Rapid seismic intensity estimation supports situational awareness, emergency response, and the assessment of earthquake-induced structural risk. Instrument-based systems provide authoritative measurements, but limited deployment density and installation cost can restrict spatial coverage. Video sensing offers a complementary route by converting the visible motion of a calibrated target into physical motion indicators; its reliability, however, depends on stable frame-wise localisation under illumination variation, blur, and target deformation. This paper presents FAMTrack, a vision-based seismic intensity prediction framework whose tracking module augments a one-stream Vision Transformer with a parallel Frequency-Aware Matching branch that encodes patch-wise Fourier amplitude and phase. A Cross-Domain Fusion Module exchanges spatial and frequency evidence through bidirectional cross-attention without modifying the detection head or tracking loss. On LaSOT, FAMTrack-256 attains a 75.1% AUC, 4.0 percentage points above OSTrack-256. On GOT-10K, it reaches 77.9% AO, a gain of 4.2 points. In the held-out seismic-video evaluation, FAMTrack reduces the displacement RMSE from 0.523 to 0.332 mm and improves window-level intensity prediction accuracy from 80.0% to 86.7% relative to OSTrack-256. Under the fixed calibrated-video pipeline and the evaluated controlled conditions, these results indicate that frequency-aware localisation is associated with improved public-benchmark tracking, motion recovery, and seismic intensity estimation.

Accurate target tracking is essential for video-based seismic-motion measurement, as tracking errors can directly affect displacement reconstruction and subsequent seismic-intensity estimation. To improve tracking accuracy and robustness under vibration scenarios, this paper proposes TopoRelTrack, a visual tracking method based on self-supervised pre-training. In the pre-training stage, a template-guided masked autoencoder (TMAE) uses the target template to guide masked reconstruction in the search region, enabling the model to learn tracking-oriented feature representations from unlabeled video sequences. A topology-preserving scanning Mamba (TPSMamba) is further introduced to efficiently capture long-range spatial dependencies while preserving spatial structural information. In addition, a sparse dynamic relation matcher (SDRM) establishes adaptive and fine-grained correspondences between the template and search regions to enhance target discrimination and suppress background interference. The TPSMamba and SDRM are then transferred to the downstream tracking network for continuous target localization and displacement reconstruction. Experiments on the LaSOT and GOT-10K benchmarks show that TopoRelTrack achieves an area under the curve (AUC) of 75.9% and a precision of 81.8% on LaSOT, as well as an AO of 78.5% on GOT-10K, outperforming the compared trackers. In controlled shaking-table video experiments, the proposed method also achieves lower displacement reconstruction errors and improved seismic-intensity prediction performance. The results demonstrate that TopoRelTrack provides accurate and robust target tracking for video-based motion measurement under controlled shaking-table excitation.