Real time motion estimation using a neural architecture implemented on GPUs

This work describes a neural network based architecture that represents and estimates object motion in videos. This architecture addresses multiple computer vision tasks such as image segmentation, object representation or characterization, motion analysis and tracking. The use of a neural network architecture allows for the simultaneous estimation of global and local motion and the representation of deformable objects. This architecture also avoids the problem of finding corresponding features while tracking moving objects. Due to the parallel nature of neural networks, the architecture has been implemented on GPUs that allows the system to meet a set of requirements such as: time constraints management, robustness, high processing speed and re-configurability. Experiments are presented that demonstrate the validity of our architecture to solve problems of mobile agents tracking and motion analysis.

Estudio y mejora de métodos de registro 3D: aceleración sobre unidades de procesamiento gráfico y caracterización del espacio de transformaciones iniciales

Durante los últimos años ha sido creciente el uso de las unidades de procesamiento gráfico, más conocidas como GPU (Graphic Processing Unit), en aplicaciones de propósito general, dejando a un lado el objetivo para el que fueron creadas y que no era otro que el renderizado de gráficos por computador. Este crecimiento se debe en parte a la evolución que han experimentado estos dispositivos durante este tiempo y que les ha dotado de gran potencia de cálculo, consiguiendo que su uso se extienda desde ordenadores personales a grandes cluster. Este hecho unido a la proliferación de sensores RGB-D de bajo coste ha hecho que crezca el número de aplicaciones de visión que hacen uso de esta tecnología para la resolución de problemas, así como también para el desarrollo de nuevas aplicaciones. Todas estas mejoras no solamente se han realizado en la parte hardware, es decir en los dispositivos, sino también en la parte software con la aparición de nuevas herramientas de desarrollo que facilitan la programación de estos dispositivos GPU. Este nuevo paradigma se acuñó como Computación de Propósito General sobre Unidades de Proceso Gráfico (General-Purpose computation on Graphics Processing Units, GPGPU). Los dispositivos GPU se clasifican en diferentes familias, en función de las distintas características hardware que poseen. Cada nueva familia que aparece incorpora nuevas mejoras tecnológicas que le permite conseguir mejor rendimiento que las anteriores. No obstante, para sacar un rendimiento óptimo a un dispositivo GPU es necesario configurarlo correctamente antes de usarlo. Esta configuración viene determinada por los valores asignados a una serie de parámetros del dispositivo. Por tanto, muchas de las implementaciones que hoy en día hacen uso de los dispositivos GPU para el registro denso de nubes de puntos 3D, podrían ver mejorado su rendimiento con una configuración óptima de dichos parámetros, en función del dispositivo utilizado. Es por ello que, ante la falta de un estudio detallado del grado de afectación de los parámetros GPU sobre el rendimiento final de una implementación, se consideró muy conveniente la realización de este estudio. Este estudio no sólo se realizó con distintas configuraciones de parámetros GPU, sino también con diferentes arquitecturas de dispositivos GPU. El objetivo de este estudio es proporcionar una herramienta de decisión que ayude a los desarrolladores a la hora implementar aplicaciones para dispositivos GPU. Uno de los campos de investigación en los que más prolifera el uso de estas tecnologías es el campo de la robótica ya que tradicionalmente en robótica, sobre todo en la robótica móvil, se utilizaban combinaciones de sensores de distinta naturaleza con un alto coste económico, como el láser, el sónar o el sensor de contacto, para obtener datos del entorno. Más tarde, estos datos eran utilizados en aplicaciones de visión por computador con un coste computacional muy alto. Todo este coste, tanto el económico de los sensores utilizados como el coste computacional, se ha visto reducido notablemente gracias a estas nuevas tecnologías. Dentro de las aplicaciones de visión por computador más utilizadas está el registro de nubes de puntos. Este proceso es, en general, la transformación de diferentes nubes de puntos a un sistema de coordenadas conocido. Los datos pueden proceder de fotografías, de diferentes sensores, etc. Se utiliza en diferentes campos como son la visión artificial, la imagen médica, el reconocimiento de objetos y el análisis de imágenes y datos de satélites. El registro se utiliza para poder comparar o integrar los datos obtenidos en diferentes mediciones. En este trabajo se realiza un repaso del estado del arte de los métodos de registro 3D. Al mismo tiempo, se presenta un profundo estudio sobre el método de registro 3D más utilizado, Iterative Closest Point (ICP), y una de sus variantes más conocidas, Expectation-Maximization ICP (EMICP). Este estudio contempla tanto su implementación secuencial como su implementación paralela en dispositivos GPU, centrándose en cómo afectan a su rendimiento las distintas configuraciones de parámetros GPU. Como consecuencia de este estudio, también se presenta una propuesta para mejorar el aprovechamiento de la memoria de los dispositivos GPU, permitiendo el trabajo con nubes de puntos más grandes, reduciendo el problema de la limitación de memoria impuesta por el dispositivo. El funcionamiento de los métodos de registro 3D utilizados en este trabajo depende en gran medida de la inicialización del problema. En este caso, esa inicialización del problema consiste en la correcta elección de la matriz de transformación con la que se iniciará el algoritmo. Debido a que este aspecto es muy importante en este tipo de algoritmos, ya que de él depende llegar antes o no a la solución o, incluso, no llegar nunca a la solución, en este trabajo se presenta un estudio sobre el espacio de transformaciones con el objetivo de caracterizarlo y facilitar la elección de la transformación inicial a utilizar en estos algoritmos.

