Robust relationship learning to illumination in a camera network

Abstract

Multi-target tracking in a camera network is one of the fastest growing fields. Re-identification is one of the most important and challenging parts of multi-target multi-camera tracking. In this paper, we introduce an approach to overcome illumination change problems. The proposed framework uses convolutional neural networks to extract features and then computes relationship matrices based on connectivity between the features. The framework is trained to avoid the ambiguity of objects that are seriously affected by illumination changes. The relationship matrix is calculated by feature matrices based on multiple brightness, which is learned by considering all the brightness cases. Experimental evaluation shows that the proposed method outperforms state-of the art competitors.