with a far detector in the Soudan Underground Laboratory in northern Minnesota. The two detectors are magnetized steel-scintillator tracking calorimeters. They are designed to be as similar as possible in order to ensure that differences in detector response have minimal impact on the comparisons of event rates, energy spectra and topologies that are essential to MINOS measurements of oscillation parameters. The design, construction, calibration and performance of the far and near detectors are described in this paper. (C) 2008 Elsevier B.V. All rights reserved.

Application of radial return mapping algorithm for finite strain elastoplasticity in a hybrid finite element and particle-in-cell model

The radial return mapping algorithm within the computational context of a hybrid Finite Element and Particle-In-Cell (FE/PIC) method is constructed to allow a fluid flow FE/PIC code to be applied solid mechanic problems with large displacements and large deformations. The FE/PIC method retains the robustness of an Eulerian mesh and enables tracking of material deformation by a set of Lagrangian particles or material points. In the FE/PIC approach the particle velocities are interpolated from nodal velocities and then the particle position is updated using a suitable integration scheme, such as the 4th order Runge-Kutta scheme[1]. The strain increments are obtained from gradients of the nodal velocities at the material point positions, which are then used to evaluate the stress increment and update history variables. To obtain the stress increment from the strain increment, the nonlinear constitutive equations are solved in an incremental iterative integration scheme based on a radial return mapping algorithm[2]. A plane stress extension of a rectangular shape J2 elastoplastic material with isotropic, kinematic and combined hardening is performed as an example and for validation of the enhanced FE/PIC method. It is shown that the method is suitable for analysis of problems in crystal plasticity and metal forming. The method is specifically suitable for simulation of neighbouring microstructural phases with different constitutive equations in a multiscale material modelling framework.