Attention-based multi-anchor direct-path detection for robust TDoA localization

Accurate indoor localization with time-differenceof-arrival (TDoA) requires reliable direct-path detection in each anchor’s channel impulse response (CIR). In multipath-rich environments, reflected components can dominate the direct path, causing peak detectors to select a biased arrival. Most non-line-of-sight (NLoS) identification and mitigation approaches treat this as a per-anchor problem and operate on scalar range measurements, leaving cross-anchor structure unexploited. We propose a joint processing approach using an attention-based neural network operating on the full multi-anchor CIR matrix. For each anchor, the network predicts a direct-path placement distribution and a visibility score, which directly parameterize a residual-weighted TDoA framework without requiring hard NLoS exclusion. Supervised by channel-derived soft labels, the model captures spatial features rather than memorizing environment-specific geometry. Evaluations on ray-traced indoor scenes demonstrate that the proposed system achieves 91.96% direct-path detection accuracy, 95.68% LoS/NLoS classification accuracy, and a 2.33 m mean localization error, outperforming all evaluated baselines in mean localization error. Furthermore, the model retains strong localization performance in an entirely unseen environment, whereas fingerprinting and range-based baselines degrade substantially.