Dans cette thèse nous étudions le problème de détection d’anomalies dans les données ouvertes utilisées par l’entreprise Qucit ; aussi bien les données métiers de ses clients, que celles permettant de les contextualiser. Dans un premier temps, nous nous sommes intéressés à la détection de vélos défectueux au sein des données de trajets du système de vélo en libre service de New York. Nous cherchons des données reflétant une anomalie dans la réalité. Des caractéristiques décrivant le comportement de chaque vélo observé sont partitionnés. Les comportements anormaux sont extraits depuis ce partitionnement et comparés aux rapports mensuels indiquant le nombre de vélos réparés ; c’est un problème d’apprentissage à sortie agrégée. Les résultats de ce premier travail se sont avérés insatisfaisant en raison de la pauvreté des données. Ce premier volet des travaux a ensuite laissé place à une problématique tournée vers la détection de bâtiments au sein d’images satellites. Nous cherchons des anomalies dans les données géographiques qui ne reflètent pas la réalité. Nous proposons une méthode de fusion de modèles de segmentation améliorant la métrique d’erreur jusqu’à +7% par rapport à la méthode standard. Nous évaluons la robustesse de notre modèle face à la suppression de bâtiments dans les étiquettes, afin de déterminer à quel point les omissions sont susceptibles d’en altérer les résultats. Ce type de bruit est communément rencontré au sein des données OpenStreetMap, régulièrement utilisées par Qucit, et la robustesse observée indique qu’il pourrait être corrigé. / In this thesis we study the problem of anomaly detection in the open data used by the Qucit company, both the business data of its customers, as well as those allowing to contextualize them.We are looking for data that reflects an anomaly in reality. Initially, we were interested in detecting defective bicycles in the trip data of New York’s bike share system. Characteristics describing the behaviour of each observed bicycle are clustered. Abnormal behaviors are extracted from this clustering and compared to monthly reports indicating the number of bikes repaired; this is an aggregate learning problem. The results of this first work were unsatisfactory due to the paucity of data. This first part of the work then gave way to a problem focused on the detection of buildings within satellite images. We are looking for anomalies in the geographical data that do not reflect reality. We propose a method of merging segmentation models that improves the error metric by up to +7% over the standard method. We assess the robustness of our model to the removal of buildings from labels to determine the extent to which omissions are likely to alter the results. This type of noise is commonly encountered within the OpenStreetMap data, regularly used by Qucit, and the robustness observed indicates that it could be corrected.

Cette thèse s'intéresse au problème de la reconnaissance d'objets dans les images vidéo et plus particulièrement à celui de leur localisation. Elle a été conduite dans le contexte d'une collaboration scientifique entre l'INRIA Rhône-Alpes et MBDA France. De ce fait, une attention particulière a été accordée à l’applicabilité des approches proposées aux images infra-rouges. La méthode de localisation proposée repose sur l'utilisation d'une fenêtre glissante incluant une cascade à deux étages qui, malgré sa simplicité, permet d'allier rapidité et précision. Le premier étage est un étage de filtrage rejetant la plupart des faux positifs au moyen d’un classifieur SVM linéaire. Le deuxième étage élimine les fausses détections laissées par le premier étage avec un classifieur SVM non-linéaire plus lent, mais plus performant. Les fenêtres sont représentées par des descripteurs HOG et Bag-of-words. La seconde contribution de la thèse réside dans une méthode permettant de combiner localisation d'objets et catégorisation d'images. Ceci permet, d'une part, de prendre en compte le contexte de l'image lors de la localisation des objets, et d'autre part de s'appuyer sur la structure géométrique des objets lors de la catégorisation des images. Cette méthode permet d'améliorer les performances pour les deux tâches et produit des détecteurs et classifieurs dont la performance dépasse celle de l'état de l'art. Finalement, nous nous penchons sur le problème de localisation de catégories d'objets similaires et proposons de décomposer la tâche de localisation d'objets en deux étapes. Une première étape de détection permet de trouver les objets sans déterminer leurs positions tandis qu’une seconde étape d’identification permet de prédire la catégorie de l'objet. Nous montrons que cela permet de limiter les confusions entre les classes, principal problème observé pour les catégories d'objets visuellement similaires. La thèse laisse une place importante à la validation expérimentale, conduites sur la base PASCAL VOC ainsi que sur des bases d’images spécifiquement réalisées pour la thèse. / This thesis addresses the problem of object recognition in images and more precisely the problem of object localization. It have been conducted in the context of a scientific collaboration between INRIA Rhônes-Alpes and MBDA France. Therefore, a particular attention was accorded to the applicability of the proposed approaches on infrared images. The localization method proposed here relies on the sliding windows mechanism combined with a two stage cascade that, despite its simplicity, allies rapidity and precision. The first stage is a filtering stage that rejects most of the false positives using a linear classifier. The second stage prunes the detections of the first classifier using a slower yet efficient non-linear classifier. Windows are represented with HOG and Bag-of-words descriptors. The second contribution of this thesis is a method that combines object localization and image categorization. This allows, on the one hand, to take into account context information in localization, and on the other hand, to rely on geometrical structure of objects while performing image categorization. This combination leads to a significant quality improvement and obtains performance superior to the state of the art for both tasks. Finally, we consider the problem of localizing visually similar object categories and suggest to decompose the task of object localization into two steps. The first is a detection step that allows to find objects without determining their category while the second step, an identification step, predicts the objects categories. We show that this approach limits inter-class confusion, which is the main difficulty faced when localizing visually similar object classes. This thesis accords an important place to experimental validation conducted on PASCAL VOC databases as well as other databases specifically introduced for the thesis.

