Visual target tracking problem consists of two topics: Obtaining targets from camera measurements and target tracking. Even though it has been studied for more than 30 years, there are still some problems not completely solved. Especially in the case of multiple targets, association of measurements to targets, creation of new targets and deletion of old ones are among those. What is more, it is very important to deal with the occlusion and crossing targets problems suitably. We believe that a slightly modified version of multiple hypothesis tracking can successfully deal with most of the aforementioned problems with sufficient success. Distance, track size, track color, gate size and track history are used as parameters to evaluate the hypotheses generated for measurement to track association problem whereas size and color are used as parameters for occlusion problem. The overall tracker has been fine tuned over some scenarios and it has been observed that it performs well over the testing scenarios as well. Furthermore the performance of the tracker is analyzed according to those parameters in both association and occlusion handling situations.

In this thesis, novel methods for background modeling, tracking, occlusion handling and event recognition via multi-camera configurations are presented. As the initial step, building blocks of typical single camera surveillance systems that are moving object detection, tracking and event recognition, are discussed and various widely accepted methods for these building blocks are tested to asses on their performance. Next, for the multi-camera surveillance systems, background modeling, occlusion handling, tracking and event recognition for two-camera configurations are examined. Various foreground detection methods are discussed and a background modeling algorithm, which is based on multi-variate mixture of Gaussians, is proposed. During occlusion handling studies, a novel method for segmenting the occluded objects is proposed, in which a top-view of the scene, free of occlusions, is generated from multi-view data. The experiments indicate that the occlusion handling algorithm operates successfully on various test data. A novel tracking method by using multi-camera configurations is also proposed. The main idea of multi-camera employment is fusing the 2D information coming from the cameras to obtain a 3D information for better occlusion handling and seamless tracking. The proposed algorithm is tested on different data sets and it shows clear improvement over single camera tracker. Finally, multi-camera trajectories of objects are classified by proposed multi-camera event recognition method. In this method, concatenated different view trajectories are used to train Gaussian Mixture Hidden Markov Models. The experimental results indicate an improvement for the multi-camera event recognition performance over the event recognition by using single camera.

In this study, primary steps of a visual surveillance system are presented: moving object detection and tracking of these moving objects. Background subtraction has been performed to detect the moving objects in the video, which has been taken from a static camera. Four methods, frame differencing, running (moving) average, eigenbackground subtraction and mixture of Gaussians, have been used in the background subtraction process. After background subtraction, using some additional operations, such as morphological operations and connected component analysis, the objects to be tracked have been acquired. While tracking the moving objects, active contour models (snakes) has been used as one of the approaches. In addition to this method / Kalman tracker and mean-shift tracker are other approaches which have been utilized. A new approach has been proposed for the problem of tracking multiple targets. We have implemented this method for single and multiple camera configurations. Multiple cameras have been used to augment the measurements. Homography matrix has been calculated to find the correspondence between cameras. Then, measurements and tracks have been associated by the new tracking method.

Video surveillance has long been in use to monitor security sensitive areas such as banks, department stores, crowded public places and borders. The rise in computer speed, availability of cheap large-capacity storage devices and high speed network infrastructure enabled the way for cheaper, multi sensor video surveillance systems. In this thesis, the problem of tracking multiple targets with multiple cameras has been discussed. Cameras have been located so that they have overlapping fields of vision. A dynamic background-modeling algorithm is described for segmenting moving objects from the background, which is capable of adapting to dynamic scene changes and periodic motion, such as illumination change and swaying of trees. After segmentation of foreground scene, the objects to be tracked have been acquired by morphological operations and connected component analysis. For the purpose of tracking the moving objects, an active contour model (snakes) is one of the approaches, in addition to a Kalman tracker. As the main tracking algorithm, a rule based tracker has been developed first for a single camera, and then extended to multiple cameras. Results of used and proposed methods are given in detail.

