Abstract
For background modeling, the conventional Gaussian Mixture Model (GMM) is a popular approach. However, because of the inappropriate parameters updating method, GMM often suffers from a problem that it cannot classify a pixel into background or foreground correctly for longtime. In the paper, we proposed a new parameters updating method for GMM, and built background model for every pixel and global foreground model for the entire image. We presented an improved object detection and tracking scheme based on the proposed approach. The experimental results show the proposed GMM parameters updating method, together with the object detection and tracking framework, give better performance than the conventional Gaussian Mixture Modeling algorithm.
Introduction
In computer vision, background subtraction is a fundamental technique for identifying moving objects within a scene. By constructing a model of the static background, systems can isolate foreground elements that deviate from this model. Gaussian Mixture Models (GMMs) have become the industry standard for this task due to their ability to handle multi-modal backgrounds, such as foliage moving in the wind or flickering lights.
Despite their popularity, conventional GMMs frequently encounter "model drift." This occurs when simplified parameter update rules, often chosen for computational speed, lead to cumulative errors. Over time, the model loses its ability to distinguish between background and foreground accurately, especially in complex or long-running sequences. This paper addresses these limitations by introducing a more rigorous statistical framework for parameter updates and a dual-model architecture.
Methodology
The proposed approach enhances GMM stability through two primary innovations: Maximum Likelihood (ML) parameter updating and a dual-model architecture.
Maximum Likelihood Parameter Updating
Rather than relying on standard online heuristic approximations, this method employs a mathematically robust ML approach. The update process involves a two-step optimization:
- Probability Calculation: Determining the probability ($\gamma_k$) that a specific pixel belongs to a particular Gaussian component.
- Parameter Re-estimation: Using these probabilities to re-estimate the mean ($\mu_k$) and variance ($\Sigma_k$) of the components.
Dual-Model Architecture
While traditional systems focus solely on local background modeling, this research maintains a background model for every pixel alongside a global foreground model for the entire image.
As illustrated in the flowchart above, when a new pixel arrives, it is first evaluated against the local background model. If it does not match with high probability, it is checked against the global foreground model. This "second opinion" helps the system differentiate between known moving objects and novel changes in the background that necessitate a model update.
Experimental Results
The algorithm was evaluated using the PETS2001 dataset, a benchmark for surveillance applications, with image resolutions of 768×576 pixels. The system was implemented in MATLAB 2013a on standard hardware (Pentium IV, 4G RAM).
The experimental results demonstrate that the proposed ML-based approach provides superior tracking stability compared to conventional GMMs. As seen in the tracking sequence, the algorithm maintains precise bounding boxes throughout the duration of the object's movement. By modeling the foreground globally, the system leverages characteristics shared by moving objects across the entire frame, increasing confidence in identification as objects move through the scene.
Conclusion
This paper presents an improved GMM framework that replaces heuristic update rules with rigorous Maximum Likelihood estimation and introduces a global foreground model. While the ML update process is more computationally intensive than conventional methods, it significantly reduces model drift and improves long-term tracking precision. Future work will focus on optimizing the parameter update process to enhance computational efficiency without compromising accuracy.
Summary updated