Le problème du dénombrement d’un grand nombre de très petits objets en mouvement dans les vidéos est un contexte applicatif jusqu’à présent peu étudié.Dans ce cadre, la difficulté réside essentiellement dans le fait qu’en raison de leurs très petites tailles apparentes dans la vidéo, il n’est pas possible de définir un modèle géométrique fiable de ces objets. Or, les travaux existants dans le domaine de la détection d’objets dans des vidéo, utilisent souvent un tel modèle géométrique des objets d’intérêt. Les méthodes de détection existantes ne sont de ce fait pas applicables directement dans le cadre de la détection de tels très petits objets. Dans le cadre de cette thèse, il est proposé une méthodologie complète permettant la détection de nombreux petits objets, avec un cadre applicatif visant plus particulièrement la détection et le comptage d’oiseaux migrateurs dans une vidéo. Le principe innovant, proposé en tant qu’une solution de ce problème, consiste à associer des coefficients de fiabilité de détection aux objets pour les dénombrer tout en évitant de prendre en compte de trop nombreuses fausses détections. Un algorithme hiérarchique analysant l’aspect spatio-temporel d’objets (leurs apparence et l’évolution dans le temps) dans une vidéo à l’aide de méthodes de traitement d’images, de statistique et de la logique floue est ainsi proposé. Le but des coefficients de fiabilité est d’estimer la probabilité que les paramètres d’une détection correspondent aux paramètres attendus pour les objets d’intérêt. Finalement, l’ensemble des coefficients est converti en une valeur qui évalue la séquence du traitement d’un objet. La somme de ces valeurs correspond au nombre d’objets d’intérêt dans une vidéo. Les résultats obtenus montrent que les bonnes détections sont pour la plupart comprises dans le dénombrement avec des coefficients de fiabilité égaux ou proche de 1, et où les fausses détections sont supprimées ou sous-pondérés avec des coefficients de fiabilité plus faible. Les résultats de comptage dans des vidéos contenant de très nombreux oiseaux sont proches de la vérité terrain, ce qui prouve la validité de la solution proposée comme un moyen de dénombrement automatique d’objets dans des vidéos. / The problem of counting of big volumes of very small moving objects in videos is a domain, which was not studied to date. The difficulty of this application consists essentially in the fact, that because of very small sizes of objects, apparent in the videos, it is impossible to define a reliable geometric model of these objects. The researches, existing in the domain of object detection in videos frequently use a geometrical model of objects of interest.For this reason, the existing methods of object detection cannot be applied for the detection of very small objects in the study case. This thesis proposes a complete methodology, allowing the detection of very small objects in videos, and designed particularly the detection and counting of migrating birds in videos. An innovative principle and the solution of this problem consist in association of coefficients of detection reliability to the objects, in order to count them, avoiding counting of many false detections. The solution proposes a hierarchical algorithm, which analyses the spatial and temporal aspects of objects (their appearance and evolution in time) in a video, by the means of methods of image processing, statistics, and fuzzy logic. The aim of the reliability coefficients is to estimate the probability, that the parameters of a detected objects conform to the expected parameters of the objects of interest. Finally, the coefficients are put together and converted into a value, which evaluates the sequence of processing, applied to detect an object. The sum of these values corresponds to the number of the objects of interest in a video. The results show, that the most of correct detections are characterized in the counting by the reliability coefficient equal or close to 1. The results show, that the most of correct detections have their reliability coefficients close to 1, and the false detection are deleted or have low reliability coefficients. The counting results in the videos with numerous groups of migrating birds are close to the ground trough. This validates the proposed solution as a method of automatic counting of objects in videos.

