Real-time Scene Identification Using Run-length Encoding of Video Feature Sequences
Akio Nagasaka and Takafumi Miyatake
Central Research laboratory, Hitachi, Ltd. 1-280 Higashi-Koigakubo, Kokubunnji-shi, Tokyo 185-8601, Japan
We propose real-time video scene identification that detects all same scenes in stored videos as the latest freelength scene, compressing the video feature sequence in an average of less than 20 bytes per second to store features for a long time. It takes less than 30ms on the average for a typical personal computer to process 1 newly input fame image even storing more than 24-hour video features. Experiments with TV showed that this method finds correct pairs of the same scenes in real time without error. This becomes the basis for active video recording based on a user’s TV viewing history and for new robotic machine vision for surveillance.
Copyright© 1999 by Fuji Technology Press Ltd. and Japan Society of Mechanical Engineers. All right reserved.