Real time objects counting

Object counting excercise. The aim is to use a low-capability camera to perform an efficient real time objects counting. It counts static objects placed in the camera FOV. This kind of problem could be solved on a variety of different devices.

Summary here: Presentation

Usage

python countobj.py [mode][fidelity][fidelityValue]

• mode: video (default) | image | h-help | info
• fidelity: activate fidelity range
• fidelityValue: [0.0, 1.0] default 0.7
• Info for authors and disclaimer

Approach

The main idea is to count contours in captured frames:

• Pro:

  • Probably a contour represents an object;
  • Efficiency;
  • A lot of known edge detectors;

• Cons:

  • Contours could be noise;
  • Complex objects can have multiple and/or inner contours;

Methodology

Frame preprocessing

• A 15x15 median filter [1] is applied to reduce noise, smoothing the color reducing light/shadow anomaly and similar-color differences. This technique preserves edges so the result is not altered;
• Frame is converted in a grayscale image and Otsu thresholding is applied;
• Morphological OPEN is used to reduce noise, remove little objects and strengthen the others;

Detect static scene

• To avoid false counting, algorithm detects movements applying a background subtraction based on Mixture of Gaussians [2]. If movement is recognized as a high pixel density in the MOG result, counting is stopped at the moment because it means that we are placing an object in the scene;

Edge detection

• Edges are detected using a Topological Structure Analysis by Border Following [3]. This technique finds a hierarchy of edges for an object;
• We used hierarchy information to get only root contours (contours with no parents);
• Each root element is an object that increases the counter;
• At the end we display the contour bounding box;

Results

It is a fast object counting algorithm suitable for real time applications. For its prototype nature it doesn’t handle light problems;

For preprocessing we tried 4 kinds of filters:

1. Normalized Box Filter: doesn’t preserve edges;
2. Median filter: the winner;
3. Bilateral filter [4]: better than median but too slow;
4. Adaptive bilateral filter [5]: as above;

Demo video

Usage on Images

It is also provided an implementation of the algorithm for images. In this case there are two choices:

1. Fidelity range: if objects have a similar size/area it is helpful to prevent false detection;
2. Normal: no area thresholding is applied;

References

[1] Median Filter: https://en.wikipedia.org/wiki/Median_filter;

[2] Mixture of Gaussians: http://www.ai.mit.edu/projects/vsam/Publications/stauffer_cvpr98_track.pdf;

[3] Suzuki, S. and Abe, K., Topological Structural Analysis of Digitized Binary Images by Border Following. CVGIP 30 1, pp 32-46 (1985);

[4] C. Tomasi and R. Manduchi, "Bilateral Filtering for Gray and Color Images", Proceedings of the 1998 IEEE International Conference on Computer Vision, Bombay, India;

[5] Buyue Zhang; Allebach, J.P., "Adaptive Bilateral Filter for Sharpness Enhancement and Noise Removal," Image Processing, IEEE Transactions on , vol.17, no.5, pp.664,678, May 2008;