Moving Object Detection Based on Segmentation of Optical Flow Field in Ir Image Sequence

Haifeng Lu,Tianxu Zhang

DOI: https://doi.org/10.1117/12.899282

2011-01-01

Abstract:Detecting regions of moving objects between successive frames is of widespread interest. The accuracy of motion field estimation strongly influences the performance of velocity field classification. An accurate optical flow estimation method should be introduced. Furthermore, most existing approaches of motion field classification are complex, and their performance depends on initialization parameters. In IR images, objects of interested usually appear as a slightly bright spots, when compared to the background terrain. We first introduce an accurate gradient-based optical flow computation method to estimate the motion filed which particularly suits the IR image sequence. Then a novel nonparametric and automatic method for motion segmentation was developed. A simple and meaningful "threshold" is chosen to separate the vector field into two classes. Meanwhile, the a priori entropies of motion field are defined, which the direction and magnitude of motion vectors are taken into account at the same time. We use maximum entropy sum method, based on the maximizing the information measure between two classes (object and background), to generate the optimal "threshold" and classify velocity field. After ordinary morphology post-processing, our approach outputs regions of moving objects and background from the scene. Experimental results demonstrate that our method is reasonable and performs well.