Extending visual perception with haptic exploration for improved scene understanding

Scene understanding has been investigated from a mainly visual information point of view. Recently depth has been provided an extra wealth of information, allowing more geometric knowledge to fuse into scene understanding. Yet to form a holistic view, especially in robotic applications, one can create even more data by interacting with the world. In fact humans, when growing up, seem to heavily investigate the world around them by haptic exploration. We show an application of haptic exploration on a humanoid robot in cooperation with a learning method for object segmentation. The actions performed consecutively improve the segmentation of objects in the scene.

Multi-scale 3D Scene Flow from Binocular Stereo Sequences

Scene flow methods estimate the three-dimensional motion field for points in the world, using multi-camera video data. Such methods combine multi-view reconstruction with motion estimation approaches. This paper describes an alternative formulation for dense scene flow estimation that provides convincing results using only two cameras by fusing stereo and optical flow estimation into a single coherent framework. To handle the aperture problems inherent in the estimation task, a multi-scale method along with a novel adaptive smoothing technique is used to gain a regularized solution. This combined approach both preserves discontinuities and prevents over-regularization-two problems commonly associated with basic multi-scale approaches. Internally, the framework generates probability distributions for optical flow and disparity. Taking into account the uncertainty in the intermediate stages allows for more reliable estimation of the 3D scene flow than standard stereo and optical flow methods allow. Experiments with synthetic and real test data demonstrate the effectiveness of the approach.

Multi-Scale 3D Scene Flow from Binocular Stereo Sequences

Scene flow methods estimate the three-dimensional motion field for points in the world, using multi-camera video data. Such methods combine multi-view reconstruction with motion estimation. This paper describes an alternative formulation for dense scene flow estimation that provides reliable results using only two cameras by fusing stereo and optical flow estimation into a single coherent framework. Internally, the proposed algorithm generates probability distributions for optical flow and disparity. Taking into account the uncertainty in the intermediate stages allows for more reliable estimation of the 3D scene flow than previous methods allow. To handle the aperture problems inherent in the estimation of optical flow and disparity, a multi-scale method along with a novel region-based technique is used within a regularized solution. This combined approach both preserves discontinuities and prevents over-regularization – two problems commonly associated with the basic multi-scale approaches. Experiments with synthetic and real test data demonstrate the strength of the proposed approach.

Efficient 3D scene modeling and mosaicing

El modelat d'escenes és clau en un gran ventall d'aplicacions que van des de la generació mapes fins a la realitat augmentada. Aquesta tesis presenta una solució completa per a la creació de models 3D amb textura. En primer lloc es presenta un mètode de Structure from Motion seqüencial, a on el model 3D de l'entorn s'actualitza a mesura que s'adquireix nova informació visual. La proposta és més precisa i robusta que l'estat de l'art. També s'ha desenvolupat un mètode online, basat en visual bag-of-words, per a la detecció eficient de llaços. Essent una tècnica completament seqüencial i automàtica, permet la reducció de deriva, millorant la navegació i construcció de mapes. Per tal de construir mapes en àrees extenses, es proposa un algorisme de simplificació de models 3D, orientat a aplicacions online. L'eficiència de les propostes s'ha comparat amb altres mètodes utilitzant diversos conjunts de dades submarines i terrestres.

Scene understanding for auto-calibration of surveillance cameras

In the last decade, several research results have presented formulations for the auto-calibration problem. Most of these have relied on the evaluation of vanishing points to extract the camera parameters. Normally vanishing points are evaluated using pedestrians or the Manhattan World assumption i.e. it is assumed that the scene is necessarily composed of orthogonal planar surfaces. In this work, we present a robust framework for auto-calibration, with improved results and generalisability for real-life situations. This framework is capable of handling problems such as occlusions and the presence of unexpected objects in the scene. In our tests, we compare our formulation with the state-of-the-art in auto-calibration using pedestrians and Manhattan World-based assumptions. This paper reports on the experiments conducted using publicly available datasets; the results have shown that our formulation represents an improvement over the state-of-the-art.

