References summarisation of surveillance videos by key-frame selection icdsc 2011
The idea in this article is roughly the same as that in the previous two articles. This article does not take this step in segments, but directly selects key frames based on a certain selection policy.
Feature Extraction: EHD edge histogram descriptor, LFE localised foreground entropy.
The EHD feature extraction procedure is as follows:
1). Convert the image into a grayscale image and divide it into 4x4 image blocks.
2) calculate each image sub-block and correspond to MF (I, j) in five directions ).
3). Combine the MF (I, j) in each direction of each graph sub-block into a vector (5*4*4) 80.
For more information, see efficient use of local edge histogram descriptor.
The LFE feature extraction procedure is as follows:
1) divide the foreground result image into N x n image blocks.
2) Calculate the entropy of each image sub-block.
3) Calculate the Mean Entropy of N x n sub-blocks to obtain the entropy of the final image.
M = N * n.
Select the local maximum value corresponding to EHD and LFE as the key frame.
This processing method is iterative. During each iteration, key frames with a smaller feature value are deleted for each pair of connected key frames until a certain number of key frames are selected, stop iteration.
Thoughts:
1). Whether EHD is used to calculate the original image sequence or the foreground result image sequence of the detection. (You need to read the article efficient use of local edge histogram descriptor)
2). The block division in LFE is even, and the weight of each image sub-block is the same when the mean value of the image sub-block entropy is obtained.
Without considering the video features of a person, the person may be more interested in the area of the image center. The weight of the content of the image center should be greater.
You can use unequal partitioning policies and assign different weights to each block.
3). The key frame selection policy is too simple and the selection of key frames is not necessarily accurate.