System and method for processing a video stream to extract highlights
With the widespread availability of video cameras, we are facing an ever-growing enormous collection of unedited and unstructured video data. Due to lack of an automatic way to generate highlights from this large collection of video streams, these videos can be tedious and time consuming to index or search. The present invention is a novel method of online video highlighting, a principled way of generating a short video highlight summarizing the most important and interesting contents of a potentially very long video, which is costly both time-wise and financially for manual processing. Specifically, the method learns a dictionary from given video using group sparse coding, and updates atoms in the dictionary on-the-fly. A highlight of the given video is then generated by combining segments that cannot be sparsely reconstructed using the learned dictionary. The online fashion of the method enables it to process arbitrarily long videos and starts generating highlights before seeing the end of the video, both attractive characteristics for practical applications.
1. A system for extracting video highlights from a video stream comprising:
storage, containing said video stream in digital form; and
a processor, in communication with said storage, said processor executing software for performing the following functions:
i. reading an initial plurality of frames from said video stream and constructing a new dictionary based on said initial plurality of frames, said dictionary containing data summarizing the content of said initial plurality of frames;
ii. breaking the remainder of said video stream into a plurality of segments of approximately equal length;
iii. reading the next segment from said video stream;
iv. detecting interest points in said segment and defining a local spatio-temporal cuboid for each of said detected interest points;
v. calculating, for each of said spatio-temporal cuboids, a plurality of feature vectors and concatenating said plurality of feature vectors into a feature representation for each of said spatio-temporal cuboids;
vi. determining if a group of feature representations representing all of said spatio-temporal cuboids in said segment can be sparsely reconstructed for a cost not exceeding a predetermined threshold, using existing entries in said dictionary, and, if not, adding said group of feature representations to said dictionary and adding said segment to said video highlights;
vii. repeating steps iii.-vi. until all segments in said video stream have been processed.
2. The system of claim 1 wherein each of said feature representations in said group of feature representations comprise a concatenation of at least two feature vectors.
3. The system of claim 2 wherein said at least two feature vectors includes a histogram of gradient and a histogram of optical flow.
4. The system of claim 1 wherein said number of detected interest points, and thus the number of feature vectors in said group of feature vectors, varies from segment to segment.
5. The system of claim 1 wherein a segment is added to said video highlights when the cost to perform said sparse reconstruction exceeds a predetermined threshold.
6. The system of claim 5 wherein the length of said video highlights can be varied by adjusting said predetermined cost threshold.
7. The system of claim 1 wherein a video segment is excluded from said video highlights when the cost to perform said sparse reconstruction is below a predetermined threshold.
8. The system of claim 1 wherein said video highlights contain said initial plurality of frames.
9. The system of claim 1 wherein said dictionary contains a matrix wherein the columns of said matrix are atoms containing groups of feature representations and wherein the rows of said matrix contain feature vectors.
10. The system of claim 9 wherein the number of atoms in said dictionary is fixed.
11. The system of claim 9 wherein said sparse reconstruction attempts to minimize the cost to reconstruct a group of feature representations for the current segment using said atoms contained in said dictionary.
12. The system of claim 11 wherein said cost to reconstruct is smaller when fewer atoms are needed to reconstruct said group of feature representations for said segment.
13. The system of claim 1 wherein said function of updating said dictionary minimizes the difference between a vector representing an event and a vector reconstructed using the dictionary, minimizes the number of dictionary items and requires that all feature vectors in nearby video segments use similar dictionary items.