Recentemente surgiram dispositivos sensores de profundidade capazes de capturar textura e geometria de uma cena em tempo real. Com isso, diversas técnicas de Visão Computacional, que antes eram aplicadas apenas a texturas, agora são passíveis de uma reformulação, visando o uso também da geometria. Ao mesmo tempo em que tais algoritmos, tirando vantagem dessa nova tecnologia, podem ser acelerados ou tornarem-se mais robustos, surgem igualmente diversos novos desafios e problemas interessantes a serem enfrentados. Como exemplo desses dispositivos podemos citar o do Projeto Vídeo 4D, do IMPA, e o Kinect (TM), da Microsoft. Esses equipamentos fornecem imagens que vêm sendo chamadas de RGBD, fazendo referência aos três canais de cores e ao canal adicional de profundidade (com a letra \'D\' vindo do termo depth, profundidade em inglês). A pesquisa descrita nesta tese apresenta uma nova abordagem não-supervisionada para a estimação de movimento a partir de vídeos compostos por imagens RGBD. Esse é um passo intermediário necessário para a identificação de componentes rígidos de um objeto articulado. Nosso método faz uso da técnica de casamento inexato (homomorfismo) entre grafos para encontrar grupos de pixels (blocos) que se movem para um mesmo sentido em quadros consecutivos de um vídeo. Com o intuito de escolher o melhor casamento para cada bloco, é minimizada uma função custo que leva em conta distâncias tanto no espaço de cores RGB quanto no XYZ (espaço tridimensional do mundo). A contribuição metodológica consiste justamente na manipulação dos dados de profundidade fornecidos pelos novos dispositivos de captura, de modo que tais dados passem a integrar o vetor de características que representa cada bloco nos grafos a serem casados. Nosso método não usa quadros de referência para inicialização e é aplicável a qualquer vídeo que contenha movimento paramétrico por partes. Para blocos cujas dimensões causem uma relativa diminuição na resolução das imagens, nossa aplicação roda em tempo real. Para validar a metodologia proposta, são apresentados resultados envolvendo diversas classes de objetos com diferentes tipos de movimento, tais como vídeos de pessoas caminhando, os movimento de um braço e um casal de dançarinos de samba de gafieira. Também são apresentados os avanços obtidos na modelagem de um sistema de vídeo 4D orientado a objetos, o qual norteia o desenvolvimento de diversas aplicações a serem desenvolvidas na continuação deste trabalho. / Depth-sensing devices have arised recently, allowing real-time scene texture and depth capture. As a result, many computer vision techniques, primarily applied only to textures, now can be reformulated using additional properties like the geometry. At the same time that these algorithms, making use of this new technology, can be accelerated or be made more robust, new interesting challenges and problems to be confronted are appearing. Examples of such devices include the 4D Video Project, from IMPA, and Kinect (TM) from Microsoft. These devices offer the so called RGBD images, being related to the three color channels and to the additional depth channel. The research described on this thesis presents a new non-supervised approach to estimate motion from videos composed by RGBD images. This is an intermediary and necessary step to identify the rigid components of an articulated object. Our method uses the technique of inexact graph matching (homomorphism) to find groups of pixels (patches) that move to the same direction in subsequent video frames. In order to choose the best matching for each patch, we minimize a cost function that accounts for distances on RGB color and XYZ (tridimensional world coordinates) spaces. The methodological contribution consists on depth data manipulation given by the new capture devices, such that these data become components of the feature vector that represents each patch on graphs to be matched. Our method does not use reference frames in order to be initialized and it can be applied to any video that contains piecewise parametric motion. For patches which allow a relative decrease on images resolution, our application runs in real-time. In order to validate the proposed methodology, we present results involving object classes with different movement kinds, such as videos with walking people, the motions of an arm and a couple of samba dancers. We also present the advances obtained on modeling an object oriented 4D video system, which guide a development of different applications to be developed as future work.

