Robust human re-identification using mean shape analysis of face images

Human re-identification is an important component in many application domains especially the automatic surveillance system. This paper proposes a robust method to re-identify persons using their face shapes based on the Active Shape Model (ASM) and the Procrustes Shape Analysis (PSA). The ASM-based technique is used to extract landmark points of each face image, as the feature. Then, the Procrustes Distance (PD) is used to measure the similarity between two ASMs of any two face images. In addition, the trained ASMs of each subject can be grouped into clusters, using the PD-based K-mean clustering. Then, the Procrustes Mean Shape (PMS) is computed for each cluster using all belonging ASMs. In stead of using ASM of individual face image, the PMS is used as the representative face model. This process is performed to increase the robustness and reduce the number of models representing each subject. The proposed method is evaluated on the well-known face datasets and the real-world scenario of the security guard re-identification under the real environment. The experimental results and comprehensive comparisons show a very promising performance of the proposed method.