Tremendous advances in wearable computing and storage technologies enable us to record not just snapshots of an event but the whole human experience for a long period of time. Such a \life-logandamp;quot; system captures important events as they happen, rather than an after-thought. Such a system has applications in many areas such as law enforcement, personal archives, police questioning, and medicine. Much of the existing eandamp;reg;orts focus on the pattern recognition and information retrieval aspects of the system. On the other hand, the privacy issues raised by such an intrusive system have not received much attention from the research community. The objectives of this research project are two-fold: andamp;macr;rst, to construct a wearable life-log video system, and second, to provide a solution for protecting the identity of the subjects in the video while keeping the video useful. In this thesis work, we designed a portable wearable life-log system that implements audio distortion and face blocking in a real time to protect the privacy of the subjects who are being recorded in life-log video. For audio, our system automatically isolates the subject's speech and distorts it using a pitch- shifting algorithm to conceal the identity. For video, our system uses a real-time face detection, tracking and blocking algorithm to obfuscate the faces of the subjects. Extensive experiments have been conducted on interview videos to demonstrate the ability of our system in protecting the identity of the subject while maintaining the usability of the life-log video.

With technological advances digital video and imaging are becoming more and more relevant. Medical, remote-learning, surveillance, conferencing and home monitoring are just a few applications of these technologies. Along with compression, there is now a need for analysis and extraction of data. During the days of film and early digital cameras the processing and manipulation of data from such cameras was transparent to the end user. This transparency has been decreasing and the industry is moving towards `smart users' - people who will be enabled to program and manipulate their video and imaging systems. Smart cameras can currently zoom, refocus and adjust lighting by sourcing out current from the camera itself to the headlight. Such cameras are used in the industry for inspection, quality control and even counting objects in jewelry stores and museums, but could eventually allow user defined programmability. However, all this will not happen without interactive software as well as capabilities in the hardware to allow programmability. In this research, compression, expansion and detail extraction from videos in the surveillance arena are addressed. Here, a video codec is defined that can embed contextual details of a video stream depending on user defined requirements creating a video summary. This codec also carries out motion based segmentation that helps in object detection. Once an object is segmented it is matched against a database using its shape and color information. If the object is not a good match, the user can either add it to the database or consider it an anomaly. RGB vector angle information is used to generate object descriptors to match objects to a database. This descriptor implicitly incorporates the shape and color information while keeping the size of the database manageable. Color images of objects that are considered `safe' are taken from various angles and distances (with the same background as that covered by the camera is question) and their RGB vector angle based descriptors constitute the information contained in the database. This research is a first step towards building a compression and detection system for specific surveillance applications. While the user has to build and maintain a database, there are no restrictions on the size of the images, zoom and angle requirements, thus, reducing the burden on the end user in creating such a database. This also allows use of different types of cameras and doesn't need a lot of up-front planning on camera location, etc.

The rise of deep learning has facilitated remarkable progress in video understanding. This thesis addresses three important tasks of video understanding: video object detection, joint object and action detection, and spatio-temporal action localization. Object class detection is one of the most important challenges in computer vision. Object detectors are usually trained on bounding-boxes from still images. Recently, video has been used as an alternative source of data. Yet, training an object detector on one domain (either still images or videos) and testing on the other one results in a significant performance gap compared to training and testing on the same domain. In the first part of this thesis, we examine the reasons behind this performance gap. We define and evaluate several domain shift factors: spatial location accuracy, appearance diversity, image quality, aspect distribution, and object size and camera framing. We examine the impact of these factors by comparing the detection performance before and after cancelling them out. The results show that all five factors affect the performance of the detectors and their combined effect explains the performance gap. While most existing approaches for detection in videos focus on objects or human actions separately, in the second part of this thesis we aim at detecting non-human centric actions, i.e., objects performing actions, such as cat eating or dog jumping. We introduce an end-to-end multitask objective that jointly learns object-action relationships. We compare it with different training objectives, validate its effectiveness for detecting object-action pairs in videos, and show that both tasks of object and action detection benefit from this joint learning. In experiments on the A2D dataset [Xu et al., 2015], we obtain state-of-the-art results on segmentation of object-action pairs. In the third part, we are the first to propose an action tubelet detector that leverages the temporal continuity of videos instead of operating at the frame level, as state-of-the-art approaches do. The same way modern detectors rely on anchor boxes, our tubelet detector is based on anchor cuboids by taking as input a sequence of frames and outputing tubelets, i.e., sequences of bounding boxes with associated scores. Our tubelet detector outperforms all state of the art on the UCF-Sports [Rodriguez et al., 2008], J-HMDB [Jhuang et al., 2013a], and UCF-101 [Soomro et al., 2012] action localization datasets especially at high overlap thresholds. The improvement in detection performance is explained by both more accurate scores and more precise localization.

