Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Open-World Drone Active Tracking with Goal-Centered Rewards

Haowei Sun, Jinwu Hu, Zhirui Zhang, Haoyuan Tian, Xinze Xie, Yufeng Wang, Xiaohua Xie, Yun Lin, Zhuliang Yu, Mingkui Tan

TL;DR

The paper tackles robust open-world drone active tracking by introducing the DAT benchmark and a reinforcement learning method GC-VAT. DAT provides 24 city-scale, high-fidelity scenes with unlimited scene generation via a digital twin, enabling rigorous evaluation of tracking under diverse, dynamic conditions. GC-VAT uses a Goal-Centered Reward and Curriculum-Based Training to overcome failures of prior distance-based rewards under tilted viewpoints, achieving superior tracking performance in both simulated and real-world tests, including notable sim-to-real transfer. The work advances open-world VAT by delivering a unified benchmark, provable reward design advantages, and practical training strategies with demonstrated robustness and real-world applicability.

Abstract

Drone Visual Active Tracking aims to autonomously follow a target object by controlling the motion system based on visual observations, providing a more practical solution for effective tracking in dynamic environments. However, accurate Drone Visual Active Tracking using reinforcement learning remains challenging due to the absence of a unified benchmark and the complexity of open-world environments with frequent interference. To address these issues, we pioneer a systematic solution. First, we propose DAT, the first open-world drone active air-to-ground tracking benchmark. It encompasses 24 city-scale scenes, featuring targets with human-like behaviors and high-fidelity dynamics simulation. DAT also provides a digital twin tool for unlimited scene generation. Additionally, we propose a novel reinforcement learning method called GC-VAT, which aims to improve the performance of drone tracking targets in complex scenarios. Specifically, we design a Goal-Centered Reward to provide precise feedback across viewpoints to the agent, enabling it to expand perception and movement range through unrestricted perspectives. Inspired by curriculum learning, we introduce a Curriculum-Based Training strategy that progressively enhances the tracking performance in complex environments. Besides, experiments on simulator and real-world images demonstrate the superior performance of GC-VAT, achieving a Tracking Success Rate of approximately 72% on the simulator. The benchmark and code are available at https://github.com/SHWplus/DAT_Benchmark.

Open-World Drone Active Tracking with Goal-Centered Rewards

TL;DR

The paper tackles robust open-world drone active tracking by introducing the DAT benchmark and a reinforcement learning method GC-VAT. DAT provides 24 city-scale, high-fidelity scenes with unlimited scene generation via a digital twin, enabling rigorous evaluation of tracking under diverse, dynamic conditions. GC-VAT uses a Goal-Centered Reward and Curriculum-Based Training to overcome failures of prior distance-based rewards under tilted viewpoints, achieving superior tracking performance in both simulated and real-world tests, including notable sim-to-real transfer. The work advances open-world VAT by delivering a unified benchmark, provable reward design advantages, and practical training strategies with demonstrated robustness and real-world applicability.

Abstract

Drone Visual Active Tracking aims to autonomously follow a target object by controlling the motion system based on visual observations, providing a more practical solution for effective tracking in dynamic environments. However, accurate Drone Visual Active Tracking using reinforcement learning remains challenging due to the absence of a unified benchmark and the complexity of open-world environments with frequent interference. To address these issues, we pioneer a systematic solution. First, we propose DAT, the first open-world drone active air-to-ground tracking benchmark. It encompasses 24 city-scale scenes, featuring targets with human-like behaviors and high-fidelity dynamics simulation. DAT also provides a digital twin tool for unlimited scene generation. Additionally, we propose a novel reinforcement learning method called GC-VAT, which aims to improve the performance of drone tracking targets in complex scenarios. Specifically, we design a Goal-Centered Reward to provide precise feedback across viewpoints to the agent, enabling it to expand perception and movement range through unrestricted perspectives. Inspired by curriculum learning, we introduce a Curriculum-Based Training strategy that progressively enhances the tracking performance in complex environments. Besides, experiments on simulator and real-world images demonstrate the superior performance of GC-VAT, achieving a Tracking Success Rate of approximately 72% on the simulator. The benchmark and code are available at https://github.com/SHWplus/DAT_Benchmark.

Paper Structure

This paper contains 32 sections, 1 theorem, 22 equations, 17 figures, 20 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Task Definition of Drone Active Tracking
  3. DAT Benchmark with Diverse Settings
  4. Diverse Scene Construction
  5. Various Trackers and Targets Construction
  6. VAT with Reinforcement Learning
  7. MDP for Drone Active Tracking
  8. Goal-Centered Reward Design
  9. Theoretical Gurantees on Reward Design
  10. Training with Curriculum Learning
  11. Experiments
  12. Experimental Settings
  13. Comparison Experiments
  14. Ablation Experiments
  15. Experiments in Real-world Scenarios
  16. ...and 17 more sections

Key Result

Proposition 1

The commonly used Euclidean distance $d(\cdot,\cdot)$ between the target and the image center proposition does not align with the deviation $\phi(\cdot,\cdot)$ of the target from the image center projection, when the camera is not at a fixed horizontal forward viewpoint. That is: where $\phi_i=\phi(P_i,C_g)$, $P_i$ are points in the projection region $\mathcal{I}_p$, $C_g$ is the image center pro

Figures (17)

  • Figure 1: A pipeline for drone VAT.
  • Figure 2: Statistics and simulator component examples of DAT. (a) Statistics on 7 complexity aspects in DAT scenes. (b) Example scenes of DAT. (c) Diverse behaviors of targets. (d) Examples of the tracking targets. More details can be found at https://github.com/SHWplus/DAT_Benchmark.
  • Figure 3: Diagram of reward acquisition.
  • Figure 4: Reward design analysis diagram.
  • Figure 5: Reward values during training.
  • ...and 12 more figures

Theorems & Definitions (1)

  • Proposition 1