Pour assurer la navigation autonome d’un robot mobile, les traitements effectués pour sa localisation doivent être faits en ligne et doivent garantir une précision suffisante pour permettre au robot d’effectuer des tâches de haut niveau pour la navigation et l’évitement d’obstacles. Les auteurs de travaux basés sur le SLAM visuel (Simultaneous Localization And Mapping) tentent depuis quelques années de garantir le meilleur compromis rapidité/précision. La majorité des solutions SLAM visuel existantes sont basées sur une représentation éparse de l’environnement. En suivant des primitives visuelles sur plusieurs images, il est possible d’estimer la position 3D de ces primitives ainsi que les poses de la caméra. La communauté du SLAM visuel a concentré ses efforts sur l’augmentation du nombre de primitives visuelles suivies et sur l’ajustement de la carte 3D, afin d’améliorer l’estimation de la trajectoire de la caméra et les positions 3D des primitives. Cependant, la localisation par SLAM visuel présente souvent des dérives dues au cumul d’erreurs, et dans le cas du SLAM visuel monoculaire, la position de la caméra n’est connue qu’à un facteur d’échelle près. Ce dernier peut être fixé initialement mais dérive au cours du temps. Pour faire face à ces limitations, nous avons centré nos travaux de thèse sur la problématique suivante : intégrer des informations supplémentaires dans un algorithme de SLAM visuel monoculaire afin de mieux contraindre la trajectoire de la caméra et la reconstruction 3D. Ces contraintes ne doivent pas détériorer les performances calculatoires de l’algorithme initial et leur absence ne doit pas mettre l’algorithme en échec. C’est pour cela que nous avons choisi d’intégrer l’information de profondeur fournie par un capteur 3D (e.g. Microsoft Kinect) et des informations géométriques sur la structure de la scène. La première contribution de cette thèse est de modifier l’algorithme SLAM visuel monoculaire proposé par Mouragnon et al. (2006b) pour prendre en compte la mesure de profondeur fournie par un capteur 3D, en proposant particulièrement un ajustement de faisceaux qui combine, d’une manière simple, des informations visuelles et des informations de profondeur. La deuxième contribution est de proposer une nouvelle fonction de coût du même ajustement de faisceaux qui intègre, en plus des contraintes sur les profondeurs des points, des contraintes géométriques d’appartenance aux plans de la scène. Les solutions proposées ont été validées sur des séquences de synthèse et sur des séquences réelles, représentant des environnements variés. Ces solutions ont été comparées aux récentes méthodes de l’état de l’art. Les résultats obtenus montrent que les différentes contraintes développées permettent d’améliorer significativement la précision de la localisation du SLAM. De plus les solutions proposées sont faciles à déployer et peu couteuses en temps de calcul. / To guarantee autonomous and safely navigation for a mobile robot, the processing achieved for its localization must be fast and accurate enough to enable the robot to perform high-level tasks for navigation and obstacle avoidance. The authors of Simultaneous Localization And Mapping (SLAM) based works, are trying since year, to ensure the speed/accuracy trade-off. Most existing works in the field of monocular (SLAM) has largely centered around sparse feature-based representations of the environment. By tracking salient image points across many frames of video, both the positions of the features and the motion of the camera can be inferred live. Within the visual SLAM community, there has been a focus on both increasing the number of features that can be tracked across an image and efficiently managing and adjusting this map of features in order to improve camera trajectory and feature location accuracy. However, visual SLAM suffers from some limitations. Indeed, with a single camera and without any assumptions or prior knowledge about the camera environment, rotation can be retrieved, but the translation is up to scale. Furthermore, visual monocular SLAM is an incremental process prone to small drifts in both pose measurement and scale, which when integrated over time, become increasingly significant over large distances. To cope with these limitations, we have centered our work around the following issues : integrate additional information into an existing monocular visual SLAM system, in order to constrain the camera localization and the mapping points. Provided that the high speed of the initial SLAM process is kept and the lack of these added constraints should not give rise to the failure of the process. For these last reasons, we have chosen to integrate the depth information provided by a 3D sensor (e.g. Microsoft Kinect) and geometric information about scene structure. The primary contribution of this work consists of modifying the SLAM algorithm proposed by Mouragnon et al. (2006b) to take into account the depth measurement provided by a 3D sensor. This consists of several rather straightforward changes, but also on a way to combine the depth and visual data in the bundle adjustment process. The second contribution is to propose a solution that uses, in addition to the depth and visual data, the constraints lying on points belonging to the plans of the scene. The proposed solutions have been validated on a synthetic sequences as well as on a real sequences, which depict various environments. These solutions have been compared to the state of art methods. The performances obtained with the previous solutions demonstrate that the additional constraints developed, improves significantly the accuracy and the robustness of the SLAM localization. Furthermore, these solutions are easy to roll out and not much time consuming.