Nous abordons le problème de la synchronisation automatique d'un contenu audiovisuel avec une procédure textuelle qui le décrit. La stratégie consiste à extraire des informations sur la structure des deux contenus puis à les mettre en correspondance. Nous proposons deux outils d'analyse vidéo qui extraient respectivement : * les limites des évènements d'intérêt à l'aide d'une méthode de quantification de type dictionnaire * les segments dans lesquels une action se répète en exploitant une méthode d'analyse fréquentielle : le YIN. Ensuite, nous proposons un système de synchronisation qui fusionne les informations fournies par ces outils pour établir des associations entre les instructions textuelles et les segments vidéo correspondants. Une "Matrice de confiance" est construite et exploitée de manière récursive pour établir ces associations en regard de leur fiabilité. / We address the problem of automatic synchronization of an audiovisual content with a procedural text that describes it. The strategy consists in extracting pieces of information about the structure from both contents, and in matching them depending on their types. We propose two video analysis tools that respectively extract: * Limits of events of interest using an approach inspired by dictionary quantization. * Segments that enclose a repeated action based on the YIN frequency analysis method. We then propose a synchronization system that merges results coming from these tools in order to establish links between textual instructions and the corresponding video segments. To do so, a "Confidence Matrix" is built and recursively processed in order to identify these links in respect with their reliability.

The thesis is divided into two parts theoretical one and practical one. The theoretical part describes methods of implementation of physical experiments, also there is discussion about curriculum that was used in the created experiments and description of video analysis with used app, at the end of this part there is also several experiments that are made to inspire teachers and to show them video analysis options. For some of the experiments there is also graphical manual. The practical part contains a set of differentiated tasks and worksheets. In this work there is also few worksheets that were filled by students.

The thesis is divided into two parts theoretical one and practical one. The theoretical part describes methods of implementation of physical experiments, also there is discussion about curriculum that was used in the created experiments, at the end of this part there is also several experiments that are made to inspire teachers and to show them video analysis options. The practical part contains a set of differentiated tasks and worksheets. In this work there is also few worksheets that were filled by students.

Le travail de thèse que nous présentons s’articule autour de l’utilisation de caméras motorisées à trois degrés de liberté, également appelées caméras PTZ. Ces caméras peuvent être pilotées suivant deux angles. L’angle de panorama (θ) permet une rotation autour d’un axe vertical et l’angle de tangage (ϕ) permet une rotation autour d’un axe horizontal. Si, théoriquement, ces caméras permettent donc une vue omnidirectionnelle, elles limitent le plus souvent la rotation suivant l’angle de panorama mais surtout suivant l’angle de tangage. En plus du pilotage des rotations, ces caméras permettent également de contrôler la distance focale permettant ainsi un degré de liberté supplémentaire. Par rapport à d’autres modèles, les caméras PTZ permettent de construire un panorama - représentation étendue d’une scène construite à partir d’une collection d’image - de très grande résolution. La première étape dans la construction d’un panorama est l’acquisition des différentes prises de vue. A cet effet, nous avons réalisé une étude théorique permettant une couverture optimale de la sphère à partir de surfaces rectangulaires en limitant les zones de recouvrement. Cette étude nous permet de calculer une trajectoire optimale de la caméra et de limiter le nombre de prises de vues nécessaires à la représentation de la scène. Nous proposons également différents traitements permettant d’améliorer sensiblement le rendu et de corriger la plupart des défauts liés à l’assemblage d’une collection d’images acquises avec des paramètres de prises de vue différents. Une part importante de notre travail a été consacrée au recalage automatique d’images en temps réel, c’est à dire que chaque étapes est effectuée en moins de 40ms pour permettre le traitement de 25 images par seconde. La technologie que nous avons développée permet d’obtenir un recalage particulièrement précis avec un temps d’exécution de l’ordre de 4ms (AMD1.8MHz). Enfin, nous proposons deux applications de suivi d’objets en mouvement directement issues de nos travaux de recherche. La première associe une caméra PTZ à un miroir sphérique. L’association de ces deux éléments permet de détecter tout objet en mouvement dans la scène puis de se focaliser sur l’un d’eux. Dans le cadre de cette application, nous proposons un algorithme de calibrage automatique de l’ensemble caméra et miroir. La deuxième application n’exploite que la caméra PTZ et permet la segmentation et le suivi des objets dans la scène pendant le mouvement de la caméra. Par rapport aux applications classiques de suivi de cible en mouvement avec une caméra PTZ, notre approche se différencie par le fait que réalisons une segmentation fine des objets permettant leur classification. / The thesis considers the of use motorized cameras with 3 degrees of freedom which are commonly called PTZ cameras. The orientation of such cameras is controlled according to two angles: the panorama angle (θ) describes the degree of rotation around on vertical axis and the tilt angle (ϕ) refers to rotation along a meridian line. Theoretically, these cameras can cover an omnidirectional field of vision of 4psr. Generally, the panorama angle and especially the tilt angle are limited for such cameras. In addition to control of the orientation of the camera, it is also possible to control focal distance, thus allowing an additional degree of freedom. Compared to other material, PTZ cameras thus allow one to build a panorama of very high resolution. A panorama is a wide representation of a scene built starting from a collection of images. The first stage in the construction of a panorama is the acquisition of the various images. To this end, we made a theoretical study to determine the optimal paving of the sphere with rectangular surfaces to minimize the number of zones of recovery. This study enables us to calculate an optimal trajectory of the camera and to limit the number of images necessary to the representation of the scene. We also propose various processing techniques which appreciably improve the rendering of the mosaic image and correct the majority of the defaults related to the assembly of a collection of images which were acquired with differing image capture parameters. A significant part of our work was used to the automatic image registration in real time, i.e. lower than 40ms. The technology that we developed makes it possible to obtain a particularly precise image registration with an computation time about 4ms (AMD1.8MHz). Our research leads directly to two proposed applications for the tracking of moving objects. The first involves the use of a PTZ camera and a spherical mirror. The combination of these two elements makes it possible to detect any motion object in the scene and to then to focus itself on one of them. Within the framework of this application, we propose an automatic algorithm of calibration of the system. The second application exploits only PTZ camera and allows the segmentation and the tracking of the objects in the scene during the movement of the camera. Compared to the traditional applications of motion detection with a PTZ camera, our approach is different by the fact that it compute a precise segmentation of the objects allowing their classification.

