Background Modeling-color feature and texture feature fusion (Continued 2)

References: background subtraction based on a combination of texture, color and intensity icsp 2008

ArticleThe background model is established based on features such as texture, color, and intensity, and the background is updated in real time. In a complex background environment, the model has a good detection effect.

Compared with the previous two articles, this article does not combine multiple features into one. Instead, it creates a background model description based on multiple features and updates the background description based on matching conditions.

Extracted features:

The texture feature is dlbp.

The color feature is HS = {H, s }.

Let HS = {H, s} denotes the color information for Pixel X,

Where H is a local3 × 3 Template centered at the pixel X for the hue channel,

S a local 3 × 3 template for the saturation channel.

The intensity feature refers to the intensity value corresponding to the pixel, that is, the gray value of the pixel.

Similarity calculation:

Dlbp similarity calculation:

Color Feature similarity calculation:

Strength Feature similarity calculation:

Background Model Description:

The background model update is similar to the Gaussian mixture model, except that the distribution is Gaussian.

Based on texture features, the I-th model with the smallest distance is calculated, and the distance between the current frame and the I-th model, including the texture, color, and intensity features, is calculated,

If the distance between the three features is less than a certain threshold value, it indicates that the distance from the third model matches with the I model, which is the background. Otherwise, it does not match and is the foreground.

If yes, update the I-th model and other models. If no match exists, add a new model or replace the model with the minimum weight.

The matching formula and update formula are as follows:

Thoughts:

1) Why do we select the model with the smallest Texture Feature Distance Based on texture features for model matching calculation.

Why is it not based on color features?

Whether texture features, color features, and intensity features can be combined into a feature vector, and then the distance of the feature vector is calculated to select the model with the smallest distance for matching calculation.

2) This method involves the selection of many parameter values, the selection of threshold values, and the selection of learning factor sizes are not adaptive. It is obtained through experiment and empirical observation.

Can we improve the threshold value selection and make it adaptive? similar to the variance of Gaussian Model, the threshold value is adjusted based on the Variance Change.