GPGPU implementation of growing neural gas: application to 3D scene reconstruction

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, especially 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 paper proposes a Graphics Processing Unit (GPU) parallel implementation of the GNG with Compute Unified Device Architecture (CUDA). In contrast to existing algorithms, the proposed GPU implementation allows the acceleration of the learning process keeping a good quality of representation. Comparative experiments using iterative, parallel and hybrid implementations are carried out to demonstrate the effectiveness of CUDA implementation. The results show that GNG learning with the proposed implementation achieves a speed-up of 6× compared with the single-threaded CPU implementation. GPU implementation has also been applied to a real application with time constraints: acceleration of 3D scene reconstruction for egomotion, in order to validate the proposal.

Semi-binary based video features for activity representation

Efficient and effective feature detection and representation is an important consideration when processing videos, and a large number of applications such as motion analysis, 3D scene understanding, tracking etc. depend on this. Amongst several feature description methods, local features are becoming increasingly popular for representing videos because of their simplicity and efficiency. While they achieve state-of-the-art performance with low computational complexity, their performance is still too limited for real world applications. Furthermore, rapid increases in the uptake of mobile devices has increased the demand for algorithms that can run with reduced memory and computational requirements. In this paper we propose a semi binary based feature detectordescriptor based on the BRISK detector, which can detect and represent videos with significantly reduced computational requirements, while achieving comparable performance to the state of the art spatio-temporal feature descriptors. First, the BRISK feature detector is applied on a frame by frame basis to detect interest points, then the detected key points are compared against consecutive frames for significant motion. Key points with significant motion are encoded with the BRISK descriptor in the spatial domain and Motion Boundary Histogram in the temporal domain. This descriptor is not only lightweight but also has lower memory requirements because of the binary nature of the BRISK descriptor, allowing the possibility of applications using hand held devices.We evaluate the combination of detectordescriptor performance in the context of action classification with a standard, popular bag-of-features with SVM framework. Experiments are carried out on two popular datasets with varying complexity and we demonstrate comparable performance with other descriptors with reduced computational complexity.

Multiple instance dictionary learning for activity representation

This paper presents an effective feature representation method in the context of activity recognition. Efficient and effective feature representation plays a crucial role not only in activity recognition, but also in a wide range of applications such as motion analysis, tracking, 3D scene understanding etc. In the context of activity recognition, local features are increasingly popular for representing videos because of their simplicity and efficiency. While they achieve state-of-the-art performance with low computational requirements, their performance is still limited for real world applications due to a lack of contextual information and models not being tailored to specific activities. We propose a new activity representation framework to address the shortcomings of the popular, but simple bag-of-words approach. In our framework, first multiple instance SVM (mi-SVM) is used to identify positive features for each action category and the k-means algorithm is used to generate a codebook. Then locality-constrained linear coding is used to encode the features into the generated codebook, followed by spatio-temporal pyramid pooling to convey the spatio-temporal statistics. Finally, an SVM is used to classify the videos. Experiments carried out on two popular datasets with varying complexity demonstrate significant performance improvement over the base-line bag-of-feature method.

