Neural forward–backward reconstruction for motion‑blur‑resilient trajectory recovery under intermittent visibility

Fast target motion poses a fundamental challenge to optical and imaging systems, as finite exposure time and intensity integration introduce motion blur and intermittent visibility, leading to irreversible information loss at the image formation stage. Under such conditions, robust recovery of point‑level motion trajectories becomes difficult, and conventional tracking‑based methods that perform well for slow or moderate motion degrade significantly at high speeds due to accumulated drift and instability. This study proposes a hybrid forward–backward trajectory reconstruction framework that addresses motion‑blur‑limited imaging by integrating data‑driven prediction with signal‑informed post‑processing. A deep learning model first estimates preliminary trajectories from incomplete visual observations, followed by temporal filtering, smoothing, and keypoint‑based bidirectional correction to suppress drift and restore trajectory continuity under intermittent observability. The proposed framework is evaluated on high‑frame‑rate pitching sequences characterized by rapid distal joint motion and partial occlusion. Comparative experiments against CoTracker, TAPIR, and MFT demonstrate consistent reductions in mean absolute error (MAE) and mean squared error (MSE), along with higher coefficients of determination (R2). A weighted keypoint‑based evaluation further highlights improved accuracy for distal joints, particularly during frames with abrupt wrist rotation and visibility interruptions. The proposed approach generalizes to other optical imaging scenarios involving fast motion, motion blur, and incomplete measurements, enabling stable trajectory reconstruction beyond the limits of conventional exposure‑based imaging.