Motion Detection Algorithm for Unmanned Aerial Vehicle Nighttime Surveillance

Huaxin Xiao, Yu Liu, Wei Wang, Maojun Zhang

Introduction

Unmanned Aerial Vehicles (UAVs) have become indispensable tools for both military and civil surveillance. However, their effectiveness is often limited by the transition from day to night. While high-resolution photoelectric cameras are preferred for their detail, low-light conditions introduce significant electronic noise from CMOS and CCD sensors. Traditional motion detection algorithms, such as background subtraction or Mixture of Gaussians (MOG), typically assume a clean video feed. In noisy nighttime environments, these methods frequently fail, mistaking random sensor fluctuations for actual movement and producing unusable "salt-and-pepper" detection masks.

Methodology: Sparse Representation and Universal Dictionaries

To address the challenge of heavy sensor noise, this research proposes a motion detection framework based on Sparse Representation. Instead of analyzing individual pixels, the algorithm represents image patches as a combination of basic structural "atoms" drawn from a pre-learned Dictionary (D).

The core innovation lies in the transition from scene-specific dictionaries to a Universal Dictionary. By training the dictionary on 20,000 to 30,000 samples from diverse environments, the system develops a robust "vocabulary" capable of accurately reconstructing almost any background structure.

Comparison of dictionary learning strategies: (a) learning from a single scene versus (b) learning a universal dictionary from diverse scenes.

Foreground Detection

The algorithm models the background as $I_B = D \times A$, where $A$ represents sparse coefficients. When a new frame arrives, the system calculates updated coefficients ($A'$). Because electronic noise is inherently random, it does not align with the structured atoms in the dictionary. In contrast, moving objects create distinct, non-random shifts in the sparse coefficients. By comparing the distribution of these coefficients using a custom difference function, the algorithm effectively filters out noise while preserving the signal of moving targets.

Experimental Setup and Datasets

The proposed algorithm was rigorously tested using a combination of real-world and synthetic data to ensure robustness across different scenarios.

  • Realistic UAV Video: Captured from an actual drone and stabilized to a consistent shooting plane.
  • Synthetic Video: Generated using an OpenGL environment to provide precise ground truth data with fixed viewpoints.
  • Noise Simulation: To replicate harsh nighttime conditions, researchers added a mixture of Gaussian and Poisson noise to the source videos, creating a high-interference testbed.
Sample frames from the evaluation datasets, including real-world UAV captures and synthetic OpenGL environments used for ground truth testing.

Key Results and Discussion

The algorithm was benchmarked against six established methods, including MOG, ViBe, and DECOLOR. While competing methods produced masks heavily contaminated by noise—often making the target invisible—the proposed sparse representation method yielded clean binary masks with near-zero noise infection.

Two primary insights emerged from the results:

  1. Structural Advantage: Sparse representation exploits the fact that while noise may have high amplitude, it lacks the spatial structure of physical objects.
  2. Dictionary Versatility: A dictionary trained on diverse scenes (Dictionary 2) proved significantly more effective at background reconstruction than one limited to a single scene, confirming the value of the "universal" approach.

Conclusion and Future Work

This research demonstrates that sparse representation is a powerful tool for UAV surveillance in low-light environments, effectively turning the detection process into a simultaneous denoising task. While the mathematical framework is robust, future work will focus on reducing computational overhead to enable real-time onboard processing and improving performance in dynamic backgrounds, such as those containing swaying vegetation.

Summary updated

BibTeX
@article{xiao2014motion,
  author = {Huaxin Xiao and Yu Liu and Wei Wang and Maojun Zhang},
  title = {{Motion Detection Algorithm for Unmanned Aerial Vehicle Nighttime Surveillance}},
  journal = {IEICE Trans. Information and Systems},
  year = {2014},
  doi = {10.1587/transinf.2014edl8129},
}
Figure