Alex Krizhevsky and his colleagues changed the world of machine vision and image processing in 2012 when their deep learning model, named Alexnet, won the Im- ageNet Large Scale Visual Recognition Challenge with more than 10.8% lower error rate than their closest competitor. Ever since, deep learning approaches have been an area of extensive research for the tasks of object detection, classification, pose esti- mation, etc...This thesis presents a comprehensive analysis of different deep learning models and architectures that have delivered state of the art performances in various machine vision tasks. These models are compared to each other and their strengths and weaknesses are highlighted. We introduce a new approach for human head and shoulder detection from RGB- D data based on a combination of image processing and deep learning approaches. Candidate head-top locations(CHL) are generated from a fast and accurate image processing algorithm that operates on depth data. We propose enhancements to the CHL algorithm making it three times faster. Different deep learning models are then evaluated for the tasks of classification and detection on the candidate head-top loca- tions to regress the head bounding boxes and detect shoulder keypoints. We propose 3 different small models based on convolutional neural networks for this problem. Experimental results for different architectures of our model are highlighted. We also compare the performance of our model to mobilenet. Finally, we show the differences between using 3 types of inputs CNN models: RGB images, a 3-channel representation generated from depth data (Depth map, Multi-order depth template, and Height difference map or DMH), and a 4 channel input composed of RGB+D data.

Being able to predict and recognize human activities is an essential element for us to effectively communicate with other humans during our day to day activities. A system that is able to do this has a number of appealing applications, from assistive robotics to health care and preventative medicine. Previous work in supervised video-based human activity prediction and detection fails to capture the richness of spatiotemporal data that these activities generate. Convolutional Long short-term memory (Convolutional LSTM) networks are a useful tool in analyzing this type of data, showing good results in many other areas. This thesis’ focus is on utilizing RGB-D Data to improve human activity prediction and recognition. A modified Convolutional LSTM network is introduced to do so. Experiments are performed on the network and are compared to other models in-use as well as the current state-of-the-art system. We show that our proposed model for human activity prediction and recognition outperforms the current state-of-the-art models in the CAD-120 dataset without giving bounding frames or ground-truths about objects.

