Multiple face tracking method in the wild using color histogram features

This paper presents a new framework for the tracking of multiple faces. In order to address this tracking problem in the challenging environments, we first adopt a robust face detector based on Multi-task Cascaded Convolutional Networks (MTCNN) and a very efficient tracker exploiting Kernelized Correlation Filters (KCF). Then, we incorporate the detector and tracker into our framework by proposing a new data association method. In our association scheme, we consider color histogram features as well as geometric overlaps, so that it works robustly in the presence of occlusions and crossovers. We conducted experiments on the selected examples of 300VW database and a challenging test video sequence (TMBS). Experimental results have shown that the proposed method works robustly for challenging scenarios in real time (49 fps).