KLTNet: Learning-Based Sparse Feature Tracking for Improved Monocular VIO

zhenjun_zhao · x · 2026-08-26

The paper presents KLTNet, a learning-based sparse feature tracker designed to replace classical KLT trackers in VIO front ends. It uses a coarse-to-fine, dense-to-sparse architecture, combining low-resolution dense optical flow for robust initialization with triplet-patch refinement for accuracy. Experiments on VINS-Mono and OpenVINS show improved tracking and odometry accuracy while maintaining real-time performance on embedded platforms.

Original post →

More from Embodied

Embodied channel →