Scholarlink

Introduction <p>Ground-target tracking using aerial robots, particularly quadrotors, is widely applied in surveillance and cinematic filming. Despite advances in machine vision, existing tracking approaches often suffer from inaccuracies caused by occlusion, motion blur, and changing illumination during UAV maneuvers.</p> Purpose <p>This study aims to develop a robust and efficient vision-based tracking framework that improves tracking accuracy and stability under challenging conditions such as occlusion and rapid motion. Specifically, the work seeks to enhance UAV tracking performance by integrating advanced visual tracking with predictive control and reducing computational complexity through learning-based control.</p> Methods <p>A vision-based tracking framework is proposed that integrates the Discriminative Model Prediction (DiMP) tracker with Model Predictive Control (MPC). The DiMP tracker estimates future target motion probabilities and adapts dynamically when the target reappears after occlusion, ensuring reliable tracking under high-speed and visually challenging conditions. Visual feedback is incorporated into MPC to enable predictive trajectory adjustments based on real-time observations. A neural imitation learning controller trained to improve system efficiency. </p> Results <p>Experimental evaluation demonstrates that the proposed DiMP–MPC framework significantly outperforms conventional MPC-based approaches. Tracking precision increased from 49.7% to 68.9%, trajectory error was reduced by approximately 3%, and smoother camera motion was achieved, particularly during occlusion recovery. The imitation learning-based controller further reduced computational cost while maintaining robust tracking performance.</p> Conclusion <p>The proposed DiMP–MPC framework provides a robust, efficient, and accurate solution for vision-based UAV target tracking. This approach offers a practical and scalable solution for real-world aerial tracking applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Neural Imitation of Model Predictive Control for Vision-Based Guidance of Agile Micro Air Vehicles in Film Production

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Mostafa Jalalnezhad,
  • Sotirios Spanogianopoulos

摘要

Introduction

Ground-target tracking using aerial robots, particularly quadrotors, is widely applied in surveillance and cinematic filming. Despite advances in machine vision, existing tracking approaches often suffer from inaccuracies caused by occlusion, motion blur, and changing illumination during UAV maneuvers.

Purpose

This study aims to develop a robust and efficient vision-based tracking framework that improves tracking accuracy and stability under challenging conditions such as occlusion and rapid motion. Specifically, the work seeks to enhance UAV tracking performance by integrating advanced visual tracking with predictive control and reducing computational complexity through learning-based control.

Methods

A vision-based tracking framework is proposed that integrates the Discriminative Model Prediction (DiMP) tracker with Model Predictive Control (MPC). The DiMP tracker estimates future target motion probabilities and adapts dynamically when the target reappears after occlusion, ensuring reliable tracking under high-speed and visually challenging conditions. Visual feedback is incorporated into MPC to enable predictive trajectory adjustments based on real-time observations. A neural imitation learning controller trained to improve system efficiency.

Results

Experimental evaluation demonstrates that the proposed DiMP–MPC framework significantly outperforms conventional MPC-based approaches. Tracking precision increased from 49.7% to 68.9%, trajectory error was reduced by approximately 3%, and smoother camera motion was achieved, particularly during occlusion recovery. The imitation learning-based controller further reduced computational cost while maintaining robust tracking performance.

Conclusion

The proposed DiMP–MPC framework provides a robust, efficient, and accurate solution for vision-based UAV target tracking. This approach offers a practical and scalable solution for real-world aerial tracking applications.