An improved kernelized-correlation-filter spatial target tracking method using variable regularization and spatio-temporal context model

The dim target tracking is essential for the spatial surveillance system. Considering that the starry image sequences acquired by imaging sensors often has low Signal-to-Noise Ratio (SNR), the brightness of a spatial target is often susceptible to the background interferences, such as the night clouds and the atmospheric turbulence, etc, and become dim and instable, its shape and profile is also blurred and lack of texture information. In order to extract the target from background, Spatio-Temporal Context Model (STCM) based filtering theory is applied in this paper and used to improve the traditional Kernelized-Correlation-Filter (KCF) target tracking method. It introduces a spatial weighting function that can pre-enhance the point target and suppresses the background interferences. So the tracking drift phenomenon is relieved when the moving object being obstructed temporarily. Considering that L1 regularization is easier to obtain sparse solutions and L2 regularization has smoothness property, the regularization function of the regressive classifiers in KCF target tracking method is renewed by using variable L1 or L2 regularization instead. The index of regularization in the improved regression model is a piecewise function, which is determined by the cost function during learning period that can distinguish the target star point from the background point by using the characteristics of points (such as brightness, etc.)The numeral simulation and actual processing results show that, comparing with the traditional Kernelized-Correlation-Filter (KCF) methods, the proposed method owns more robustness and precision in the starry images with low signal-to-noise ratio and complex background.