Pour les véhicules intelligents autonomes ou semi-autonomes, la perception constitue la première tâche fondamentale à accomplir avant la décision et l’action. Grâce à l’analyse des données vidéo, Lidar et radar, elle fournit une représentation spécifique de l’environnement et de son état, à travers l’extraction de propriétés clés issues des données des capteurs. Comparé à d’autres modalités de perception telles que le GPS, les capteurs inertiels ou les capteurs de distance (Lidar, radar, ultrasons), les caméras offrent la plus grande quantité d’informations. Grâce à leur polyvalence, les caméras permettent aux systèmes intelligents d’extraire à la fois des informations contextuelles de haut niveau et de reconstruire des informations géométriques de la scène observée et ce, à haute vitesse et à faible coût. De plus, la technologie de détection passive des caméras permet une faible consommation d’énergie et facilite leur miniaturisation. L’utilisation des caméras n’est toutefois pas triviale et pose un certain nombre de questions théoriques liées à la façon dont ce capteur perçoit son environnement. Dans cette thèse, nous proposons un système de détection d’objets mobiles basé seule- ment sur l’analyse d’images. En effet, dans les environnements observés par un véhicule intelligent, les objets en mouvement représentent des obstacles avec un risque de collision élevé, et ils doivent être détectés de manière fiable et robuste. Nous abordons le problème de la détection d’objets mobiles à partir de l’extraction du contexte local reposant sur une segmentation de la route. Après transformation de l’image couleur en une image invariante à l’illumination, les ombres peuvent alors être supprimées réduisant ainsi leur influence négative sur la détection d’obstacles. Ainsi, à partir d’une sélection automatique de pixels appartenant à la route, une région d’intérêt où les objets en mouvement peuvent apparaître avec un risque de collision élevé, est extraite. Dans cette zone, les pixels appartenant à des objets mobiles sont ensuite identifiés à l’aide d’une approche plan+parallaxe. À cette fin, les pixels potentiellement mobiles et liés à l’effet de parallaxe sont détectés par une méthode de soustraction du fond de l’image; puis trois contraintes géométriques différentes: la contrainte épipolaire, la contrainte de cohérence structurelle et le tenseur trifocal, sont appliquées à ces pixels pour filtrer ceux issus de l’effet de parallaxe. Des équations de vraisemblance sont aussi proposées afin de combiner les différents contraintes d’une manière complémentaire et efficace. Lorsque la stéréovision est disponible, la segmentation de la route et la détection d’obstacles peuvent être affinées en utilisant une segmentation spécifique de la carte de disparité. De plus, dans ce cas, un algorithme de suivi robuste combinant les informations de l’image et la profondeur des pixels a été proposé. Ainsi, si l’une des deux caméras ne fonctionne plus, le système peut donc revenir dans un mode de fonctionnement monoculaire ce qui constitue une propriété importante pour la fiabilité et l’intégrité du système de perception. Les différents algorithmes proposés ont été testés sur des bases de données d’images publiques en réalisant une évaluation par rapport aux approches de l’état de l’art et en se comparant à des données de vérité terrain. Les résultats obtenus sont prometteurs et montrent que les méthodes proposées sont efficaces et robustes pour différents scénarios routiers et les détections s’avèrent fiables notamment dans des situations ambiguës. / For autonomous or semi-autonomous intelligent vehicles, perception constitutes the first fundamental task to be performed before decision and action/control. Through the analysis of video, Lidar and radar data, it provides a specific representation of the environment and of its state, by extracting key properties from sensor data with time integration of sensor information. Compared to other perception modalities such as GPS, inertial or range sensors (Lidar, radar, ultrasonic), the cameras offer the greatest amount of information. Thanks to their versatility, cameras allow intelligent systems to achieve both high-level contextual and low-level geometrical information about the observed scene, and this is at high speed and low cost. Furthermore, the passive sensing technology of cameras enables low energy consumption and facilitates small size system integration. The use of cameras is however, not trivial and poses a number of theoretical issues related to how this sensor perceives its environmen. In this thesis, we propose a vision-only system for moving object detection. Indeed,within natural and constrained environments observed by an intelligent vehicle, moving objects represent high risk collision obstacles, and have to be handled robustly. We approach the problem of detecting moving objects by first extracting