Weakly-Supervised Crowdsourcing Localization Based on Self-Learning Spatial-Temporal Attention Network

Crowdsourcing offers an effective solution to alleviate the heavy burden of data collection and annotation in fingerprint-based localization, particularly for Sixth-Generation (6G) communications. However, extracting robust and reliable features from sparsely labeled crowdsourced data remains highly challenging, significantly limiting localization accuracy. To address this issue, this paper proposes a Weakly-Supervised Crowd-sourcing Localization (WSCL) framework, which is based on a Self-learning Spatial-Temporal Attention Network (SSTAN). The framework consists of two phases: self-learning phase and fine-tuning phase. By applying spatial-temporal supervisory signals, a linear embedding attention encoder, and a linear projection layer, SSTAN extracts features rich in spatial-temporal information via self-learning. In the fine-tuning phase, a small amount of labeled data is used to learn the mapping relationship from the feature space to the location space, achieving satisfactory localization results. Experiments conducted in a real-world underground parking lot environment reveal that our proposed WSCL framework achieves a localization accuracy of 1.41m within an area of 1223m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> using only 10% labeled fingerprint samples.