Neighborhood Difference-Based Passive Localization With Automatic Annotation in Visible Light Sensing

Abstract

The rapid development of visible light (VL) communication has stimulated increasing interest in VL-based passive positioning. However, false target positioning remains a major challenge in multitarget sensing scenarios. Fingerprint-based approaches are widely used for false-target recognition, but two issues limit their effectiveness: the high labor cost of labeling training data, and the degradation of recognition accuracy caused by region-dependent variations in shadowed link distributions. To address these challenges, this article proposes a novel multitarget passive sensing framework. First, a Hungarian algorithm-based data annotation mechanism is introduced to automatically partition training samples into true- and false-target datasets, significantly reducing manual labeling effort. Second, a multidimensional neighborhood-difference feature extraction method is developed to enhance the identification accuracy of false targets for target counting. Extensive simulation results demonstrate that the proposed method reduces the mean counting error by approximately <inline-formula> <tex-math notation="LaTeX">${48}\%$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">${41}\%$ </tex-math></inline-formula> compared with the nonfeature-difference-based fingerprint method and the successive cancellation method, respectively.