A New System for Event Detection from Video Surveillance Sequences - Advanced Concepts for Intelligent Vision Systems - page 117

specific features within a given feature modality is shown to be consistent across
all concepts/events. However, the relative importance of one feature modality
vs. another may change from one concept/event to the other. The following de-
scriptors had the top overall performance for both search and concept modeling
experiments:
- Color Histogram: global color represented as 128-dimensional histogram in
HSV color space.
- Color Moments: localized color extracted from 3x3 grid and represented by
the first 3 moments for each grid region in Lab color space as normalized
255-dimensional vector.
- Co-occurence Texture: global texture represented as a normalized
96-dimentional vector of entropy, energy, contrast and homogeneity extracted
from the image gray-scale co-occurence matrix at 24 orientation.
- Gabor Texture: Gabor functions are Gaussians modulated by complex sin-
isoids. The Gabor filter masks can be considred as orientation and scale-tunable
and line detectors. The statistics of these micro-features in a given region can
be used to characterize the underlying texture information. We take 4 scales
and 6 orientations of Gabor textures and further use their mean and standard
deviation to represent the whole frame and result in 48 textures.
- Fourier: Features based on the Fourier transform of the binarized edge image.
The 2- dimensional amplitude spectrum is smoothed and down-sampled to
form a feature vector of 512 parameters.
- Sift:The SIFT descriptor [7] is consistently among the best performing inter-
est region descriptors. SIFT describes the local shape of the interest region
using edge histograms. To make the descriptor invariant, while retaining
some positional information, the interest region is divided into a 4x4 grid
and every sector has its own edge direction histogram (8 bins). The grid is
aligned with the dominant direction of the edges in the interest region to
make the descriptor rotation invariant.
- Combined Sift and Gabor.
- Wavelet Transform for texture descriptor: Wavelets are hybrids that are
waves within a region of the image, but otherwise particles. Another impor-
tant distinction is between particles that have place tokens and those that do
not. Although all particles have places in the image, it does not follow these
places will be represented by tokens in feature space. It is entirely feasible
to describe some images as a set of particles, of unknown position. Some-
thing like this happens in many description of texture. We performe 3 levels
of a Daubechies wavelet [4] decomposition for each frame and calculate the
energy level for each scale, which resulted in 10 bins features data.
- Hough Transform: As descriptor of shape we employ a histogram based on
the calculation of Hough transform [9]. This histogram gives information
better than those given by the edge histogram. We obtain a combination of
behavior of the pixels in the image along the straight lines.
- Motion Activity: We use the information calculated by the optical flow,
through concentrating on movements of the various objects (people or