the local contextusing a color-based road segmentation. After transforming the color image into illuminant invariant image, shadows as well as their negative influence on the detection process can be removed. Hence, according to the feature automatically selected onthe road, a region of interest (ROI), where the moving objects can appear with a high collision risk, is extracted. Within this area, the moving pixels are then identified usin ga plane+parallax approach. To this end, the potential moving and parallax pixels a redetected using a background subtraction method; then three different geometrical constraints : the epipolar constraint, the structural consistency constraint and the trifocaltensor are applied to such potential pixels to filter out parallax ones. Likelihood equations are also introduced to combine the constraints in a complementary and effectiveway. When stereo vision is available, the road segmentation and on-road obstacles detection can be refined by means of the disparity map with geometrical cues. Moreover, in this case, a robust tracking algorithm combining image and depth information has been proposed. If one of the two cameras fails, the system can therefore come back to a monocular operation mode, which is an important feature for perception system reliability and integrity. The different proposed algorithms have been tested on public images data set with anevaluation against state-of-the-art approaches and ground-truth data. The obtained results are promising and show that the proposed methods are effective and robust on the different traffic scenarios and can achieve reliable detections in ambiguous situations.

Dans ce manuscrit de thèse, nous présentons un descripteur robuste pour la soustraction d’arrière-plan qui est capable de décrire la texture à partir d’une séquence d’images. Ce descripteur est moins sensible aux bruits et produit un histogramme court, tout en préservant la robustesse aux changements d’éclairage. Un autre descripteur pour la reconnaissance dynamique des textures est également proposé. Le descripteur permet d’extraire non seulement des informations de couleur, mais aussi des informations plus détaillées provenant des séquences vidéo. Enfin, nous présentons une approche de sélection de caractéristiques basée sur le principe d'apprentissage par ensemble qui est capable de sélectionner les caractéristiques appropriées pour chaque pixel afin de distinguer les objets de premier plan de l’arrière plan. En outre, notre proposition utilise un mécanisme pour mettre à jour l’importance relative de chaque caractéristique au cours du temps. De plus, une approche heuristique est utilisée pour réduire la complexité de la maintenance du modèle d’arrière-plan et aussi sa robustesse. Par contre, cette méthode nécessite un grand nombre de caractéristiques pour avoir une bonne précision. De plus, chaque classificateur de base apprend un ensemble de caractéristiques au lieu de chaque caractéristique individuellement. Pour compenser ces limitations, nous avons amélioré cette approche en proposant une nouvelle méthodologie pour sélectionner des caractéristiques basées sur le principe du « wagging ». Nous avons également adopté une approche basée sur le concept de « superpixel » au lieu de traiter chaque pixel individuellement. Cela augmente non seulement l’efficacité en termes de temps de calcul et de consommation de mémoire, mais aussi la qualité de la détection des objets mobiles. / In this thesis, we present a robust descriptor for background subtraction which is able to describe texture from an image sequence. The descriptor is less sensitive to noisy pixels and produces a short histogram, while preserving robustness to illumination changes. Moreover, a descriptor for dynamic texture recognition is also proposed. This descriptor extracts not only color information, but also a more detailed information from video sequences. Finally, we present an ensemble for feature selection approach that is able to select suitable features for each pixel to distinguish the foreground objects from the background ones. Our proposal uses a mechanism to update the relative importance of each feature over time. For this purpose, a heuristic approach is used to reduce the complexity of the background model maintenance while maintaining the robustness of the background model. However, this method only reaches the highest accuracy when the number of features is huge. In addition, each base classifier learns a feature set instead of individual features. To overcome these limitations, we extended our previous approach by proposing a new methodology for selecting features based on wagging. We also adopted a superpixel-based approach instead of a pixel-level approach. This does not only increases the efficiency in terms of time and memory consumption, but also can improves the segmentation performance of moving objects.

