TensorMoG: A Tensor-Driven Gaussian Mixture Model with Dynamic Scene Adaptation for Background Modelling

Synh Viet-Uyen Ha, Nhat Minh Chung, Hung Ngoc Phan, Cuong Tien Nguyen2020 · DOI

TL;DR

This paper introduces TensorMoG, a novel tensor-based Gaussian Mixture Model (GMM) for background modeling in surveillance systems. It addresses the challenge of sudden scene changes and illumination variations by adaptively updating the background model, achieving a good balance between accuracy and processing speed.

What Problem Does This Paper Address?

Background modeling is a crucial task in intelligent surveillance systems for identifying "foreground" objects of interest by separating them from the static "background." Traditional GMM-based approaches, while successful, struggle with dynamic scenes that experience sudden changes like illumination shifts or camera jitter. These variations can lead to erroneous background models and inaccurate foreground detection. Existing deep learning methods often require extensive pretraining and are computationally expensive, making them less practical for real-time applications. This paper aims to bridge this gap by proposing an unsupervised, adaptive, and computationally efficient background modeling technique.

Methods

TensorMoG utilizes a tensor-based framework to represent GMM parameters (mean, variance, weight) in a parallelizable manner, allowing for efficient pixel-wise statistical learning. The core innovation lies in its dynamic scene adaptation mechanism. It employs an adaptive learning threshold based on Shannon entropy to estimate the uncertainty of the background model. This entropy estimation signals when the model is likely to be corrupted by scene changes. Based on this, TensorMoG operates in two states:

  • Updating State: The primary state where the model actively updates its parameters.
  • Suspending State: Activated during significant scene changes, where the model suspends primary updates and uses implicit updates to adapt, preventing distortions.

This selective update strategy allows the model to ignore transient, irrelevant variations while retaining important background components.

Datasets

The authors evaluated TensorMoG on several publicly available benchmarking datasets:

  • CDnet 2014: A large-scale dataset for motion detection across various scenarios.
  • CAMO-UOW: Designed for evaluating camouflaged object segmentation.
  • SBMnet: Focuses on modeling background scenes.
  • SBI: Features sequences where background images are occluded by foreground objects.

Key Insights and Ideas

  • Tensor-based parallelism: Representing GMM parameters as tensors allows for efficient parallel processing, significantly speeding up computations, especially on GPUs.
  • Entropy-driven adaptation: Using Shannon entropy to quantify uncertainty provides a robust mechanism for detecting and responding to dynamic scene changes.
  • Dual-state model: The "Updating" and "Suspending" states enable a more intelligent adaptation strategy, balancing responsiveness to changes with stability.
  • Selective updates: The ability to selectively update or discard incoming frames based on entropy estimations prevents model corruption from anomalies.

Key Results

TensorMoG demonstrated strong performance across various datasets:

  • CDnet 2014: Achieved top-1 performance among unsupervised approaches and was within the top-3 for eight out of eleven scenarios in terms of F-measure. It also showed high precision and low error rates.
  • CAMO-UOW: Ranked third best in F-measure, comparable to leading methods, and showed good performance in handling camouflaging effects.
  • SBI and SBMnet: Showed competitive results, particularly in MS-SSIM, indicating good perceived visual quality of the generated backgrounds.
  • Speed: TensorMoG achieved a processing speed of 302.5261 frames-per-second, making it one of the fastest methods, second only to GMM-Zivkovic. This speed advantage is crucial for real-time applications.

Limitations and Future Work

The authors acknowledge that TensorMoG struggles in scenarios with:

  • Lack of information for background construction.
  • High degrees of motion noises.
  • Spatial translations in imaging signals.

For future work, they propose incorporating convolutional supervised training for spatio-temporal smoothing and filtering within the tensor-driven framework to address these limitations.

Summary updated

BibTeX
@misc{ha2020tensormog,
  author = {Synh Viet-Uyen Ha and Nhat Minh Chung and Hung Ngoc Phan and Cuong Tien Nguyen},
  title = {{TensorMoG: A Tensor-Driven Gaussian Mixture Model with Dynamic Scene Adaptation for Background Modelling}},
  year = {2020},
  doi = {10.3390/s20236973},
}
Figure