SEARISE

adaptive_visual_attention_to_salient_events

2011-03-03T13:39:46+02:00

Adaptive visual attention to salient events A model component for the computation of dynamic salience maps and recognition of salient events will be developed in this WP. This component is at the highest cognitive level in the hierarchical architecture of the overall visual model. The output from this component will direct saccadic moves of the binocular cameras in order to switch the attention of the Smart-Eyes system to salient events with the highest priority. At the same time this compone…

benchmark_datasets

2011-03-10T10:03:57+02:00

Benchmark Datasets Here we provide links to benchmark datasets created in the context of the SEARISE project which are demeed useful for the wider community to evaluate performance in tasks which were dealt with in the project. * Real-world stereo sequences with ground truth for benchmarking full (i.e., 2D) disparity estimation in active binocular systems * Psychophysical motion benchmark dataset to evaluate models of visual cortical functions

camera_hardware_development_and_control

2010-06-07T15:24:04+02:00

A trinocular robotic head (SmartEyes) with 4 degrees of freedom has been designed (see Fig. left), by taking into account the specifications of the two application scenarios: the long-range one, where the SmartEyes will be fixed on the arena ceiling overlooking opposite spectacular ranks, and the short-range one, where the robotic head will be fixed above railway platform overlooking the train carriages. The SmartEyes is equipped with 3 cameras, a wide-angle fixed one, and two active cameras wit…

collection_of_ground_truth_and_validation

2011-03-01T11:34:05+02:00

Long-range scenario: the Tübingen hooligan simulator A benchmark for behavioural pattern recognition should be comprised of a collection of videos pertinent to the application domain as defined for the SEARISE system. For the stadium application in the Esprit (former LTU) Arena, one would therefore like to use recordings from real soccer matches as benchmark data. However, this approach is problematic for several reasons:

contact

2008-04-11T10:00:42+02:00

Dr Marina Kolesnik, Coordinator Fraunhofer Gesellschaft, Institute for Applied Information Technology FIT Schloss Birlinghoven D-53754 Sankt Augustin, Germany Tel.: +49 2241 14 3421 Fax: +49 2241 14 1506 Web: Email:

demos

2011-02-24T17:57:00+02:00

Saliency detection Shape Based Pedestrian Detection visualization of learned motion prototypes color coded velocity hypothesis tracked salient region and optical flow tracked salient region and optical flow event fixation and pursuit left luggage person detection in crowd

gen

2009-04-14T14:57:41+02:00

gen restricted

home

2009-04-15T13:25:36+02:00

01.03.2008 – 28.02.2011 The SEARISE project develops a trinocular active cognitive vision system, the Smart-Eyes, for detection, tracking and categorization of salient events and behaviours. Unlike other approaches in video surveillance, the system will have human-like capability to learn continuously from the visual input, self-adjust to ever changing visual environment, fixate salient events and follow their motion, categorize salient events dependent on the context. Inspired by the human v…

internalmotion-driven_attention_and_grouping

2008-10-13T15:18:58+02:00

The primary objective here is to develop neural computational models of visual cortex to extract temporal visual information as well as static form information for feature-based grouping into segments and attention processes for incremental binding of fragments for figure-ground segregation, target tracking and detection of sudden scene events. Models for What and Where pathways will be employed to build feature maps for further processing. In order to employ the system with fast reaction capabi…

low_level_motion_and_contrast_detection

2010-05-28T11:01:49+02:00

Low level motion and contrast detection A large initial filter bank inspired by the visual cortex is computed on an graphic processing unit (gpu), feeding both contrast and motion detection systems. The motion and contrast detection proceed via separate two level hierarchies with competitive local interaction and excitatory feedback projections, and diffuse upwards projections, where information from the form stream is used to guide diffusion in the motion stream. The output is an array of mot…

neural_dynamics_and_interactions_with_higher_visual_areas

2008-10-13T15:26:29+02:00

Goal of this WP is to investigate dynamic interactions between different neural layers to link it to well-established variational approaches. Using neurophysiological evidence on cortical feedbacks and maps, a unifying framework for biological approaches of motion, segmentation, learning and feed-forward / feedback connectivity between different levels of hierarchy will be developed. Specifically, this framework aims at linking variational approaches and neural network models, making the above f…

objectives

2008-04-11T09:32:10+02:00

SEARISE project pursues three major objectives: * Achieve comparable to human level of performance in identification, rough categorization, fixation and pursuit of salient events in complex videos for surveillance applications. * Develop the Smart Eyes prototype. * Prove the Smart Eyes functionality in real-life environment for observation of large public spaces and restricted in-door areas.

