Background Model-color feature and texture feature fusion (Continued 3)

References multi-layer background subtraction based on color and texture CVPR-VS 2007

And the previous articleArticleThe general idea is the same. texture features and color features are extracted, the background model is established, and the background model is updated in real time.

Texture features.

Color Features: The Distribution Model of color space is used for reference. The angle between the current pixel and the background pixel and the minimum and maximum values are used as the color features.

Texture Feature similarity calculation:

Color Feature similarity calculation:

The final similarity calculation is the weighted sum of texture features and color feature similarity.

Background Model Description:

The background update is similar to the previous article, except that the learning factor of its weight update is adjusted based on the maximum weight of the model.

If the current weight is small, but the model that once had the power value is matched, the weight increase is very fast.

If the current weight value is small, but a model with a previous power value does not match, the reduction of the weight is very slow.

A model that once had a powerful value may actually belong to the background model,

It is reasonable to quickly restore the weights of pixels that match these models,

It is reasonable to slowly reduce the weight of a model that does not match this type.

When a matching model exists, the corresponding K model update formula is as follows:

For other models, only their weights are updated, while others remain unchanged,

Steps for foreground Detection:

1) Calculate the minimum distance between K models and the background model. The model that records the minimum distance is the I model.

2) determine the I-th model, whether it is a background model, and whether it is a trusted background model.

3) If yes, the minimum distance remains unchanged. If not, the minimum distance is updated.

4). Use the cross bilateral filter to smooth the minimum distance value.

5) Determine whether the corresponding pixel is foreground or background based on the minimum distance value and threshold value.

The formula is as follows:

Innovation in this article:

1). Color Feature selection.

2). The update method of the weight value uses the historical information corresponding to the weight value.

3) introduce the concept of layering and use the historical information of the model to be matched as a trusted background model to determine the credibility of the current model match.

4) perform smooth processing after calculating the minimum distance value, instead of obtaining the foreground and background.

Weaknesses: The Impact of parameter selection and Threshold Selection on detection results and how to select the most appropriate set of parameter values.

 

Multi-layer differential Moving Target Detection Based on color and texture feature Background ModelAlgorithmComputer Application 2009

When I first read this article, I feel that there are many similarities with the above English article. However, there are still some differences.

1) in this paper, texture features and color features are matched separately to obtain the results under the two features, and then the two results are fused.

2) in this paper, texture feature matching and updating are the same as those of the Gaussian mixture model. The variance is used to adjust the matching threshold.