Adaptive Background Mixture Models for Real-Time Tracking(1999) by Chris Stauffer, W. Eric L. Grimson
Background Subtraction Using Markov Thresholds(2005) by Joshua Migdal, W. Eric L. Grimson
In section 4.5(Experimental setup), they have mentioned the thresholds used.
K = 2, alpha = 0.005, rho = 0.05, T = 0.5
They could have mentioned in 1999 paper also.
Foreground-Adaptive Background Subtraction(2009) by J Mike McHugh et al
paper link in Konrad's page
Master Thesis explains in detail.
Statistical Background Subtraction for a Mobile Observer (2003) by Eric Hayman and Eklundh
In Section 2.1(review of Stauffer and Grimson's algorithm), they mention that the pixel value in current frame is assigned to only one background model. But I think in the Stauffer's algorithm, there's no such constraint. Whichever model(s) satisfies the constraint of having the pixel within 2.5 sigma, its considered as matched model.
In their method, Hayman & Eklundh, for the first few frames, they consider only a single background model. Probably that's the reason why they thought of Grimson's method as like that.
Note: alpha stands for weight in this paper, unlike in Stauffer & Grimson's paper where it was learning rate. Beware of this confusion.
A statistical Approach for real-time Robust Background Subtraction and Shadow Detection (1999) by T. Horprasert et al
1. test_background_subtraction_pixel.m -> implementation of Stauffer & Grimson'e 1999 paper
2. test_background_subtraction_frame.m -> implementation of Hayman & Eklundh's 2003 paper [[ Having some doubts in the implementation ]]
Test video: 1st video of http://cvrc.ece.utexas.edu/SDHA2010/videos/human_interaction.zip