A Discriminatively Learning Model with Illumination Transfer for Inter-Camera Pedestrians Association
Shijun Zhong, Chunyan Yu, Jiali Lin · 2018
Inter-camera pedestrians association always employs appearance features to merge tractlets of the same pedestrian into a whole. However, appearance features are always view- and illumination- sensitive. In this paper, we present a method to solve inter-camera pedestrian association via discriminative learned feature in a stable way with illumination transfer. First, we proposed a discriminative feature learning model which is a convolution siamese network that combines the verification and identification losses. Furthermore, we introduce color brightness transfer reduce color distortions under different illumination. To learn proper brightness transfer function, a fuzzy color cluster is used to model the change of color brightness between different cameras. The experiments show the effectiveness of the proposed method and achieve the state-of-the-art in the benchmark NLRP_MCT dataset.