User-centric incremental learning model of dynamic personal identification for mobile devices

This study presents a user-centric incremental learning model based on the proposed output selection strategy (OSS) and multiview body direction estimation for dynamic personal identification systems on mobile devices. First, the OSS filters primitive results generated from the classifier, so that the refined information can be used to update the learning model. Second, the robustness of the model is enhanced by using different views of faces as system input, which allows the learning model to adapt itself when either of facial views is not available. In addition, the body direction estimation method is proposed for estimating multiple views of a person by matching templates of human shapes and skin colors. An experiment on 168,000 test samples (20 classes with three facial views) is conducted to evaluate the proposed system. The experimental results show that the proposed method improves accuracy by more than 40 % compared to baseline, and correspondingly confirms the effectiveness of the proposed idea.