A study on video scene filtering using queue models for personalized semantic content consumption

Abstract

A real-time content filtering system for live broadcasts in TV terminals enables a TV viewer to obtain desired video scenes from multiple channel broadcasts. In order to achieve stable and reliable scene filtering for multiple video inputs, real-time filtering requirements such as frame sampling rate per channel, number of input channels, and buffer condition should be considered to meet the limited capacity of the terminal. In this thesis, we propose analysis models of selecting those requirements by modeling a filtering system using two queues, i.e., D/D/1 and D/M/1 queues. By minimizing the queueing time of input video frames, the proposed models can maximize filtering capacity for multiple video inputs in real-time. To verify the proposed filtering system and analysis models, we perform experiments on soccer videos with the proposed model. The experiment results show that the proposed real-time scene filtering system is suitable to TV terminals for personalized semantic content consumption, and give that its analysis models allow stable filtering within limited resources.