Extracting full-field subpixel structural displacements from videos via deep learning

Conventional displacement sensing techniques (e.g., laser, linear variable differential transformer) have been widely used in structural health monitoring the past two decades. Though these are capable of measuring time histories with high accuracy, distinct shortcoming remains such as point-to-point contact which limits its applicability real-world problems. Video cameras years due to advantages that include low price, agility, spatial resolution, and non-contact. Compared target tracking approaches digital image correlation, template matching, etc.), phase-based method is powerful for detecting small subpixel motions without use paints or markers on structure surface. Nevertheless, complex computational procedure real-time inference capacity. To address this fundamental issue, we develop a deep learning framework based convolutional neural networks (CNNs) enable extraction full-field displacements from videos. In particular, new CNN architectures designed trained dataset generated by motion single lab-recorded high-speed video dynamic structure. As only reliable regions sufficient texture contrast, sparsity field induced mask considered via network architecture design loss function definition. Results show that, supervision full sparse field, identifying pixels contrast well their motions. The performance tested various videos other structures extract histories), indicates generalizability accurately contrast.