Change Detection for Video

The increased presence of digital multimedia in numerous applications, such as security, surveillance, the semantic web, has rendered the automated characterization of video content necessary. The localization of different activities/events in video content is of particular interest, however it is quite challenging to achieve in a principled manner and with no prior knowledge or training. This work presents an original, principled solution to the problem of detecting changes of activity in video, based on sequential change detection techniques. Initially, a binary mask of the active pixels, the Activity Area, is extracted in a pre-processing step by estimating the kurtosis values of inter-frame illumination variations. Sequential change detection is then applied to the illumination changes of the active pixels over time, leading to the separation of the video sequence into segments corresponding to different activities, which can be further processed for classification and recognition. Experiments with various indoors and outdoors videos demonstrate the system’s good performance.