Multiple faces tracking based on joint kernel density estimation and robust feature descriptors

In this paper, we present a robust implementation of multi-face tracker using joint feature model, Kalman filter-based mean-shift and speeded-up robust features (SURF), which can tolerate interference caused by objects of similar color, partial occlusion, total occlusion, rotation and scale change. The joint feature model for each person combines the non-parametric distribution of colors in the face region and gradient information of face, Mean-shift based on Kalman filter is adopted to update the position and velocity of the object in real-time and predict the locations in the subsequent frame, and SURF solves the object-recovery problem in occlusion. Experimental results demonstrate the efficiency of the tracking algorithm and the recovery capability even in case of total occlusion.