Chaos and MPEG-7 based feature vector for video object classification

We propose a method to generate unique feature vectors for video objects using chaos theory and MPEG-7 visual descriptors. We consider each feature element of visual descriptors as a dynamic system. The proposed method performs feature binding of the re-constructed trajectory of simulated chaotic attractors using histogram analysis. The binding derives two feature vectors different from the original MPEG-7 one, for each video object. We use the new vectors for object classification. Low (e.g., Logistic Map) and high (e.g., Mackey-Glass) dimensional chaotic attractors are used. Dynamic feature reduction (35.44% on average) in the proposed feature vectors are achieved from the MPEG-7 feature vector. Cross validation accuracy with different classifiers shows significant (87.6% on average) improvement with the proposed feature vectors over that (73.2% on average) of the MPEG-7 feature vector.