Automatic recognition of unpredictable events in videos

The presented approach allows us to recognize frames in video sequences that are significantly different from the previous and/or following frames. In this way we are able to detect unpredictable events in videos. We map a video sequence to a polygonal trajectory by mapping each frame to a feature vector and joining the vectors representing consecutive frames by line segments. Shape analysis of the obtained polygonal curve allows us to detect frames representing unpredictable events. We demonstrate the performance of our approach on surveillance videos.