BieColl - Bielefeld eCollections
Not a member yet
1004 research outputs found
Sort by
Towards a Human-like Vision System for Resource-Constrained Intelligent Cars
Get PDFResearch on computer vision systems for driver assistance resulted in a variety of approaches mainly performing reactive tasks like, e.g., lane keeping. However, for a full understanding of generic traffic situations, integrated and more flexible approaches are needed. We present a system inspired by the human visual system. Based on combining task-dependent tunable visual saliency, an object recognizer, and a tracker it provides warnings in dangerous situations
Self-Organisation of Neural Topologies by Evolutionary Reinforcement Learning
Get PDFIn this article we present EANT, "Evolutionary Acquisition of Neural Topologies", a method that creates neural networks (NNs) by evolutionary reinforcement learning. The structure of NNs is developed using mutation operators, starting from a minimal structure. Their parameters are optimised using CMA-ES. EANT can create NNs that are very specialised; they achieve a very good performance while being relatively small. This can be seen in experiments where our method competes with a different one, called NEAT, "NeuroEvolution of Augmenting Topologies", to create networks that control a robot in a visual serving scenario
SOM-based experience representation for Dextrous Grasping
Get PDFWe present an approach to dextrous robot grasping which combines a purely tactile-driven algorithm with an implicit representation of grasp experience to yield an algorithm which can handle arbitrary, partially unknown grasp situations. During the grasp movement, the obtained contact information is used to dynamically adapt the grasping control by targeting the best matching posture from the experience base. Thus, the robot recalls and actuates a grasp it already successfully performed in a similar tactile context. To efficiently represent the experience, we introduce the Grasp Manifold assuming that grasp postures form a smooth manifold in hand posture space. We present a simple way of providing approximations of Grasp Manifolds using Self-Organising Maps (SOMs) and study the properties of the represented grasp manifolds concerning their smoothness and robustness against clustered training data
Supervised Pixel-Based Texture Classification with Gabor Wavelet Filters
No full textThis paper proposes an efficient technique for pixel-based texture classification based on multichannel Gabor wavelet filters. The proposed technique is general enough to be applicable to other texture feature extraction methods that also characterize the texture around image pixels through feature vectors. During the training stage, a clustering technique is applied in order to compute a suitable set of prototypes that model every given texture pattern. Multisize evaluation windows are also utilized for improving the accuracy of the classifier near boundaries between regions of different texture. Experimental results with Brodatz compositions show the benefits of the proposed scheme in contrast with alternative approaches in terms of efficiency, memory and classification rates
Self-organizing homotopy network
Get PDFIn this paper, we propose a conceptual learning algorithm called the 'self-organizing homotopy (SOH)' together with an implementation thereof. As in the case of the SOM, our SOH organizes a homotopy in a self-organizing manner by giving a set of data episodes. Thus it is an extension of the SOM, moving from a 'map' to a 'homotopy'. From a geometrical viewpoint, the SOH represents a set of (i.e. multiple) data distributions by a fiber bundle, whereas the SOM represents a single data distribution by a manifold. One of the solutions to the SOH is SOM², in which every reference vector unit of the conventional SOM is itself replaced by an SOM. Consequently SOM² has the ability to represent a fiber bundle, i.e. a product manifold, by using a product space of SOM x SOM. It is expected that SOHs will play important roles in the fields of pattern recognition, adaptive functions, context understanding, and others, in which nonlinear manifolds and the homotopy play crucial roles
A Comprehensive System for 3D Modeling from Range Images Acquired from a 3D ToF Sensor
Get PDFDeveloping a system which generates a 3D representation of a whole scene is a difficult task. Several new technologies of 3D time-of-flight (ToF) imaging have been developed in recent years, which overcome various limitations of other 3D imaging systems, such as laser/ra\-dar/sonar scanners, structured light and stereo rigs. However, only limited work got published upon computer vision applications based on such ToF sensors. We present in this paper a new complete system for 3D modeling from a sequence of range images acquired during an arbitrary flight of a 3D ToF sensor. First, comprehensive preprocessing steps are performed to improve the quality of range images. An initial estimate of the transformation between two 3D point clouds, which are computed from two consecutive range images respectively, is achieved through feature extraction and tracking based on three kinds of images delivered by the 3D sensor. During the initial estimation, a RANSAC sampling algorithm is implemented to filter out outlier correspondences. At last the transformation is further optimized through registering the two 3D point clouds using a robust variation of the Iterative Closest Point (ICP) algorithm, the so-called Picky ICP. Extensive experimental results are provided in the paper and show the efficiency and robustness of the proposed system
Gaze shift reflex in a humanoid active vision system
Get PDFFull awareness of sensory surroundings requires active attentional and behavioural exploration. In visual animals, visual, auditory and tactile stimuli elicit gaze shifts (head and eye movements) aimed at optimising visual perception of stimuli. Such gaze shifts can either be top-down attention driven (e.g. visual search) or they can be reflex movements triggered by unexpected changes in the surroundings. Here we present a model active vision system with focus on multi-sensory integration and the generation of desired gaze shift commands. Our model is based on recent data from studies of primate superior colliculus and is developed as part of the sensory-motor control of the humanoid robot CB
Integrating Behavior-based Prediciton for Tracking Vehicles in Traffic Videos
Get PDFRoad vehicles usually remain within marked lanes. Such an hypothesis reflects a longer temporal perspective than the frequently used assumption that a vehicle continues with the currently estimated speed and direction. We study the first, more general, hypothesis in particular to track road vehicles through extended periods of occlusion "without", however, relying on 3D-models of occluding foreground bodies. A potential onset of occlusion is detected by a fuzzy conjunction of large, "facet-specific" color changes and a low ratio of the number of pixels with a prediction-compatible Optical-Flow (OF) vector relative to the total number of pixels within a facet of the 3D-polyhedral vehicle model. Experimental results for the entire approach are presented
An Adaptive Multidimensional Scaling and Principled Nonlinear Manifold
Get PDFThe self-organizing map (SOM) and some of its variants such as visualization induced SOM (ViSOM) have been shown to yield similar results to multidimensional scaling (MDS). However the exact connection has yet been established. In this paper we first examine their relationship with (generalized) MDS from their cost functions in the aspect of data visualization and dimensionality reduction. The SOM is shown to produce a quantized, qualitative or nonmetric scaling and while the ViSOM is a quantitative metric scaling. Then we propose a way to use the core principle of the ViSOM, i.e. local distance preserving, to adaptively and incrementally construct a metric local scaling and to extract nonlinear manifold. Comparison with other methods such as ISOMAP and LLE has been made, especially in mapping highly nonlinear subspaces. The advantages over other methods are also discussed
An Adaptive Vision System for Tracking Soccer Players from Various Camera Settings
Get PDFIn this paper we present Aspogamo, a vision system capable of estimating motion trajectories of soccer players taped on video. The system performs well in a multitude of application scenarios because of its adaptivity to various camera setups, such as single or multiple camera settings, static or dynamic ones. Furthermore, Aspogamo can directly process image streams taken from TV broadcast, and extract all valuable information despite scene interruptions and cuts between different cameras. The system achieves a high level of robustness through the use of modelbased vision algorithms for camera estimation and player recognition and a probabilistic multi-player tracking framework capable of dealing with occlusion situations typical in team-sports. The continuous interplay between these submodules is adding to both the reliability and the efficiency of the overall system