Classification of surveillance video objects using chaotic series
Hanif Azhar, A. Amer · IET Image Processing · 2012
The authors propose a framework for binary classification of challenging objects (e.g. incomplete, partial occluded, background over-lapped, scaled, outdoor) in surveillance video. The framework uses feature binding of MPEG-7 visual descriptors via chaotic series simulation. Diverse video objects are tested in multiple binary classifiers for generic classes (e.g. has_person, has_group_of_persons, has_vehicle and has_unknown). Object classification accuracy is verified with both low- and high-dimensional chaotic series-based feature binding. With high-dimensional chaotic series simulation: (i) the classification accuracy significantly improves on average, 83% compared with the 62% with the original MPEG -7 visual descriptors; (ii) ‘vehicle’ objects are clustered well, which leads to above 99% accuracy for only vehicles against other objects; and (iii) drifts in high-dimensional chaotic series, because of transient, allow the training feature vector to include subtle variations in MPEG-7 descriptor coefficients for video objects in a class. A higher variance in training feature vector, using high-dimensional chaotic series simulation, manifests these subtle variations.