Dans ce manuscrit de thèse, nous introduisons les avancées récentes sur la décomposition en matrices (et tenseurs) de rang faible et parcimonieuse ainsi que les contributions pour faire face aux principaux problèmes dans ce domaine. Nous présentons d’abord un aperçu des méthodes matricielles et tensorielles les plus récentes ainsi que ses applications sur la modélisation d’arrière-plan et la segmentation du premier plan. Ensuite, nous abordons le problème de l’initialisation du modèle de fond comme un processus de reconstruction à partir de données manquantes ou corrompues. Une nouvelle méthodologie est présentée montrant un potentiel intéressant pour l’initialisation de la modélisation du fond dans le cadre de VSI. Par la suite, nous proposons une version « double contrainte » de l’ACP robuste pour améliorer la détection de premier plan en milieu marin dans des applications de vidéo-surveillance automatisées. Nous avons aussi développé deux algorithmes incrémentaux basés sur tenseurs afin d’effectuer une séparation entre le fond et le premier plan à partir de données multidimensionnelles. Ces deux travaux abordent le problème de la décomposition de rang faible et parcimonieuse sur des tenseurs. A la fin, nous présentons un travail particulier réalisé en conjonction avec le Centre de Vision Informatique (CVC) de l’Université Autonome de Barcelone (UAB). / This thesis introduces the recent advances on decomposition into low-rank plus sparse matrices and tensors, as well as the main contributions to face the principal issues in moving object detection. First, we present an overview of the state-of-the-art methods for low-rank and sparse decomposition, as well as their application to background modeling and foreground segmentation tasks. Next, we address the problem of background model initialization as a reconstruction process from missing/corrupted data. A novel methodology is presented showing an attractive potential for background modeling initialization in video surveillance. Subsequently, we propose a double-constrained version of robust principal component analysis to improve the foreground detection in maritime environments for automated video-surveillance applications. The algorithm makes use of double constraints extracted from spatial saliency maps to enhance object foreground detection in dynamic scenes. We also developed two incremental tensor-based algorithms in order to perform background/foreground separation from multidimensional streaming data. These works address the problem of low-rank and sparse decomposition on tensors. Finally, we present a particular work realized in conjunction with the Computer Vision Center (CVC) at Autonomous University of Barcelona (UAB).

The topic of this thesis is the recognition and detection of moving object and persons in video sequence and in the static image. Designed application uses the combination of background model for movement detection, histograms of oriented gradients method for person recognition and Lucas-Kanade method for object tracking.

In this dissertation, we discuss the problem of robust linear subspace estimation using low-rank optimization and propose three formulations of it. We demonstrate how these formulations can be used to solve fundamental computer vision problems, and provide superior performance in terms of accuracy and running time. Consider a set of observations extracted from images (such as pixel gray values, local features, trajectories . . . etc). If the assumption that these observations are drawn from a liner subspace (or can be linearly approximated) is valid, then the goal is to represent each observation as a linear combination of a compact basis, while maintaining a minimal reconstruction error. One of the earliest, yet most popular, approaches to achieve that is Principal Component Analysis (PCA). However, PCA can only handle Gaussian noise, and thus suffers when the observations are contaminated with gross and sparse outliers. To this end, in this dissertation, we focus on estimating the subspace robustly using low-rank optimization, where the sparse outliers are detected and separated through the `1 norm. The robust estimation has a two-fold advantage: First, the obtained basis better represents the actual subspace because it does not include contributions from the outliers. Second, the detected outliers are often of a specific interest in many applications, as we will show throughout this thesis. We demonstrate four different formulations and applications for low-rank optimization. First, we consider the problem of reconstructing an underwater sequence by removing the iii turbulence caused by the water waves. The main drawback of most previous attempts to tackle this problem is that they heavily depend on modelling the waves, which in fact is ill-posed since the actual behavior of the waves along with the imaging process are complicated and include several noise components; therefore, their results are not satisfactory. In contrast, we propose a novel approach which outperforms the state-of-the-art. The intuition behind our method is that in a sequence where the water is static, the frames would be linearly correlated. Therefore, in the presence of water waves, we may consider the frames as noisy observations drawn from a the subspace of linearly correlated