Levees provide protection for vast amounts of commercial and residential properties. However, these structures degrade over time, due to the impact of severe weather, sand boils, subsidence of land, seepage, etc. In this research, we focus on detecting sand boils. Sand boils occur when water under pressure wells up to the surface through a bed of sand. These make levees especially vulnerable. Object detection is a good approach to confirm the presence of sand boils from satellite or drone imagery, which can be utilized to assist in the automated levee monitoring methodology. Since sand boils have distinct features, applying object detection algorithms to it can result in accurate detection. To the best of our knowledge, this research work is the first approach to detect sand boils from images. In this research, we compare some of the latest deep learning methods, Viola Jones algorithm, and other non-deep learning methods to determine the best performing one. We also train a Stacking-based machine learning method for the accurate prediction of sand boils. The accuracy of our robust model is 95.4%.

In this paper we present a component based person detection system that is capable of detecting frontal, rear and near side views of people, and partially occluded persons in cluttered scenes. The framework that is described here for people is easily applied to other objects as well. The motivation for developing a component based approach is two fold: first, to enhance the performance of person detection systems on frontal and rear views of people and second, to develop a framework that directly addresses the problem of detecting people who are partially occluded or whose body parts blend in with the background. The data classification is handled by several support vector machine classifiers arranged in two layers. This architecture is known as Adaptive Combination of Classifiers (ACC). The system performs very well and is capable of detecting people even when all components of a person are not found. The performance of the system is significantly better than a full body person detector designed along similar lines. This suggests that the improved performance is due to the components based approach and the ACC data classification structure.

Ground penetrating radar (GPR) systems are increasingly being used for the detection and location of buried objects within the upper regions of the earth’s surface. The antenna is the most critical component of such a system. This thesis presents a comprehensive study of resistor-loaded planar dipole antennas for GPR applications using both theory and experiments. The theoretical analysis is performed using the finite difference time domain (FDTD) technique. The analysis starts with the most popular planar dipole, the bow-tie. A parametric study is done to find out how the flare angle, length, and lumped resistors of the antenna should be selected to achieve broadband properties and good target detection with less clutter. The screening of the antenna and the position of transmitting and receiving antennas with respect to each other and ground surface are also studied. A number of other planar geometrical shapes are considered and compared with the bow-tie in order to find what geometrical shape gives the best performance. The FDTD simulations are carried out for both lossless and lossy, dispersive grounds. Also simulations are carried out including surface roughness and natural clutter like rocks and twigs to make the modeling more realistic. Finally, a pair of resistor-loaded bow-tie antennas is constructed and both indoor and outdoor measurements are carried out to validate the simulation results. / <p>QC 20100923</p>

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.

In this thesis, detection of traffic signs is studied. Since, both color and shape properties of traffic signs are distinctive / these two properties have been employed for detection. Detection using color properties is studied in two different color domains in order to examine and compare the advantages and the disadvantages of these domains for detection purposes. In addition to their color information, shape information is also employed for detection purpose. Edge information (obtained by using the Sobel Operator) of the images/frames is considered as search domain to find triangular, rectangular, octagonal and circular traffic signs. In order to improve the performance of detection process a joint implementation of shape and color based algorithms is utilized. Two different methods have been used v in order to combine these two features. Both of the algorithms help reducing the number of pixels to check whether they belong to a sign or not. This, of course, reduces the processing time of detection process. Each utilized algorithm is tested and compared with the others by using both static images from different sources and video streams. Images having adverse properties are used in order to state algorithms response for some specific conditions such as bad illumination and shadow. After implementation, results show that joint implementation of the color and shape based detection algorithms produces more accurate results. Moreover, joint implementation reduces the processing time of the detection process when compared to application of algorithms individually since it diminishes the search domain.