Abstract
Detecting camouflaged moving foreground objects has been known to be difficult due to the similarity between the foreground objects and the background. Conventional methods cannot distinguish the foreground from background due to the small differences between them and thus suffer from under-detection of the camouflaged foreground objects. In this paper, we present a fusion framework to address this problem in the wavelet domain. We first show that the small differences in the image domain can be highlighted in certain wavelet bands. Then the likelihood of each wavelet coefficient being foreground is estimated by formulating foreground and background models for each wavelet band. The proposed framework effectively aggregates the likelihoods from different wavelet bands based on the characteristics of the wavelet transform. Experimental results demonstrated that the proposed method significantly outperformed existing methods in detecting camouflaged foreground objects. Specifically, the average F-measure for the proposed algorithm was 0.87, compared to 0.71 to 0.8 for the other state-of-the-art methods.
Introduction
Moving object detection is a fundamental task in computer vision, typically addressed through background subtraction. This process involves building a statistical model of the static scene and identifying deviations as foreground. However, this approach frequently fails in "camouflaged" scenarios, where the foreground object shares nearly identical intensity and hue with the background. In such cases, pixel-level differences are often indistinguishable from sensor noise, leading standard models like Gaussian Mixture Models (GMM) or Local Binary Patterns (LBP) to under-detect the moving object.
This paper introduces FWFC (Fusion in the Wavelet domain for Foreground detection in Camouflaged scenes), a framework designed to amplify subtle texture differences invisible in the standard color domain. By shifting the analysis to the wavelet domain, the system can isolate high-frequency signatures that differentiate a foreground object from its background, even when their colors match.
Methodology: The FWFC Framework
The core methodology relies on the observation that while a camouflaged object may match the background's color, it rarely matches its fine-grained texture across all scales. The FWFC framework processes video frames through several distinct stages:
- Stationary Wavelet Transform (SWT): To preserve spatial resolution, the framework utilizes a non-decimated SWT. This ensures that every wavelet coefficient maps directly back to a specific pixel, avoiding the downsampling artifacts of the standard Discrete Wavelet Transform.
- Multi-Level Decomposition: Images are decomposed into six levels, each containing four bands: LL (low-frequency/intensity), LH (horizontal details), HL (vertical details), and HH (diagonal details).
- Band-Specific Modeling: Rather than modeling the image as a whole, the system maintains a separate GMM for every coefficient in every band. This allows for precise likelihood estimation based on the historical behavior of specific frequency components.
- Two-Stage Fusion:
- Level Fusion: Likelihoods are aggregated across the six levels for each band type using translation weights (accounting for spatial representation) and noise-induced weights (to suppress sensor noise).
- Band Fusion: The final stage combines results from intensity-focused LL bands with texture-focused high-frequency bands (LH, HL, HH) to produce a robust binary foreground decision.
Experimental Evaluation
The authors evaluated the FWFC framework using both existing benchmarks and a newly introduced dataset:
- Existing Benchmarks: The synthetic "Camouflage" dataset and the real-world SBM-RGBD dataset (utilizing RGB channels).
- CAMO_UOW Dataset: A new collection of 10 videos featuring real-world camouflaged scenes, such as individuals wearing clothes that match indoor or outdoor backgrounds. This dataset includes manually labeled ground truth for every frame, providing a rigorous test for detection accuracy.
Results and Discussion
The proposed FWFC method was compared against eight state-of-the-art algorithms, including MOG2, SuBSENSE, and DECOLOR.
Quantitative Performance
FWFC achieved an average F-measure of 0.87, significantly higher than the 0.71 to 0.80 range achieved by competing methods. The framework demonstrated particularly high recall, successfully identifying camouflaged body parts that other methods missed entirely.
Qualitative Analysis
As shown in the experimental results, the wavelet-based approach effectively handles the "aperture problem" by using multi-level information to fill in the centers of objects that edge-based detectors often leave hollow. While the method is computationally more intensive than basic GMMs (processing a 640x384 frame in 0.275s), it remains significantly faster than optimization-based methods like DECOLOR (4.99s), suggesting potential for near-real-time application with GPU acceleration.
Conclusion
The FWFC framework demonstrates that the wavelet domain is a powerful representation for detecting camouflaged moving objects. By modeling background statistics directly in this domain and employing a two-stage fusion process, the method effectively amplifies subtle textural differences. Future work may involve integrating depth information from RGB-D sensors or exploring deep learning architectures that incorporate multi-scale wavelet fusion to further enhance robustness in dynamic or texture-less environments.
Summary updated