An Approach to Real-time Color-based Object Tracking

Abstract

Object tracking is of great interest in different areas of industry, security and defense. Tracking moving objects based on color information is more robust than systems utilizing motion cues. In order to maintain the lock on the object as the surrounding conditions vary, the color model needs to be adapted in real-time. In this paper an on-line learning method for the color model is implemented using fuzzy adaptive resonance theory (ART). Fuzzy ART is a type of neural network that is trained based on competitive learning principle. The color model of the target region is regularly updated based on the vigilance criteria (which is a threshold) applied to the pixel color information. The target location in the next frame is predicted using evolving extended Takagi-Sugeno (exTS) model to improve the tracking performance. The results of applying exTS for prediction of the position of the moving target were compared with the usually used solution based on Kalman filter. The experiments with real footage demonstrate over a variety of scenarios the superiority of the exTS as a predictor comparing to the Kalman filter. Further investigation concentrates on using evolving clustering for realizing computationally efficient simultaneous tracking of different segments in the object