Kernel Similarity Modeling of Texture Pattern Flow for Motion Detection in Complex Background

2011 IEEE transection on CSVT

Baochang Zhang, Yongsheng Gao, Sanqiang Zhao, Bineng Zhong

Outline TPF Operator Kernel Similarity Modeling Experiment Result Conclusion

TPF Operator-Spatial Given a gray-scale image sequence To capture the spatial variations in x and

y directions, two threshold functions, and , are employed to encode the gradient information into binary representations

TPF Operator-Temporal The temporal derivative is defined as

A pixel value lying within 2.5 standard deviations of a distribution is defined as a match

match

TPF Operator By integrating both spatial and temporal

information, the TPF is defined as

TPF reveals the relationship between derivative directions in both spatial and temporal domains

Flowchart for one pixel

Integral Histogram

Integral Histogram of TPF

Using a neighborhood region provides certain robustness against noise

When the local region is too large, the more details will be lost

Building Background Model Use GMM to model the background If a match has been found for the pixel,

update mean and variance of the matched Gaussian distribution

If none of the K Gaussian distributions match the current pixel value, the least probable distribution is replaced with a new distribution whose mean is the current pixel value

Kernel Similarity Measurement

We use k to represent the result of kernel similarity

With the information of kernel similarity, we can get an adaptive threshold to classify the input pixel

: mean of the th Gaussian distribution at time t: variance of the th Gaussian distribution at time t : model integral histogram : learning rate

Update the Background Model

If the pixel is labeled as background, the background model histogram with the highest similarity value will be updated with the new data

: input integral histogram : 1 for the best-matched distribution, 0 for the other distributions

Experiment Results All the experiments in this paper are

conducted on gray-level values For simplicity, 3 Gaussian distributions

and 3 model integral histograms are used to describe all the Gaussian mixture models

= 0.7, = 0.01

Experiment 1

Experiment 2Wallflower video(a)GMM(b)CMU(c) LBP(d)TPF(e)KSM-TPF

Experiment 2

GMM CMU LBP TPF KSM-TPF

Conclusion KSM-TPF is much more robust to

significant background variations However, it is less computationally

efficient than the GMM method or LBP method