Traditional computer vision algorithms try to understand the world using visible light cameras. However, there are inherent limitations of this type of data source. First, visible light images are sensitive to illumination changes and background clutter. Second, the 3D structural information of the scene is lost when projecting the 3D world to 2D images. Recovering the 3D information from 2D images is a challenging problem. Range sensors have existed for over thirty years, which capture 3D characteristics of the scene. However, earlier range sensors were either too expensive, difficult to use in human environments, slow at acquiring data, or provided a poor estimation of distance. Recently, the easy access to the RGBD data at real-time frame rate is leading to a revolution in perception and inspired many new research using RGBD data. I propose algorithms to detect persons and understand the activities using RGBD data. I demonstrate the solutions to many computer vision problems may be improved with the added depth channel. The 3D structural information may give rise to algorithms with real-time and view-invariant properties in a faster and easier fashion. When both data sources are available, the features extracted from the depth channel may be combined with traditional features computed from RGB channels to generate more robust systems with enhanced recognition abilities, which may be able to deal with more challenging scenarios. As a starting point, the first problem is to find the persons of various poses in the scene, including moving or static persons. Localizing humans from RGB images is limited by the lighting conditions and background clutter. Depth image gives alternative ways to find the humans in the scene. In the past, detection of humans from range data is usually achieved by tracking, which does not work for indoor person detection. In this thesis, I propose a model based approach to detect the persons using the structural information embedded in the depth image. I propose a 2D head contour model and a 3D head surface model to look for the head-shoulder part of the person. Then, a segmentation scheme is proposed to segment the full human body from the background and extract the contour. I also give a tracking algorithm based on the detection result. I further research on recognizing human actions and activities. I propose two features for recognizing human activities. The first feature is drawn from the skeletal joint locations estimated from a depth image. It is a compact representation of the human posture called histograms of 3D joint locations (HOJ3D). This representation is view-invariant and the whole algorithm runs at real-time. This feature may benefit many applications to get a fast estimation of the posture and action of the human subject. The second feature is a spatio-temporal feature for depth video, which is called Depth Cuboid Similarity Feature (DCSF). The interest points are extracted using an algorithm that effectively suppresses the noise and finds salient human motions. DCSF is extracted centered on each interest point, which forms the description of the video contents. This descriptor can be used to recognize the activities with no dependence on skeleton information or pre-processing steps such as motion segmentation, tracking, or even image de-noising or hole-filling. It is more flexible and widely applicable to many scenarios. Finally, all the features herein developed are combined to solve a novel problem: first-person human activity recognition using RGBD data. Traditional activity recognition algorithms focus on recognizing activities from a third-person perspective. I propose to recognize activities from a first-person perspective with RGBD data. This task is very novel and extremely challenging due to the large amount of camera motion either due to self exploration or the response of the interaction. I extracted 3D optical flow features as the motion descriptor, 3D skeletal joints features as posture descriptors, spatio-temporal features as local appearance descriptors to describe the first-person videos. To address the ego-motion of the camera, I propose an attention mask to guide the recognition procedures and separate the features on the ego-motion region and independent-motion region. The 3D features are very useful at summarizing the discerning information of the activities. In addition, the combination of the 3D features with existing 2D features brings more robust recognition results and make the algorithm capable of dealing with more challenging cases. / text

Visual object tracking is a classical and very popular problem in computer vision with a plethora of applications such as vehicle navigation, human computer interface, human motion analysis, surveillance, auto-control systems and many more. Given the initial state of a target in the first frame, the goal of tracking is to predict states of the target over time where the states describe a bounding box covering the target. Despite numerous object tracking methods that have been proposed in recent years [1-4], most of these trackers suffer a degradation in performance mainly because of several challenges that include illumination changes, motion blur, complex motion, out of plane rotation, and partial or full occlusion, while occlusion is usually the most contributing factor in degrading the majority of trackers, if not all of them. This thesis is devoted to the advancement of generic object trackers tackling different challenges through different proposed methods. The work presented propose four new state-of-the-art trackers. One of which is 3D based tracker in a particle filter framework where both synchronization and registration of RGB and depth streams are adjusted automatically, and three works in correlation filters that achieve state-of-the-art performance in terms of accuracy while maintaining reasonable speeds.

Spatial Augment Reality (SAR) has presented itself to be an interesting tool for not only interesting ways to visualize information but to develop creative works in performance arts. The main challenge is to determine accurate geometry of a projection space and determine an efficient and effective way to project digital media and information to create an augmented space. In our previous implementation of SAR, we developed a projector-camera calibration approach using infrared markers. However, the projection suffers severe distortion due to the lack of depth information in the projection space. For this research, we propose to develop a RGBD sensor - projector system to replace our current projector-camera SAR system. Proper calibration between the camera or sensor and projector links vision to projection, answering the question of which point in camera space maps to what point in the space of projection. Calibration will resolve the problem of capturing the geometry of the space and allow us to accurately augment the surfaces of volumetric objects and features. In this work three calibration methods are examined for performance and accuracy. Two of these methods are existing adaptations of 2D camera - projector calibrations (calibration using arbitrary planes and ray-plane intersection) with our third proposed novel technique which utilizes point cloud information from the RGBD sensor directly. Through analysis and evaluation using re-projection error, results are presented, identifying the proposed method as practical and robust. / Electrical and Computer Engineering

