Abstract
Computer vision technologies for people detection and tracking are widely required in various fields such as surveillance system applications for security system and marketing services. In this paper, we propose a method for tracking a person using several video frames in multi-camera network with non-overlapping views. In many surveillance camera systems, it is desirable to determine if a given individual has been previously observed multi-camera environment. However, since the human features are unstable under the change of viewpoint and posture, this is likely due to the difficulty in identifying individuals with non-overlapping cameras. Therefore, this paper presents a method of people re-identification in multi-camera based on co-visible features using observed crossover region between people. The method is evaluated using a public viewpoint invariant pedestrian recognition dataset (VIPeR) and the results are shown to be superior to the previous benchmarks.
