Real-time Tracking of Non-rigid Objects Using Modified Kernel-based Mean Shift and Optimal Predictoin

An efficient scheme for real-time color-based tracking of non-rigid objects is proposed. The central computational module is based on mean shift iterations. It computes the most probable target position in the current frame, while the prediction of the next target location is computed using a Kalman filter. The dissimilarity between the target model and the target candidates is expressed by a metric based on the Bhattacharyya coefficient. In this work, we have adapted the kernel profile (used in calculating the feature histogram) with a binary mask generated by proposed adaptive background subtraction scheme. The modified kernel calculates the feature histogram only for foreground pixels and prevents background pixels from causing the estimation process to deviate. The adaptive background subtraction algorithm may fail under varying illumination and shadow conditions. To overcome this problem, we have decomposed the incoming image into its intrinsic components (illuminance and reflectance), and have designed an adaptive background subtraction scheme using the reflectance image. The experimental results show the capability of the proposed tracker to handle real-time partial occlusions, significant clutter, and also target scale variations.