Real-time depth camera tracking with CAD models and ICP

In recent years, depth cameras have been widely utilized in camera tracking for augmented and mixed reality. Many of the studies focus on the methods that generate the reference model simultaneously with the tracking and allow operation in unprepared environments. However, methods that rely on predefined CAD models have their advantages. In such methods, the measurement errors are not accumulated to the model, they are tolerant to inaccurate initialization, and the tracking is always performed directly in reference model's coordinate system. In this paper, we present a method for tracking a depth camera with existing CAD models and the Iterative Closest Point (ICP) algorithm. In our approach, we render the CAD model using the latest pose estimate and construct a point cloud from the corresponding depth map. We construct another point cloud from currently captured depth frame, and find the incremental change in the camera pose by aligning the point clouds. We utilize a GPGPU-based implementation of the ICP which efficiently uses all the depth data in the process. The method runs in real-time, it is robust for outliers, and it does not require any preprocessing of the CAD models. We evaluated the approach using the Kinect depth sensor, and compared the results to a 2D edge-based method, to a depth-based SLAM method, and to the ground truth. The results show that the approach is more stable compared to the edge-based method and it suffers less from drift compared to the depth-based SLAM.

A three-dimensional representation method for noisy point clouds based on growing self-organizing maps accelerated on GPUs

The research described in this thesis was motivated by the need of a robust model capable of representing 3D data obtained with 3D sensors, which are inherently noisy. In addition, time constraints have to be considered as these sensors are capable of providing a 3D data stream in real time. This thesis proposed the use of Self-Organizing Maps (SOMs) as a 3D representation model. In particular, we proposed the use of the Growing Neural Gas (GNG) network, which has been successfully used for clustering, pattern recognition and topology representation of multi-dimensional data. Until now, Self-Organizing Maps have been primarily computed offline and their application in 3D data has mainly focused on free noise models, without considering time constraints. It is proposed a hardware implementation leveraging the computing power of modern GPUs, which takes advantage of a new paradigm coined as General-Purpose Computing on Graphics Processing Units (GPGPU). The proposed methods were applied to different problem and applications in the area of computer vision such as the recognition and localization of objects, visual surveillance or 3D reconstruction.

INVESTIGATIONS INTO MECHANISMS UNDERLYING EXTREME WAVE FORMATIONS AND COMPUTATIONALLY INTENSIVE SIMULATIONS

Various mechanisms have been proposed to explain extreme waves or rogue waves in an oceanic environment including directional focusing, dispersive focusing, wave-current interaction, and nonlinear modulational instability. The Benjamin-Feir instability (nonlinear modulational instability), however, is considered to be one of the primary mechanisms for rogue-wave occurrence. The nonlinear Schrodinger equation is a well-established approximate model based on the same assumptions as required for the derivation of the Benjamin-Feir theory. Solutions of the nonlinear Schrodinger equation, including new rogue-wave type solutions are presented in the author's dissertation work. The solutions are obtained by using a predictive eigenvalue map based predictor-corrector procedure developed by the author. Features of the predictive map are explored and the influences of certain parameter variations are investigated. The solutions are rescaled to match the length scales of waves generated in a wave tank. Based on the information provided by the map and the details of physical scaling, a framework is developed that can serve as a basis for experimental investigations into a variety of extreme waves as well localizations in wave fields. To derive further fundamental insights into the complexity of extreme wave conditions, Smoothed Particle Hydrodynamics (SPH) simulations are carried out on an advanced Graphic Processing Unit (GPU) based parallel computational platform. Free surface gravity wave simulations have successfully characterized water-wave dispersion in the SPH model while demonstrating extreme energy focusing and wave growth in both linear and nonlinear regimes. A virtual wave tank is simulated wherein wave motions can be excited from either side. Focusing of several wave trains and isolated waves has been simulated. With properly chosen parameters, dispersion effects are observed causing a chirped wave train to focus and exhibit growth. By using the insights derived from the study of the nonlinear Schrodinger equation, modulational instability or self-focusing has been induced in a numerical wave tank and studied through several numerical simulations. Due to the inherent dissipative nature of SPH models, simulating persistent progressive waves can be problematic. This issue has been addressed and an observation-based solution has been provided. The efficacy of SPH in modeling wave focusing can be critical to further our understanding and predicting extreme wave phenomena through simulations. A deeper understanding of the mechanisms underlying extreme energy localization phenomena can help facilitate energy harnessing and serve as a basis to predict and mitigate the impact of energy focusing.

