A biologically inspired optical flow system for motion detection and object identification
Metadata[+] Show full item record
Optical flow is possibly the best known method for motion segmentation. However its application is restricted to offline processing as it requires extensive computational resources and time. This thesis explores an optical flow method derived from observation on vision system of diptereous insect. The proposed method , Biological Optical flow (BioOF) was implemented using series of first order filters, and, therefore is much faster than any existing machine coded optical flow algorithm beside being hardware implement able. Like other optical flow methods, the output of proposed BioOF has two components: horizontal optical flow and vertical optical flow; both of them can be combined in order to get a better final result in terms of motion segmentation. Unfortunately, this combined output of the BioOF can be heavily coupled with noise. So, in order to remove the noise, intensive image processing had to be performed. The result was an algorithm that can provide a good contour of the segmented object in an image. Finally the object contour is converted to a Fourier feature space leading to a representation that is rotational and translational invariant. Over this feature space various classification algorithms including SVM, feature subset forward selection, Scatter matrix, and a simple linear classifier using principal component analysis and Mahanabolis distance were investigated.