Our main approaches to feature extraction have concerned low- and high-level image features. Low-level feature extraction concerns the edges of an object or its curvature, for moving shapes we can find the motion too. There are techniques for filtering the image to improve its quality. The standard techniques are described in our textbook. This extends to high-level feature extraction, which is about how we find shapes and objects in computer images. Our research has developed more advanced techniques, often driven by our interest in biometrics, and especially in walking people.
Context-aware super pixles Symmetry for Finding People
Object extraction by image ray transform
For examples of lower level image-shape analysis, we have developed new approaches to detecting symmetry, especially the symmetry of moving people. If human motion was to be that of a pendulum (or grandfather clock), there would be perfect symmetry. But human motion is not perfectly symmetric and actually the differences can be used to recognise people.
Some of our approaches have considered using physical analogies for shape extraction. We have developed techniques based on the propagation of heat (which led to a new technique for moving-edge detection) and water (a bit like the watershed algorithm). We have also modelled the use of light for object shape extraction in a technique we have named the image ray transform. It’s a bit like shining torches positioned at random positions in an image and the pixel intensities control the refractive index. You are welcome to try our online demonstration of the image ray transform.
Finding people by GHT with motion
Motion Modelling Finding the face