Contributions to 3D Data Registration and Representation

Nowadays, new computers generation provides a high performance that enables to build computationally expensive computer vision applications applied to mobile robotics. Building a map of the environment is a common task of a robot and is an essential part to allow the robots to move through these environments. Traditionally, mobile robots used a combination of several sensors from different technologies. Lasers, sonars and contact sensors have been typically used in any mobile robotic architecture, however color cameras are an important sensor due to we want the robots to use the same information that humans to sense and move through the different environments. Color cameras are cheap and flexible but a lot of work need to be done to give robots enough visual understanding of the scenes. Computer vision algorithms are computational complex problems but nowadays robots have access to different and powerful architectures that can be used for mobile robotics purposes. The advent of low-cost RGB-D sensors like Microsoft Kinect which provide 3D colored point clouds at high frame rates made the computer vision even more relevant in the mobile robotics field. The combination of visual and 3D data allows the systems to use both computer vision and 3D processing and therefore to be aware of more details of the surrounding environment. The research described in this thesis was motivated by the need of scene mapping. Being aware of the surrounding environment is a key feature in many mobile robotics applications from simple robotic navigation to complex surveillance applications. In addition, the acquisition of a 3D model of the scenes is useful in many areas as video games scene modeling where well-known places are reconstructed and added to game systems or advertising where once you get the 3D model of one room the system can add furniture pieces using augmented reality techniques. In this thesis we perform an experimental study of the state-of-the-art registration methods to find which one fits better to our scene mapping purposes. Different methods are tested and analyzed on different scene distributions of visual and geometry appearance. In addition, this thesis proposes two methods for 3d data compression and representation of 3D maps. Our 3D representation proposal is based on the use of Growing Neural Gas (GNG) method. This Self-Organizing Maps (SOMs) has been successfully used for clustering, pattern recognition and topology representation of various kind of data. Until now, Self-Organizing Maps have been primarily computed offline and their application in 3D data has mainly focused on free noise models without considering time constraints. Self-organising neural models have the ability to provide a good representation of the input space. In particular, the Growing Neural Gas (GNG) is a suitable model because of its flexibility, rapid adaptation and excellent quality of representation. However, this type of learning is time consuming, specially for high-dimensional input data. Since real applications often work under time constraints, it is necessary to adapt the learning process in order to complete it in a predefined time. This thesis proposes a hardware implementation leveraging the computing power of modern GPUs which takes advantage of a new paradigm coined as General-Purpose Computing on Graphics Processing Units (GPGPU). Our proposed geometrical 3D compression method seeks to reduce the 3D information using plane detection as basic structure to compress the data. This is due to our target environments are man-made and therefore there are a lot of points that belong to a plane surface. Our proposed method is able to get good compression results in those man-made scenarios. The detected and compressed planes can be also used in other applications as surface reconstruction or plane-based registration algorithms. Finally, we have also demonstrated the goodness of the GPU technologies getting a high performance implementation of a CAD/CAM common technique called Virtual Digitizing.

Análisis de la implementación de un conformador software de señales ultrasónicas para tiempo real

Este trabajo analiza la implementacion software en un sistema de imagen ultrasónica del Total Focusing Method para la compensación dinámica en tiempo real de los tiempos de vuelo para emisión y recepción de todos los puntos de la imagen. Para ello, haciendo uso de técnicas GPGPU, se analizan dos diferentes alternativas de implementacion, mostrando como una planificación adecuada de acceso a los datos permite mejorar los tiempos de ejecución del algoritmo.

Using multiple GPUs to accelerate string searching for digital forensic analysis

String searching within a large corpus of data is an important component of digital forensic (DF) analysis techniques such as file carving. The continuing increase in capacity of consumer storage devices requires corresponding im-provements to the performance of string searching techniques. As string search-ing is a trivially-parallelisable problem, GPGPU approaches are a natural fit – but previous studies have found that local storage presents an insurmountable performance bottleneck. We show that this need not be the case with modern hardware, and demonstrate substantial performance improvements from the use of single and multiple GPUs when searching for strings within a typical forensic disk image.