partners

2010-06-23T15:25:47+02:00

FhG-FIT Coordinator The Fraunhofer-Gesellschaft undertakes applied research of direct utility to private and public enterprises. At present, the Fraunhofer-Gesellschaft maintains roughly 80 research units, including 57 Fraunhofer Institutes, at over 40 different locations in Germany. The Institute for Applied Information Technology FIT, pursues a user-centered approach to information and cooperation systems design. The group of Life Science Informatics is involved in the research and deve…

publications

2011-03-14T13:49:00+02:00

* D. Oberhoff. Hierarchical Bayesian Image Models. In “Object Recognition”, INTECH, accepted for publication in 2011. * M. Hoeffken, D. Oberhoff and M. Kolesnik. Temporal Prediction and Spatial Regularization in Differential Optical Flow. Submitted to ACIVS 2011. * D. Oberhoff, D. Endres, M. Kolesnik and M. Giese. Gates for Handling Occlusion in Hierarchical Bayesian Models of Images. Submitted to ICML 2011. * A. Christensen, W. Ilg, H.-O. Karnath, M. A. Giese: Specific influences o…

research

2011-03-10T09:47:39+02:00

[overview2.png] * Camera hardware development and control * Visual learning of shapes and motion patterns * Low level motion and contrast detection * Adaptive visual attention to salient events * System integration and synchronization * System integration and synchronization * System integration and synchronization

software

2011-02-25T10:43:48+02:00

Parts of the software produced during the SEARISE project is hereby released under the Cecill-C license, which is closely related to the LGPL license, thus allows use in commercial code without requiring source release of that code. Note that various parts of the software are licensed independently, usually one directory in the archive relates to one licensed module, except for auxiliary directories such as doc and doxygen which are for documentation purporses.

start

2008-04-11T10:10:45+02:00

Introduction The SEARISE project develops a trinocular active cognitive vision system: Smart-Eyes. This system will be able to detect, track and categorize salient events and behaviours. Like humans Smart-Eyes will be capable of learning continuously from visual input and of self-adjusting to changing visual environments. It will be able to fixate salient events following their motion, and to categorize salient visual events dependent on the context. Inspired by the human visual systems, a cyc…

system_integration_and_synchronization

2010-05-28T16:22:33+02:00

Saccadic movement and tracking The high-level surprise detection results in a saliency map which is fused with task bias information. The combined saliency map is used to simulate saccadic eye movement. To this end, the saliency in a fixation region is enchanced. This enchancement is dilatated over time, increasing the probability of a saccading move to a more salient region. The salient region coordinate is passed to a tracking module. The tracking employs optical flow and saliency map to p…

vision_system

2010-05-26T12:27:32+02:00

Smart-Eyes system is a three-camera system, with a wide-view global camera performing general monitoring of events and active binocular stereo cameras fixating at particular events of interest – salient events. The monitoring camera is fixed and its visual field is adjusted to overlook the whole surveying area. Detection of salient events takes place in the global view of the monitoring camera. The active stereo cameras are then guided to focus at detected salient events and to acquire their c…

visual_learning_of_shapes_and_motion_patterns

2010-05-31T11:41:43+02:00

Visual learning of shapes and motion patterns The objective here is the development of biologically inspired self-organizing hierarchical architecture for the learning-based recognition of objects and transitive and intransitive actions (i.e. actions that are directed towards a goal object or not). The focus is placed on the integration of the shape learning process, which takes place in the ventral stream and the learning of motion patterns mediated by neurons in dorsal stream. The neural arc…