Hierarchical Data Structure for Real-Time Background Subtraction
One of the drawbacks of background subtraction methods for motion detection is a slow processing time. Consequently, real-time motion detection at 30 frames per second with an image size of 480 by 640 is not possible with these methods. In an effort to solve this problem, a quadtree-based hierarchical framework was developed to sample a small percentage of the image pixels in order to decrease processing time. The results of using our method for motion detection has quality as high as that of the original algorithm. Another benefit of our method is that it can be applied to most existing background subtraction algorithms with little modification.
The following images show some of the results obtained using our method. The method was tested on four sequences: Seq1 and Seq2 are from the MUSCLE project (European Union MUSCLE Network of Excellence FP6-507752, http://muscle.prip.tuwien.ac.at/ ), Seq3 is an outdoor scene with moving cars and trees, while Seq4 is a scene of people moving in a laboratory.
The following image progression briefly describes our method:
The images below show some experimental results comparing our method with some established methods. RA is running average, RA with tree is running average with our hierarchical structure, MOG stands for Mixture of Gaussians (Stauffer and Grimson, 1999 and 2000), and MOG with tree is the MOG technique tested with our hierarchical data strucuture.
• Johnny Park
Results are offered in two video formats: AVI and WMV (Windows Media Viewer). The AVI format is the best quality, though some may not be able to view this format. If you use the WMV format, make sure that you have Windows Media Player version 10 installed.
Results using the Running Average algorithm and the hierarchical data structure:
Seq1
MUSCLE project (European Union MUSCLE Network of Excellence (FP6-507752),
Recorded at 30 fps, video play speed is 15 fps
Seq2
MUSCLE project (European Union MUSCLE Network of Excellence (FP6-507752),
Recorded at 30 fps, video play speed is 15 fps
Seq3, Garage scene
Recorded at 30 fps, video play speed is 15 fps
Seq4, High Resolution camera, Lab scene
Recorded at 15 fps, video play speed is 10 fps
Results using the Mixture of Gaussians algorithm and the hierarchical data structure
Seq1
MUSCLE project (European Union MUSCLE Network of Excellence (FP6-507752),
Recorded at 30 fps, video play speed is 15 fps
Seq2
MUSCLE project (European Union MUSCLE Network of Excellence (FP6-507752),
Recorded at 30 fps, video play speed is 15 fps
Seq3, Garage scene
Recorded at 30 fps, video play speed is 15 fps
Seq4, High Resolution camera, Lab scene
Recorded at 15 fps, video play speed is 10 fps
J. Park, A. Tabb, and A. C. Kak, "Hierarchical Data Structure for Real-Time Background Subtraction." IEEE International Conference on Image Processing, 2006. [pdf]
