Analyzing smoothness of oscillatory motion from NIR video via optical flow vector field

Abstract

This thesis develops a non-contact framework for assessing the smoothness of oscillatory motion using near-infrared (NIR) video. Traditional vibration monitoring relies on contact-based sensors such as accelerometers, which alter system dynamics through mass loading and provide only point-wise measurements. To overcome these constraints, this research introduces a new methodology that integrates NIR video acquisition with geometric analysis of motion vector fields in the polar complex plane. The proposed framework extracts motion vector fields from sequential NIR video frames using optical flow and represents them in the complex plane. The spatial Jacobian of each vector field is computed, and its eigenvalues are extracted. Motion is classified as smooth when the maximum eigenvalue magnitude lies within the unit circle, and jerky when it exceeds this boundary. This operational criterion is validated through extensive comparison with Inertial Measurement Unit (IMU) data. A key finding from our companion study shows that the maximum eigenvalue derived from video correlates strongly with accelerometer peak readings, establishing a direct link between non-contact video analysis and traditional vibration metrics. Three complementary measures are introduced: (1) a frame-by-frame smoothness classification based on the unit-circle criterion; (2) a characteristic distance metric that quantifies the similarity between different motion states; and (3) a smoothness index providing an intuitive summary of overall motion regularity. Experimental validation employs NIR video recordings of human locomotion (walking, running, and loaded running) and vehicle suspension motion across different terrains and payloads. Results demonstrate strong agreement between the video-based method and IMU data (correlation coefficients exceeding 0.90). A significant positive correlation is identified between mechanical loading and motion smoothness, with η increasing systematically with load in both human and vehicle experiments (p < 0.001, effect sizes > 0.4). This finding aligns with physical damping principles: increased load enhances effective damping, producing smoother, more regular motion. The framework establishes that NIR video sensing achieves accuracy comparable to contact-based methods while eliminating sensor-induced artifacts and providing spatial resolution. This work offers a scalable, non-invasive tool for motion smoothness assessment, with applications in biomechanical load optimization, rehabilitation monitoring, and mechanical overload diagnostics.