TRACE-Loc: Causal Trajectory-Constrained Cross-View Localization for GNSS-Denied UAVs

UAV-to-satellite visual localization can provide absolute position observations when Global Navigation Satellite System (GNSS) signals are denied or lost. Without a reliable initial position, however, cross-view differences and repetitive textures can cause single-frame retrieval to jump to distant but visually similar regions. This paper presents TRACE-Loc, an online trajectory-constrained cross-view retrieval framework for initial localization. A DINOv3 frontend with view-conditioned low-rank adaptation (VC-LoRA) generates current-frame candidates from a global satellite gallery. The online reranker projects recent location hypotheses using heading, elapsed time, and causal visual speed, then recursively selects a motion-consistent path from the union of current retrievals and historical projections. The path recursion without historical projection coincides with a beam-width-limited Viterbi decode over the retrieval candidates, so the method can be read as beam Viterbi plus causal candidate projection. Experiments use a unified gallery of 15,130 UAV-VisLoc satellite tiles and 1215 queries from two regions excluded from both training and parameter selection. Across three independent training runs, TRACE-Loc raises C@1 from 48.67 ± 0.41% to 58.33 ± 2.84%, a paired gain of 9.66 ± 2.69 percentage points; the single representative run used for detailed comparisons increases C@1 from 49.05% to 61.23%, reduces median localization error from 150.92 m to 93.75 m, and raises D < 300 m from 53.66% to 70.62%. The same reranker improves C@1 across six retrieval frontends. In reset-at-every-start cold-start evaluation, TRACE-Loc increases the Kaplan–Meier probability of obtaining a stable fix within 30 s over single-frame retrieval by 2.33, 9.27, and 5.89 percentage points at the 50, 100, and 300 m thresholds (paired block-bootstrap 95% CIs excluding zero at every threshold), while the particle filter attains statistically indistinguishable early-fix probabilities and recovers faster over longer observation periods; the reported gains therefore concern acquisition from an empty state relative to single-frame retrieval, not an advantage over the particle filter. These results show that candidate-level trajectory constraints improve framewise localization in a large gallery with millisecond-level reranking overhead and increase the probability of an early stable fix.