Research in Computational Vision at CB, CSC, KTH

We humans are extremely good at recognizing objects in the world and the activities they are involved in. A main goal of computer vision research is to be able to equip artificial vision systems with corresponding perceptual abilities.

Applications of such computer vision modules include (i) visually guided inspection and surveillance, (ii) visually guided robots and vehicles, including safety systems for driving, (iii) automatic or semi-automatic construction of object models from multiple views (iv) remote sensing for geographic information systems and environmental monitoring (v) human-computer interaction and (vi) retrieval of images from large image databases.

A deeper understanding of the structure of the computational vision problem is also important for gaining further insights into biological vision.

A main subject of our research is to perform basic research in this area to develop new and more powerful methods for object recognition and recognition of spatio-temporal events, and to develop scale-space theory for the underlying image measurements from needs for recognition task.

During recent years, substantial progress has been made in these areas. Specifically, the use of view-based representations in terms of receptive field responses has emerged as a highly promising paradigm for visual recognition and scale-space methodologies have been demonstrated to be highly useful in this context for robust computation of image features and for adapting local image features to scale variations and perspective deformations.

Currently, we focus on furthering the notion of view-based recognition for both spatial and spatio-temporal recognition, and to co-develop scale-space theory to obtain more discriminatory image features and image descriptors for spatial and spatio-temporal recognition.

The underlying image features in terms of Gaussian derivatives and related operators have interesting similarities to receptive fields registered in biological vision. Parallel to the development of computational models and algorithms, we are therefore also relating our methods to biological vision.

In the area of computational biology, we have also performed previous work on analysing medical image data as are obtained from magnetic resonance imaging MRI och positron emission tomography PET and participated in a neuroinformatics project for constructing a database of functional and cytoarchitectonical regions in the human brain.

Previously, we have also worked on capturing human gestures by a camera for purposes in human-computer interaction.