Tracking unknown moving targets on omnidirectional vision

Abstract

This paper presents an integrated method by using optical flow and kernel particle filter (KPF) to detect and track moving targets in omnidirectional vision. According to the circle character in omnidirectional image, the algorithms of optical flow fields and kernel particle filter are improved based on the polar coordinates at the omnidirectional center. The edge of a motion object can be detected by optical flow fields and is surrounded by a reference region. In order to resolve some shape distortions such as rotation and scaling in the omnidirectional image a dynamic elliptical template with affine transformations is constructed and its motion model is established to predict particle state. Histograms are used as the features in the reference region and particle regions. The Bhattacharyya distance is computed for particle weights. Gaussian kernel function is used in kernel particle filter. Experiment results show that the method can detect and track moving objects and has better performance at real-time and accuracy.