<p>Structural health monitoring (SHM) of power equipment faces unique challenges in substation environments, including electromagnetic interference and sensor deployment constraints. To address these challenges, this paper presents an innovative real-time visual monitoring method for seismic responses of power equipment, leveraging a deep learning-based Siamese Region Proposal Network (SiamRPN). A novel frame-modified scale (FMS) approach was developed to reconstruct displacement responses with enhanced accuracy, eliminating the need for pre-calibration. Experimental validation through seismic simulation shaking table tests on a converter transformer bushing demonstrated that the FMS method substantially reduced displacement reconstruction errors compared to conventional methods, with low sensitivity to video resolution. The SiamRPN-based tracking achieved real-time processing, outperforming traditional computer vision algorithms. Dual-target tracking further mitigated camera motion effects, ensuring reliable measurements under dynamic conditions. This non-contact, real-time solution enables sensor-free monitoring through existing surveillance infrastructure, significantly advancing post-disaster assessment efficiency.</p>