Key Frame Extraction-Based on Motion Information (Continued 1)

References traffic video segmentation and Key Frame Extraction Using Improved global K-means clustering isise 2010

And previousArticleThe idea is similar. Extract Motion Information Features and perform segmentation, and extract a certain policy for each segment parameter to extract key frames.

Motion Information Features: dissimilarity of the current frame image and background image.

The Movement Information of the current frame is measured based on the dissimilarity.

Segmentation strategy: uses a sliding window to measure the movement trend of the current frame. Based on the movement trend changes, the sliding window size is 2L.

When the FTI and the Fti-1 are different, the point can be used as a candidate segment, that is, the moving trend of the two frames is different and used as a candidate segment.

Further determining the segmentation point: Use the foreground target to enter and exit the image judgment, and select the final segmentation point from the candidate points.

1). Define the detection line. The area inside the Detection Line belongs to the detection area.

2). Measure the inbound and outbound movement information in the detection area.

3) determine the foreground target's entry and exit status based on the amount of incoming and outgoing movement information.

4). select a candidate segmentation point with an in/out status as the final candidate point.

Senter and sexit indicate the corresponding inbound and outbound status.

Key Frame selection policy: clustering method, improved global K-means clustering

Select the number of candidate segmentation points in each segment as the K value of clustering. For details, refer to the original article.

Segmentation aggregation: if excessive segmentation exists, segmentation merging is required.

When the number of frames stored in a segment is less than Tn = 50, the current segment and connected segments need to be aggregated.

Calculate the similarity between the current segment and the first and second video segments. The similarity is measured by the similarity between the current frame and the second frame.

Merge a segment with a large similarity until excessive segmentation does not exist.

References:

1). Use of Sliding Window and segment point selection strategy.

2) how to set the Virtual Detection Line and determine the moving status of the foreground target

(The method mentioned in this article is not necessarily reliable, but provides the idea of judging the moving status of the foreground target)

3) solve the idea of excessive segmentation.