Les travaux présentés dans ce manuscrit traitent de la détection de personnes dans des séquences d’images et de l’analyse de leur activité. Ces travaux ont été menés au sein de l’institut PRISME dans le cadre du projet CAPTHOM du pôle de compétitivité S2E2. Après un état de l’art sur l’analyse de séquences d’images pour l’interprétation automatique de scènes et une étude comparative de modules de vidéo-surveillance, nous présentons la méthode de détection de personnes proposée dans le cadre du projet CAPTHOM. Celle-ci s’articule autour de trois étapes : la détection de changement, le suivi d’objets mobiles et la classification. Chacune de ces étapes est décrite dans ce manuscrit. Ce système a été évalué sur une large base de vidéos correspondant à des scénarios de cas d’usage de CAPTHOM établis par les partenaires du projet. Ensuite, nous présentons des méthodes permettant d’obtenir, à partir du flux vidéo d’une ou deux caméras, d’autres informations de plus haut-niveau sur l’activité des personnes détectées. Nous présentons tout d’abord une mesure permettant de quantifier leur activité. Ensuite, un système de stéréovision multi-capteurs combinant une caméra infrarouge et une caméra visible est utilisé pour augmenter les performances du système de détection mais aussi pour permettre la localisation dans l’espace des personnes et donc accéder à une cartographie de leurs déplacements. Finalement, une méthode de détection d’événements anormaux, basée sur des statistiques de distributions spatiales et temporelles des pixels de l’avant-plan est détaillée. Les méthodes proposées offrent un panel de solutions performantes sur l’extraction d’informations haut-niveau à partir de séquences d’images. / The work presented in this manuscript deals with people detection and activity analysis in images sequences. This work has been done in the PRISME institut within the framework of the CAPTHOM project of the French Cluster S2E2. After a state of the art on video analysis and a comparative study of several video surveillance tools, we present the people detection method proposed within the framework of the CAPTHOM project. This method is based on three steps : change detection, mobile objects tracking and classification. Each steps is described in this thesis. The system was assessed on a wide videos dataset. Then, we present methods used to obtain other high-level information concerning the activity of detected persons. A criterion for characterizing their activity is presented. Then, a multi-sensors stereovision system combining an infrared and a daylight camera is used to increase performances of the people detection system but also to localize persons in the 3D space and so build the moving cartography. Finally, an abnormal events detection method based on statistics about spatio-temporal foreground pixel distribution is presented. These proposed methods offer robust and efficient solutions on high-level information extraction from images sequences.

Height estimation of objects can be implemented both for soft-biometrics and as an object tracking feature. In first case we can eliminate some possible subjects having considerably different height than the observed one, and focus on determining more distinctive remote identification features, like colour, face or ear, and search for similar subjects in a smaller set of possible candidates. For object tracking it can be used for temporal and spatial correspondence analysis as well or simultaneously for both in case of having different cameras. In this thesis we propose a novel method for automatic estimation of height using an uncalibrated camera. Nowadays such cameras can be found in any corner for different purposes like as for security reasons. A crucial moment in height estimation is finding vanishing points. In the method we use RANSAC to estimate best vanishing point from several estimated candidate points. The method has the new advantages that from different frames and their respective height estimations, automatically determines certain reasonable heights depending on height measurements distribution. With spreading of camera implementation in common applications, we believe this new software can be widely applied in as many fields as it can be imagined.

The article introduces a general survey of concepts used in traffic monitoring applications. It describes different approaches for solving particular steps of vehicle detection process. Analysis of these methods was performed. Furthermore this work focuses on the design and realization of complex robust algorithm for real-time vehicle detection. It is based on analysis of video-sequence acquired from static camera situated on highway. Processing consists of many steps. It starts with background subtraction and ends with traffic monitoring results, i.e. average speed, number of cars, level of service etc.