Automatic detection of salient objects for a video database system

Abstract

Recently, the increase in the amount of multimedia data has unleashed the development of storage techniques. Multimedia databases is one of the most popular of these techniques because of its scalability and ability to be queried by the media features. One downside of these databases is the necessity for processing of the media for feature extraction prior to storage and querying. Ever growing pile of media makes this processing harder to be completed manually. This is the case with BilVideo Video Database System, as well. Improvements on computer vision techniques for object detection and tracking have made automation of this tedious manual task possible. In this thesis, we propose a tool for the automatic detection of objects of interest and deriving spatio-temporal relations between them in video frames. The proposed framework covers the scalable architecture for video processing and the stages for cut detection, object detection and tracking. We use color histograms for cut detection. Based on detected shots, the system detects salient objects in the scene, by making use of color regions and camera focus estimation. Then, the detected objects are tracked based on their location, shape and estimated speed.