Bayesian scene analysis for multi-camera 3D tracking and camera positioning

Los sistemas de seguimiento mono-cámara han demostrado su notable capacidad para el análisis de trajectorias de objectos móviles y para monitorización de escenas de interés; sin embargo, tanto su robustez como sus posibilidades en cuanto a comprensión semántica de la escena están fuertemente limitadas por su naturaleza local y monocular, lo que los hace insuficientes para aplicaciones realistas de videovigilancia. El objetivo de esta tesis es la extensión de las posibilidades de los sistemas de seguimiento de objetos móviles para lograr un mayor grado de robustez y comprensión de la escena. La extensión propuesta se divide en dos direcciones separadas. La primera puede considerarse local, ya que está orientada a la mejora y enriquecimiento de las posiciones estimadas para los objetos móviles observados directamente por las cámaras del sistema; dicha extensión se logra mediante el desarrollo de un sistema multi-cámara de seguimiento 3D, capaz de proporcionar consistentemente las posiciones 3D de múltiples objetos a partir de las observaciones capturadas por un conjunto de sensores calibrados y con campos de visión solapados. La segunda extensión puede considerarse global, dado que su objetivo consiste en proporcionar un contexto global para relacionar las observaciones locales realizadas por una cámara con una escena de mucho mayor tamaño; para ello se propone un sistema automático de localización de cámaras basado en las trayectorias observadas de varios objetos móviles y en un mapa esquemático de la escena global monitorizada. Ambas líneas de investigación se tratan utilizando, como marco común, técnicas de estimación bayesiana: esta elección está justificada por la versatilidad y flexibilidad proporcionada por dicho marco estadístico, que permite la combinación natural de múltiples fuentes de información sobre los parámetros a estimar, así como un tratamiento riguroso de la incertidumbre asociada a las mismas mediante la inclusión de modelos de observación específicamente diseñados. Además, el marco seleccionado abre grandes posibilidades operacionales, puesto que permite la creación de diferentes métodos numéricos adaptados a las necesidades y características específicas de distintos problemas tratados. El sistema de seguimiento 3D con múltiples cámaras propuesto está específicamente diseñado para permitir descripciones esquemáticas de las medidas realizadas individualmente por cada una de las cámaras del sistema: esta elección de diseño, por tanto, no asume ningún algoritmo específico de detección o seguimiento 2D en ninguno de los sensores de la red, y hace que el sistema propuesto sea aplicable a redes reales de vigilancia con capacidades limitadas tanto en términos de procesamiento como de transmision. La combinación robusta de las observaciones capturadas individualmente por las cámaras, ruidosas, incompletas y probablemente contaminadas por falsas detecciones, se basa en un metodo de asociación bayesiana basado en geometría y color: los resultados de dicha asociación permiten el seguimiento 3D de los objetos de la escena mediante el uso de un filtro de partículas. El sistema de fusión de observaciones propuesto tiene, como principales características, una gran precisión en términos de localización 3D de objetos, y una destacable capacidad de recuperación tras eventuales errores debidos a un número insuficiente de datos de entrada. El sistema automático de localización de cámaras se basa en la observación de múltiples objetos móviles y un mapa esquemático de las áreas transitables del entorno monitorizado para inferir la posición absoluta de dicho sensor. Para este propósito, se propone un novedoso marco bayesiano que combina modelos dinámicos inducidos por el mapa en los objetos móviles presentes en la escena con las trayectorias observadas por la cámara, lo que representa un enfoque nunca utilizado en la literatura existente. El sistema de localización se divide en dos sub-tareas diferenciadas, debido a que cada una de estas tareas requiere del diseño de algoritmos específicos de muestreo para explotar en profundidad las características del marco desarrollado: por un lado, análisis de la ambigüedad del caso específicamente tratado y estimación aproximada de la localización de la cámara, y por otro, refinado de la localización de la cámara. El sistema completo, diseñado y probado para el caso específico de localización de cámaras en entornos de tráfico urbano, podría tener aplicación también en otros entornos y sensores de diferentes modalidades tras ciertas adaptaciones. ABSTRACT Mono-camera tracking systems have proved their capabilities for moving object trajectory analysis and scene monitoring, but their robustness and semantic possibilities are strongly limited by their local and monocular nature and are often insufficient for realistic surveillance applications. This thesis is aimed at extending the possibilities of moving object tracking systems to a higher level of scene understanding. The proposed extension comprises two separate directions. The first one is local, since is aimed at enriching the inferred positions of the moving objects within the area of the monitored scene directly covered by the cameras of the system; this task is achieved through the development of a multi-camera system for robust 3D tracking, able to provide 3D tracking information of multiple simultaneous moving objects from the observations reported by a set of calibrated cameras with semi-overlapping fields of view. The second extension is global, as is aimed at providing local observations performed within the field of view of one camera with a global context relating them to a much larger scene; to this end, an automatic camera positioning system relying only on observed object trajectories and a scene map is designed. The two lines of research in this thesis are addressed using Bayesian estimation as a general unifying framework. Its suitability for these two applications is justified by the flexibility and versatility of that stochastic framework, which allows the combination of multiple sources of information about the parameters to estimate in a natural and elegant way, addressing at the same time the uncertainty associated to those sources through the inclusion of models designed to this end. In addition, it opens multiple possibilities for the creation of different numerical methods for achieving satisfactory and efficient practical solutions to each addressed application. The proposed multi-camera 3D tracking method is specifically designed to work on schematic descriptions of the observations performed by each camera of the system: this choice allows the use of unspecific off-the-shelf 2D detection and/or tracking subsystems running independently at each sensor, and makes the proposal suitable for real surveillance networks with moderate computational and transmission capabilities. The robust combination of such noisy, incomplete and possibly unreliable schematic descriptors relies on a Bayesian association method, based on geometry and color, whose results allow the tracking of the targets in the scene with a particle filter. The main features exhibited by the proposal are, first, a remarkable accuracy in terms of target 3D positioning, and second, a great recovery ability after tracking losses due to insufficient input data. The proposed system for visual-based camera self-positioning uses the observations of moving objects and a schematic map of the passable areas of the environment to infer the absolute sensor position. To this end, a new Bayesian framework combining trajectory observations and map-induced dynamic models for moving objects is designed, which represents an approach to camera positioning never addressed before in the literature. This task is divided into two different sub-tasks, setting ambiguity analysis and approximate position estimation, on the one hand, and position refining, on the other, since they require the design of specific sampling algorithms to correctly exploit the discriminative features of the developed framework. This system, designed for camera positioning and demonstrated in urban traffic environments, can also be applied to different environments and sensors of other modalities after certain required adaptations.

