Research Analyzer
← Back ICRA 2026

Lightweight Learning-Based Feature Selection for Real-Time Optical Flow Navigation on a Quadrotor Platform

Ali Abosaad, Jinjun Shan

PDF

AI summary

Key figure (auto-extracted from paper)
A lightweight CNN-based feature pruning framework with adaptive thresholding significantly improves quadrotor navigation accuracy and reduces computation time while maintaining real-time performance on embedded hardware.
Feature selection CNN pruning optical flow quadrotor navigation GPS-denied real-time estimation

Problem

Conventional visual-inertial odometry relies on dense feature extraction that is computationally heavy and prone to drift from low-quality features, while existing learning-based detectors are too heavy for real-time embedded deployment on resource-constrained quadrotors.

Approach

The authors propose a compact CNN that scores and prunes unreliable ORB features using an adaptive threshold, integrated with Lucas-Kanade optical flow, rotational false-velocity compensation, and an Extended Kalman Filter for real-time state estimation.

Key results

  • Up to 75–80% reduction in position RMSE compared to Fourier-based Phase Correlation
  • 25–30% reduction in computation time while sustaining over 120 Hz processing frequency
  • CNN pruning reduces baseline position error by 88% in ablation studies
  • Adaptive thresholding dynamically maintains a stable feature count across varying environments

Why it matters

Enables robust, drift-resistant, and computationally efficient GPS-denied navigation for resource-constrained aerial robots in safety-critical applications.

Abstract

Accurate state estimation in GPS-denied environ- ments is critical for autonomous quadrotor navigation. Conven- tional visual-inertial odometry (VIO) pipelines rely on dense feature extraction and tracking, which increases computational cost and is prone to drift when low-quality features dominate. Although learning-based detectors improve robustness, most are too computationally heavy for embedded deployment. This paper proposes a lightweight learning-based feature selection framework that prunes unreliable features to enable effi- cient optical flow navigation. A compact Convolutional Neu- ral Network (CNN) is employed, with its pruning threshold adaptively adjusted to maintain a stable number of reliable features. The CNN augments ORB and Lucas–Kanade optical flow in a multithreaded pipeline with rotational false-velocity compensation and EKF fusion. Experiments on the Quanser QDrone2 demonstrate up to 75–80% reduction in position RMSE and approximately 25–30% reduction in computation time compared to the Fourier-based Phase Correlation (FPC) method, while sustaining real-time performance above 120 Hz without reliance on external localization systems.

Index terms

Vision-Based Navigation Deep Learning for Visual Perception Sensor Fusion

Related papers