frames. However, the noise introduced by the water waves is not sparse, and thus cannot directly be detected using low-rank optimization. Therefore, we propose a data-driven two-stage approach, where the first stage “sparsifies” the noise, and the second stage detects it. The first stage leverages the temporal mean of the sequence to overcome the structured turbulence of the waves through an iterative registration algorithm. The result of the first stage is a high quality mean and a better structured sequence; however, the sequence still contains unstructured sparse noise. Thus, we employ a second stage at which we extract the sparse errors from the sequence through rank minimization. Our method converges faster, and drastically outperforms state of the art on all testing sequences. Secondly, we consider a closely related situation where an independently moving object is also present in the turbulent video. More precisely, we consider video sequences acquired in a desert battlefields, where atmospheric turbulence is typically present, in addition to independently moving targets. Typical approaches for turbulence mitigation follow averaging or de-warping techniques. Although these methods can reduce the turbulence, they distort the independently moving objects which can often be of great interest. Therefore, we address the iv problem of simultaneous turbulence mitigation and moving object detection. We propose a novel three-term low-rank matrix decomposition approach in which we decompose the turbulence sequence into three components: the background, the turbulence, and the object. We simplify this extremely difficult problem into a minimization of nuclear norm, Frobenius norm, and `1 norm. Our method is based on two observations: First, the turbulence causes dense and Gaussian noise, and therefore can be captured by Frobenius norm, while the moving objects are sparse and thus can be captured by `1 norm. Second, since the object’s motion is linear and intrinsically different than the Gaussian-like turbulence, a Gaussian-based turbulence model can be employed to enforce an additional constraint on the search space of the minimization. We demonstrate the robustness of our approach on challenging sequences which are significantly distorted with atmospheric turbulence and include extremely tiny moving objects. In addition to robustly detecting the subspace of the frames of a sequence, we consider using trajectories as observations in the low-rank optimization framework. In particular, in videos acquired by moving cameras, we track all the pixels in the video and use that to estimate the camera motion subspace. This is particularly useful in activity recognition, which typically requires standard preprocessing steps such as motion compensation, moving object detection, and object tracking. The errors from the motion compensation step propagate to the object detection stage, resulting in miss-detections, which further complicates the tracking stage, resulting in cluttered and incorrect tracks. In contrast, we propose a novel approach which does not follow the standard steps, and accordingly avoids the aforementioned diffi- culties. Our approach is based on Lagrangian particle trajectories which are a set of dense trajectories obtained by advecting optical flow over time, thus capturing the ensemble motions v of a scene. This is done in frames of unaligned video, and no object detection is required. In order to handle the moving camera, we decompose the trajectories into their camera-induced and object-induced components. Having obtained the relevant object motion trajectories, we compute a compact set of chaotic invariant features, which captures the characteristics of the trajectories. Consequently, a SVM is employed to learn and recognize the human actions using the computed motion features. We performed intensive experiments on multiple benchmark datasets, and obtained promising results. Finally, we consider a more challenging problem referred to as complex event recognition, where the activities of interest are complex and unconstrained. This problem typically pose significant challenges because it involves videos of highly variable content, noise, length, frame size . . . etc. In this extremely challenging task, high-level features have recently shown a promising direction as in [53, 129], where core low-level events referred to as concepts are annotated and modelled using a portion of the training data, then each event is described using its content of these concepts. However, because of the complex nature of the videos, both the concept models and the corresponding high-level features are significantly noisy. In order to address this problem, we propose a novel low-rank formulation, which combines the precisely annotated videos used to train the concepts, with the rich high-level features. Our approach finds a new representation for each event, which is not only low-rank, but also constrained to adhere to the concept annotation, thus suppressing the noise, and maintaining a consistent occurrence of the concepts in each event. Extensive experiments on large scale real world dataset TRECVID Multimedia Event Detection 2011 and 2012 demonstrate that our approach consistently improves the discriminativity of the high-level features by a significant margin.