Person Re-Identification Between Visible and Thermal Camera Images Based on Deep Residual CNN Using Single Input

초록

In recent years, numerous studies have been undertaken regarding person re-identification (ReID), an important issue for intelligent surveillance systems. Person ReID, however, is an extremely difficult problem because of variables such as different viewpoints and poses, and varying lighting in person regions in images that have been captured from remote distances. A majority of the studies have been performed for visible-light camera-based person ReID, which can be used only in a limited environment owing to the characteristics of a visible-light camera that are considerably dependent on the illumination. To overcome this problem, studies have been conducted for multimodal camera-based person ReID. However, because the previous studies used two or more input images, the computational complexity was high. This paper proposes a novel person ReID method that simplifies the convolutional neural network (CNN) structure by combining visible-light and thermal images as a single input. This method overcomes the limitation of visible-light camera-based person ReID using both a visible-light and thermal camera. To verify the performance of the proposed method, two open databases, the DBPerson-Recog-DB1, and Sun Yat-sen University multiple modality Re-ID (SYSU-MM01) databases were used. The method proposed in this study demonstrated excellent performance compared to the conventional methods.

키워드