High-resolution synthetic aperture radar (SAR) enables persistent monitoring under challenging illumination and weather conditions, but reliable vehicle detection remains difficult due to speckle, geometry-dependent scattering, and the scarcity of annotated SAR datasets for site- and mode-specific scenarios. This paper presents an unsupervised refinement pipeline for vehicle detection in TerraSAR-X imagery. Candidate regions are first generated within a region of interest using ordered-statistics CFAR (OS-CFAR). Each candidate is then represented by a feature embedding extracted from a frozen pretrained MobileNetV2 backbone, and embeddings collected across a short multi-temporal stack are clustered with K-means to separate vehicle-like signatures from other high-response candidates. Evaluation is supported by co-registered UAV imagery for qualitative verification. Across seven TerraSAR-X acquisitions acquired at approximately two-week intervals (including one held-out acquisition reported separately for generalization), the proposed clustering stage reduces false positives relative to OS-CFAR-only candidate sets and improves end-to-end detection quality, yielding an effective F1-score of 0.736 while preserving a low-annotation operating regime.