Improved MOG Algorithm based on Adaptive Threshold for Effective Background Classification
Jeong-Su Oh · TECHART Journal of Arts and Imaging Science · 2014
In a conventional mixture of Gaussians (MOG) approach for background classification, a small threshold for background decisions causes a recognition delay in a periodic background (PBG) and a large threshold causes passing objects to be considered part of the background in a stationary background. This paper discusses an improved MOG algorithm using an adaptive threshold. Specifically, the proposed scheme evaluates the number of model changes and dominant models for both short and long-term periods; classifies backgrounds as static, non-periodic dynamic, or periodic dynamic; and assigns the appropriate threshold for each case. The simulation results confirm that the proposed approach can significantly reduce a background recognition delay, from 137 to 4 frames for a PBG, and prevent a moving object from being recognized as part of the stationary background. Therefore, we believe that this improved algorithm can be a useful tool for object tracking or background-segmentation-related fields in outdoor applications.