People apparel model for human appearance description

Many applications of computer vision, such as people identification and tracking in surveillance system, demand to describe the appearance of people. However, factors such as illumination, people pose and background may greatly affect the performance of the people appearance description. In this paper, a semantic model to describe people in surveillance cameras is introduced. Observing that in a short period of time, the clothes of people does not change, therefore, we construct a people apparel model based on their clothes colors, clothes pattern and dress style. These are all semantic traits which are invariant to illumination and people pose change, as well as various camera specifications. So they can provide robust description. Traits are extracted from two parts of the human body for more accurate description. Experiments are carried out on videos captured in public areas and the results show the good performance of this semantic people apparel model.