Efficient background subtraction for tracking in embedded camera networks

Yiran Shen, Wen Hu, Mingrui Yang, Junbin Liu, Chun Tung Chou2012 · DOI

TL;DR

This paper tackles the challenge of real-time object tracking in resource-constrained embedded camera networks by proposing a computationally efficient background subtraction method. They achieve this by integrating Compressive Sensing (CS) with the established Mixture of Gaussians (MoG) model, significantly reducing computational load while maintaining comparable accuracy.

What Problem Does This Paper Address?

Real-time object tracking is crucial for many computer vision applications, but it often demands substantial computational power and energy. This poses a significant hurdle for embedded camera networks, which have limited processing capabilities and battery life. Traditional background subtraction techniques, while robust, are frequently too computationally intensive for these platforms. The authors aim to bridge this gap by developing a background subtraction method that is both accurate and highly efficient, making it suitable for deployment on embedded systems.

Methods

The core innovation of this paper is the integration of Compressive Sensing (CS) into the Mixture of Gaussians (MoG) background subtraction model, creating a new method called CS-MoG.

Here's how it works:

  1. Image Blocking and Projection: The input image is divided into small blocks (e.g., 8x8 pixels). For each block, a random projection vector (significantly smaller than the original block size, e.g., 8 projections for a 64-pixel block) is generated. This process effectively reduces the dimensionality of the data, as represented by the equation $Y = \Phi X$, where $X$ is the vectorized block, $\Phi$ is a random projection matrix, and $Y$ is the compressed projection vector.
  2. MoG Modeling on Projections: Instead of modeling the original pixel values, CS-MoG models each of these compressed projection vectors using a MoG. This means the background model is built on these lower-dimensional representations.
  3. Foreground Detection and Fusion: A pixel block is flagged as foreground if its projection does not fit the learned background MoG model. The authors explore different fusion strategies (max, min, majority voting) to make a final decision for the block, finding that "min voting" yields the best results.

The key novelty lies in applying CS to reduce the dimensionality of image blocks before applying the MoG model, thereby drastically cutting down on the computations required for background modeling and subtraction.

Datasets

The authors evaluated their method on a custom outdoor pedestrian tracking dataset consisting of 400 images with a resolution of 320x240 pixels. No new benchmarks were introduced; the focus was on evaluating their method against existing techniques on this specific dataset.

Key Insights and Ideas

  • Compressive Sensing for Efficiency: The core insight is that CS can be effectively leveraged to reduce the computational burden of background subtraction without a significant loss in accuracy. By compressing the data before modeling, the complexity of the MoG updates and comparisons is dramatically reduced.
  • MoG Robustness Preserved: The CS-MoG method successfully retains the robustness of the original MoG algorithm, which is known for handling complex backgrounds and gradual illumination changes.
  • Block-based Processing: Dividing the image into blocks and applying CS to each block allows for parallel processing and further enhances efficiency.
  • Min Voting for Fusion: The finding that "min voting" is the optimal fusion strategy for determining foreground blocks is a specific, actionable insight for implementing CS-MoG.

Key Results

  • Computational Efficiency: CS-MoG achieved a significant speedup, being approximately 7 times faster than the original MoG method when implemented on a Blackfin DSP camera node. Specifically, it reduced processing time from 250-280ms to 37-42ms per 320x240 image.
  • Accuracy: The paper demonstrates that CS-MoG achieves comparable accuracy to the original MoG. In terms of Probability of Detection (PD) vs. False Alarm (FA), CS-MoG achieved over 98% PD with a FA below 3%, outperforming a CS-based subtraction method (CSBS) which achieved only about 75% PD under the same FA.
  • Projection Ratio: The performance of CS-MoG plateaus when the number of projections exceeds 8 per block (for an 8x8 block), suggesting an optimal trade-off between compression and accuracy.

Limitations and Future Work

  • Limitations: The paper acknowledges that the current CS-MoG method might not directly yield pixel-scale results from compressed projections, especially for blocks containing both foreground and background.
  • Future Work: The authors plan to:
  • Develop methods to directly obtain pixel-scale results from compressed projections, particularly for mixed blocks.
  • Evaluate the proposed CS-MoG method on a wider variety of datasets.
  • Implement the complete tracking system using CS-MoG on an embedded camera network.

Summary updated

Abstract

Background subtraction is often the first step in many computer vision applications such as object localisation and tracking. It aims to segment out moving parts of a scene that represent object of interests. In the field of computer vision, researchers have dedicated their efforts to improve the robustness and accuracy of such segmentations but most of their methods are computationally intensive, making them non-viable options for our targeted embedded camera platform whose energy and processing power is significantly more constrained. To address this problem as well as maintain an acceptable level of performance, we introduce Compressive Sensing (CS) to the widely used Mixture of Gaussian to create a new background subtraction method. The results show that our method not only can decrease the computation significantly (a factor of 7 in a DSP setting) but remains comparably accurate.
BibTeX
@misc{shen2012efficient,
  author = {Yiran Shen and Wen Hu and Mingrui Yang and Junbin Liu and Chun Tung Chou},
  title = {{Efficient background subtraction for tracking in embedded camera networks}},
  year = {2012},
  doi = {10.1145/2185677.2185698},
}

Loading related content...

Figure