Lightweight Learning-Based Feature Selection for Real Time Optical Flow Navigation

Abstract

Accurate state estimation is essential for autonomous quadrotor navigation in GPSdenied environments. Although Visual-Inertial Odometry (VIO) systems perform well, their accuracy depends on the quality and stability of visual features. Classical detectors are efficient but often produce redundant or unstable features, while learning-based methods are more robust but computationally expensive for embedded platforms. This thesis presents a lightweight fully onboard navigation framework that improves estimation accuracy by enhancing feature quality rather than feature density. A compact adaptive convolutional neural network predicts feature trackability and removes unreliable features before Lucas–Kanade optical flow tracking. An adaptive threshold regulates the retained feature count. Visual velocity estimates, with inertial rotational compensation, are fused with IMU and range measurements using an Extended Kalman Filter. Experiments on the Quanser QDrone2 show reduced computational cost, improved estimation accuracy, and real-time performance above 60 Hz.