A spatiotemporal attentive recurrent neural network for robust RSS based indoor localization

Abstract

Indoor localization remains challenging in GPS-denied environments due to weak signal reception and non-line-of-sight conditions. Among various techniques, WLAN Fingerprinting using Received Signal Strength (RSS) has emerged as a cost-effective solution that leverages the widespread WLAN infrastructure. However, existing machine learning approaches struggle with localization accuracy due to inherent RSS variability and temporal dynamics in complex indoor environments. This paper introduces STaRLoc-Net, a novel spatio-temporal recurrent neural network framework that addresses these limitations through hierarchical deep learning and hybrid localization strategies. The proposed architecture employs four cascaded recurrent blocks integrating Gated Recurrent Units (GRU), Bidirectional Long Short-Term Memory (BiLSTM), and LSTM layers to effectively model both short-term fluctuations and long-term temporal dependencies in RSS fingerprints. Each block incorporates batch normalization, dropout, ReLU activation, and fully connected networks to progressively refine discriminative features while mitigating overfitting. A distinctive hybrid approach combines extracted deep spatio-temporal features with K-Nearest Neighbors (KNN) regression for enhanced position estimation and local consistency. Extensive experiments in two indoor environments (30 × 30 m 2 and 50 × 50 m 2 ) demonstrate that STaRLoc-Net consistently outperforms classical machine learning methods and deep learning baselines, achieving the lowest 50th and 95th-percentile error across both scenarios. Comprehensive ablation studies confirm that the fusion of multiple recurrent components enables robust generalization across varying environment scales and complexities, maintaining resilience under diverse multipath conditions. The results validate STaRLoc-Net as a scalable, environment-adaptive solution that establishes a new benchmark for RSS-based indoor localization systems.