Lost World
John See, Suyin Tan · 2014
Video surveillance over a long span of time is no longer a luxury in this day and age. The abundance of video data captured over time presents a slew of new problems in computer vision. One potential challenge involves the task of finding anomalous tracks over a long period of time. In this work, we propose a new time-scale framework for mining anomalous track patterns in long-term surveillance videos. Track clustering is performed at two separate temporal levels to better represent the common modes of behaviour. A probabilistic anomaly prediction algorithm is also introduced to evaluate the abnormality of new tracks. In our preliminary work, experiments conducted on the LOST dataset offer insights into how track anomalies can be mined and classified. We hope this work will provide the impetus for further advancements in this direction.