Bayesian scene analysis for multi-camera 3D tracking and camera positioning

Los sistemas de seguimiento mono-cámara han demostrado su notable capacidad para el análisis de trajectorias de objectos móviles y para monitorización de escenas de interés; sin embargo, tanto su robustez como sus posibilidades en cuanto a comprensión semántica de la escena están fuertemente limitadas por su naturaleza local y monocular, lo que los hace insuficientes para aplicaciones realistas de videovigilancia. El objetivo de esta tesis es la extensión de las posibilidades de los sistemas de seguimiento de objetos móviles para lograr un mayor grado de robustez y comprensión de la escena. La extensión propuesta se divide en dos direcciones separadas. La primera puede considerarse local, ya que está orientada a la mejora y enriquecimiento de las posiciones estimadas para los objetos móviles observados directamente por las cámaras del sistema; dicha extensión se logra mediante el desarrollo de un sistema multi-cámara de seguimiento 3D, capaz de proporcionar consistentemente las posiciones 3D de múltiples objetos a partir de las observaciones capturadas por un conjunto de sensores calibrados y con campos de visión solapados. La segunda extensión puede considerarse global, dado que su objetivo consiste en proporcionar un contexto global para relacionar las observaciones locales realizadas por una cámara con una escena de mucho mayor tamaño; para ello se propone un sistema automático de localización de cámaras basado en las trayectorias observadas de varios objetos móviles y en un mapa esquemático de la escena global monitorizada. Ambas líneas de investigación se tratan utilizando, como marco común, técnicas de estimación bayesiana: esta elección está justificada por la versatilidad y flexibilidad proporcionada por dicho marco estadístico, que permite la combinación natural de múltiples fuentes de información sobre los parámetros a estimar, así como un tratamiento riguroso de la incertidumbre asociada a las mismas mediante la inclusión de modelos de observación específicamente diseñados. Además, el marco seleccionado abre grandes posibilidades operacionales, puesto que permite la creación de diferentes métodos numéricos adaptados a las necesidades y características específicas de distintos problemas tratados. El sistema de seguimiento 3D con múltiples cámaras propuesto está específicamente diseñado para permitir descripciones esquemáticas de las medidas realizadas individualmente por cada una de las cámaras del sistema: esta elección de diseño, por tanto, no asume ningún algoritmo específico de detección o seguimiento 2D en ninguno de los sensores de la red, y hace que el sistema propuesto sea aplicable a redes reales de vigilancia con capacidades limitadas tanto en términos de procesamiento como de transmision. La combinación robusta de las observaciones capturadas individualmente por las cámaras, ruidosas, incompletas y probablemente contaminadas por falsas detecciones, se basa en un metodo de asociación bayesiana basado en geometría y color: los resultados de dicha asociación permiten el seguimiento 3D de los objetos de la escena mediante el uso de un filtro de partículas. El sistema de fusión de observaciones propuesto tiene, como principales características, una gran precisión en términos de localización 3D de objetos, y una destacable capacidad de recuperación tras eventuales errores debidos a un número insuficiente de datos de entrada. El sistema automático de localización de cámaras se basa en la observación de múltiples objetos móviles y un mapa esquemático de las áreas transitables del entorno monitorizado para inferir la posición absoluta de dicho sensor. Para este propósito, se propone un novedoso marco bayesiano que combina modelos dinámicos inducidos por el mapa en los objetos móviles presentes en la escena con las trayectorias observadas por la cámara, lo que representa un enfoque nunca utilizado en la literatura existente. El sistema de localización se divide en dos sub-tareas diferenciadas, debido a que cada una de estas tareas requiere del diseño de algoritmos específicos de muestreo para explotar en profundidad las características del marco desarrollado: por un lado, análisis de la ambigüedad del caso específicamente tratado y estimación aproximada de la localización de la cámara, y por otro, refinado de la localización de la cámara. El sistema completo, diseñado y probado para el caso específico de localización de cámaras en entornos de tráfico urbano, podría tener aplicación también en otros entornos y sensores de diferentes modalidades tras ciertas adaptaciones. ABSTRACT