Human identification by using human part similarity

Identifying whether or not two persons in two set of different videos are the same person can be useful in many applications such as a video surveillance system. However, the problem is very difficult because in actual applications two videos can contain different amount of information. This paper proposes an approach to answer the question in term of a similarity score with level of confidence based on the available amount of information. The proposed approach is based on the assumption that there are multiple cameras per scene taken in different views. For this paper, two cameras perpendicular to each other are set for taking a set of videos of a person. Two persons in two different set of videos are compared by decomposing the person figure into 3 parts, the upper, the middle, and the lower parts. In each part, the volume estimated by the stereo views of the body is used as a static feature of a person. Besides the volume, other static or dynamic features are considered. In the upper part, the shapes of the part seen from both cameras are used. In the middle and the lower part, the movement of hands and the movement of feet are used, respectively. The hand or foot movement is represented in term of its period and maximum width. In the comparison and scoring step, the similarity scores of all feature are computed first before weighted summation them to get the total similarity score. The experiments were carried out using 5 persons walking in two different angles with and without a bag for the total of 192 pairs of video sets. The results showed that all similarity scores of pairs with different persons have lower similarity scores than the scores of pairs with the same person.