Robust human authentication using appearance and holistic anthropometric features

We propose here decision-level fusion using neural networks and feature-level fusion using boosting for the purpose of robust human authentication vis-à-vis face occlusion and disguise. Holistic anthropometric and appearance-based features feed the data fusion stage. In addition to standard head and face geometric measurements, the proposed holistic anthropometric features include additional measurements below the face to describe the neck and shoulder and their contextual relations to head and face. The appearance-based features include standard PCA or Fisherfaces. Experimental data shows the feasibility and utility of the proposed hybrid (extended geometry + appearance) approach for robust human authentication vis-à-vis occluded and/or degraded face biometrics. The authentication results presented compare favorably against both appearance-based methods and hybrid methods with anthropometric features confined to face and head. The methods proposed can train on clean data and authenticate on corrupt data, or train on corrupt data and authenticate on clean data.