本論文基於具備捕捉影像深度的RGBD影音串流裝置－Kinect感測器，在簡報場域中，作為擷取簡報者肢體動作、表情、以及語言表現模式的設備。首先我們提出在特定時段內的表現模式，可以經由大眾的評估，而具有喜歡／不喜歡的性質，我們將其分別命名為Period of Like(POL)以及Period of Dislike(POD)。論文中並以三種Kinect SDK所提供的影像特徵：動畫單元、骨架關節點、以及3D臉部頂點，輔以35位評估者所提供之評估資料，以POD/POL取出的特徵模式，分析是否具有一致性，以及是否可用於未來預測。最後將研究結果開發應用於原型程式，期許這樣的預測系統，能夠為在簡報中表現不佳而困擾的人們，提點其優劣之處，以作為後續改善之依據。 / In this thesis, we capture body movements, facial expressions, and voice data of subjects in the presentation scenario using RGBD-capable Kinect sensor. The acquired videos were accessed by a group of reviewers to indicate their preferences/aversions to the presentation style. We denote the two classes of ruling as Period of Like (POL) and Period of Dislike (POD), respectively. We then employ three types of image features, namely, animation units (AU), skeletal joints, and 3D face vertices to analyze the consistency of the evaluation result, as well as the ability to classify unseen footage based on the training data supplied by 35 evaluators. Finally, we develop a prototype program to help users to identify their strength/weakness during their presentation so that they can improve their skills accordingly.

Cette thèse a pour objectif la conception de méthodes pour la reconnaissance automatique des activités chirurgicales. Cette reconnaissance est un élément clé pour le développement de systèmes réactifs au contexte clinique et pour des applications comme l’assistance automatique lors de chirurgies complexes. Nous abordons ce problème en utilisant des méthodes de Vision puisque l’utilisation de caméras permet de percevoir l’environnement sans perturber la chirurgie. Deux types de vidéos sont utilisées : des vidéos laparoscopiques et des vidéos multi-vues RGBD. Nous avons d’abord étudié les résultats obtenus avec les méthodes de l’état de l’art, puis nous avons proposé des nouvelles approches basées sur le « Deep learning ». Nous avons aussi généré de larges jeux de données constitués d’enregistrements de chirurgies. Les résultats montrent que nos méthodes permettent d’obtenir des meilleures performances pour la reconnaissance automatique d’activités chirurgicales que l’état de l’art. / The main objective of this thesis is to address the problem of activity recognition in the operating room (OR). Activity recognition is an essential component in the development of context-aware systems, which will allow various applications, such as automated assistance during difficult procedures. Here, we focus on vision-based approaches since cameras are a common source of information to observe the OR without disrupting the surgical workflow. Specifically, we propose to use two complementary video types: laparoscopic and OR-scene RGBD videos. We investigate how state-of-the-art computer vision approaches perform on these videos and propose novel approaches, consisting of deep learning approaches, to carry out the tasks. To evaluate our proposed approaches, we generate large datasets of recordings of real surgeries. The results demonstrate that the proposed approaches outperform the state-of-the-art methods in performing surgical activity recognition on these new datasets.

Indiana University-Purdue University Indianapolis (IUPUI) / An RGBZ synthetic dataset consisting of five object classes in a variety of virtual environments and orientations was combined with a small sample of real-world image data and used to train the Mask R-CNN (MR-CNN) architecture in a variety of configurations. When the MR-CNN architecture was initialized with MS COCO weights and the heads were trained with a mix of synthetic data and real world data, F1 scores improved in four of the five classes: The average maximum F1-score of all classes and all epochs for the networks trained with synthetic data is F1∗ = 0.91, compared to F1 = 0.89 for the networks trained exclusively with real data, and the standard deviation of the maximum mean F1-score for synthetically trained networks is σ∗ = 0.015, compared to σ_F1 = 0.020 for the networks trained exclusively with real F1 data. Various backgrounds in synthetic data were shown to have negligible impact on F1 scores, opening the door to abstract backgrounds and minimizing the need for intensive synthetic data fabrication. When the MR-CNN architecture was initialized with MS COCO weights and depth data was included in the training data, the net- work was shown to rely heavily on the initial convolutional input to feed features into the network, the image depth channel was shown to influence mask generation, and the image color channels were shown to influence object classification. A set of latent variables for a subset of the synthetic datatset was generated with a Variational Autoencoder then analyzed using Principle Component Analysis and Uniform Manifold Projection and Approximation (UMAP). The UMAP analysis showed no meaningful distinction between real-world and synthetic data, and a small bias towards clustering based on image background.