Home Research BCCN I Project Details Computer Vision and Image Analysis
Personal tools

Computer Vision and Image Analysis

Computational Neuroscience

The goal of this project is to develop computer vision software based on visual cortical operational principles as well as to understand the statistics of images. It is in part the continuation of the European funded ECOVISION project.

Images are represented in a sparsified way by so-called Primitives which are elements that capture at one location multiple features like orientation, color, optic flow, stereo disparity, etc. These Primitives can be combined into line-like segments and also traced across multiple camera frames yielding a highly accurate 3-D representation about the scene.

primitives

Illustration of Primitives on an example scene. 3D reconstruction is also shown.

Currently, we are working to improve optic-flow analysis, as well as 3-D interpolation into structureless image regions. 

Our cooperation partners (group of N. Krüger) are at the moment trying to push the system into fast real-time.

Free, Calibrated Data Base with Ground Truth: A set of highly calibrated 3-D movies with laser range-finder ground truth is being provided for the community suitable for general benchmarking of computer vision approaches.  The laser range finder finds the distance of a reflective object by sending a narrow laser pulse towards the objects and measuring the time taken by the pulse to be reflected off and returned to the sender. In this way, a 3D model of the world is constructed. Information about the 3D set up of a visual scene is important to find both the true disparity of stereo pairs and  the true optic flow of image sequences .

rangedata

An example scene and the corresponding 3D information acquired with laser range scanner.

Methods for Optic-Flow Estimation from Global Fourier Components: We are developing a novel Fourier-based technique for the estimation of  optic-flow fields from image sequences.  Here, the instantaneous velocity of an image point is inferred directly from the global (3D) Fourier components of the image sequence. Unlike other techniques,  this method utilizes the fact that precise spatiotemporal information is encoded in the 3D Fourier components of the image sequence. Hence, spatiotemporal windowing is not required when estimating local velocities. Currently, the performance of the technique is being evaluated using 3-D movies from our laser range finder.
 Walking


                          Estimated optic-flow field of a walking person derived from the global 3D Fourier components of the image sequence.



Main cooperation Partners:

Norbert Krüger (Odense, Denmark) and Nicolas Pugeault (Edinburgh, Scotland).



Belongs to Group(s):