Investigating fast re-identification for multi-camera indoor person tracking

Abstract In person tracking applications involving multiple cameras, person re-identification is an important step for ensuring accurate tracking of individuals as they move between camera views. However, changes in camera parameters and environmental conditions can make re-identification challenging. This is especially difficult in resource-constrained environments, as is often the case in many real-world intelligent applications. In this paper, we explore dimensionality reduction, metric learning, and classification for achieving re-identification in a computationally efficient way. We report that the covariance metric transformation is a sufficient distance metric for achieving good linear separability between identity classes, and produces better results than more complex approaches across two re-identification datasets. We also explore one-shot learning methods for performing classification, show that our Sequential k-Means algorithm outperforms other fast one-shot learning approaches, and discuss parameter tuning to improve accuracy.