Claims:I/WE CLAIM:

1. One or more non-transitory computer-readable media storing instructions executed by one or more processors to perform a sequence comprising:
detecting shot boundaries between a first and second segment in video and between the second and a third segment in video, wherein a segment includes a plurality of frames;
storing feature vectors based on characteristics of frames in the second segment;
using an object detection technique based on features of the feature vectors to identify one or more objects in frames of the second segment to develop a frame score; and
selecting a representative frame for the second segment based on at least one frame score.

2. The media of claim 1 further comprising scoring each frame on a first criterion based on how well the frame represents other frames in a segment.

3. The media of claim 2 further including scoring each frame based on a second criterion different from the first criterion.

4. The media of claim 1 further including using a histograms of oriented gradient object detection technique.

5. The media of claim 3 wherein the second criterion is different from how well the frame represents other frames.
, Description:CLAIM FOR PRIORITY

This application claims priority to U.S. Patent Application Serial No. 14/477,595, filed on 04 September, 2014, titled "REAL TIME VIDEO SUMMARIZATION", and which is incorporated by reference in entirety.

BACKGROUND

A digital camera is a component often included in commercial electronic media device platforms. Digital cameras are now available in wearable form factors (e.g., video capture earpieces, video capture headsets, video capture eyeglasses, etc.), as well as embedded within smartphones, tablet computers, and notebook computers, etc.

The introduction of streaming video from mobile digital cameras has ushered in an era having an unprecedented volume of video data. Consider an application where the user wears a pair of glasses fitted with a video camera. The camera captures video streams depicting the activities of the user throughout the day. The user may be interested in viewing a synopsis of the main events of a particular day. However, manual analysis of such large amounts of data is intractable and automated data processing techniques have not kept pace with the need.

For example, automated video summarization algorithms that attempt to abstract the main occurrences and provide a succinct representation of a captured video typically require access to an entire pre-recorded video file and generate a summary of this static atomic unit. Such a technique however relies on large processing/storage resources to first encode, archive, and decode the entire stream duration originally captured at potentially high frame rates (e.g., 25 frames/second, or more). If such processing/storage resources are not present natively within the video recording platform, which may become the default for ultra-low power platforms (e.g., headsets, and other wearable forms), an entire video stream needs to be transferred from a camera platform to a backend machine (e.g., cloud server) for processing. Such a transfer however entails a heavy load on communication channels and is impractical for many device platforms and use cases.

Video recording platform architectures and techniques capable of automatically summarizing video in real-time (on-the-fly) as a video stream arrives progressively over time from camera module hardware present on the platform are therefore advantageous.