Mono-camera tracking systems have proved their capabilities for moving object trajectory analysis and scene monitoring, but their robustness and semantic possibilities are strongly limited by their local and monocular nature and are often insufficient for realistic surveillance applications. This thesis is aimed at extending the possibilities of moving object tracking systems to a higher level of scene understanding. The proposed extension comprises two separate directions. The first one is local, since is aimed at enriching the inferred positions of the moving objects within the area of the monitored scene directly covered by the cameras of the system; this task is achieved through the development of a multi-camera system for robust 3D tracking, able to provide 3D tracking information of multiple simultaneous moving objects from the observations reported by a set of calibrated cameras with semi-overlapping fields of view. The second extension is global, as is aimed at providing local observations performed within the field of view of one camera with a global context relating them to a much larger scene; to this end, an automatic camera positioning system relying only on observed object trajectories and a scene map is designed. The two lines of research in this thesis are addressed using Bayesian estimation as a general unifying framework. Its suitability for these two applications is justified by the flexibility and versatility of that stochastic framework, which allows the combination of multiple sources of information about the parameters to estimate in a natural and elegant way, addressing at the same time the uncertainty associated to those sources through the inclusion of models designed to this end. In addition, it opens multiple possibilities for the creation of different numerical methods for achieving satisfactory and efficient practical solutions to each addressed application. The proposed multi-camera 3D tracking method is specifically designed to work on schematic descriptions of the observations performed by each camera of the system: this choice allows the use of unspecific off-the-shelf 2D detection and/or tracking subsystems running independently at each sensor, and makes the proposal suitable for real surveillance networks with moderate computational and transmission capabilities. The robust combination of such noisy, incomplete and possibly unreliable schematic descriptors relies on a Bayesian association method, based on geometry and color, whose results allow the tracking of the targets in the scene with a particle filter. The main features exhibited by the proposal are, first, a remarkable accuracy in terms of target 3D positioning, and second, a great recovery ability after tracking losses due to insufficient input data. The proposed system for visual-based camera self-positioning uses the observations of moving objects and a schematic map of the passable areas of the environment to infer the absolute sensor position. To this end, a new Bayesian framework combining trajectory observations and map-induced dynamic models for moving objects is designed, which represents an approach to camera positioning never addressed before in the literature. This task is divided into two different sub-tasks, setting ambiguity analysis and approximate position estimation, on the one hand, and position refining, on the other, since they require the design of specific sampling algorithms to correctly exploit the discriminative features of the developed framework. This system, designed for camera positioning and demonstrated in urban traffic environments, can also be applied to different environments and sensors of other modalities after certain required adaptations.

A unified 2D-3D video scene change detection framework for mobile camera platforms

In this paper, we present a novel scene change detection algorithm for mobile camera platforms. Our approach integrates sparse 3D scene background modelling and dense 2D image background modelling into a unified framework. The 3D scene background modelling identifies inconsistent clusters over time in a set of 3D cloud points as the scene changes. The 2D image background modelling further confirms the scene changes by finding inconsistent appearances in a set of aligned images using the classical MRF background subtraction technique. We evaluate the performance of our proposed system on a number of challenging video datasets obtained from a camera placed on a moving vehicle and the experiments show that our proposed method outperforms previous works in scene change detection, which suggested the feasibility of our approach.