Interactive visualization-based surveillance video synopsis

摘要

Video synopsis is an effective technique for the efficient analysis of long videos in a short time. To generate a compact video, multiple tracks of moving objects, which we call as tubes are displayed simultaneously by rearranging them along the time axis. Contemporaneous video synopsis approaches focus on collision avoidance, or preservation of chronological order among tubes. However, generation of an adaptive personalized user-oriented synopsis video congruent to users’ preferences has yet not been thoroughly experimented. This paper propounds a framework for personalized visualization of synopsis video, integrating pertinent object attributes such as color, type, size, speed, travel path and direction towards generation of synopsis video for precise inference of user needs. The framework motivates users to interactively define queries for creation of the targeted synopsis. User queries are classified into visual-queries, temporal-queries, spatial-queries, and spatio-temporal queries concomitant with the visual and spatio-temporal attributes. Tubes relevant to a user-query are selected, and grouped according to original behavioral interactions followed by their rearrangement, to generate synopsis video with fewer false collisions. To evaluate the proffered technique, two evaluation metrics are proposed and extensive experiments of publicly available surveillance videos are conducted. The experimental results demonstrate the propriety and usability of the newer approach.