Abnormal Video Event Detection

Many surveillance applications require analysis of the events taking place in video streams recorded in specific situations, in order to find suspicious or abnormal actions, which might present a threat and should be signaled to a human operator. For example, in a traffic monitoring system, it is useful to detect vehicles that make U-turns or brake suddenly, pedestrians trespassing the street, and other abnormal traffic behavior or violations.

A fundamental issue in detecting abnormal events is that abnormality cannot be specified, because anything that deviates from normality is abnormal. The clustering-based approach has been investigated in this project. This approach is based on the fact that an abnormal event is associated with the distinctness of the activity and a normal event indicates the commonality. For instance, people running is abnormal if most of the crowd is walking, and a car moving in a different way than the most other traffic is also abnormal. Obviously, what characterizes the normality is the high recurrence of some similar events. And usually there are only a few such normal patterns in a specific surveillance scenario. Therefore, unsupervised clustering can be performed on all the video events. Those events clustered into dominant (e.g., large) groups can be identified as normal. And those that cannot be explained by the dominant groups (e.g., distant from all cluster centers) are abnormal. This clustering-based approach is actually a mining process.

Currently, we are trying to find solutions of several problems in this approach, including:

  • How to characterize video events in a parameter space and then measure the dissimilarity between any two events;
  • The number of clusters is unknown, as it is not clear how many kinds of normal patterns are present in the video;
  • Criteria are needed to identify abnormal and normal event clusters from clustering results;
  • This approach should also be able to detect abnormalities in videos outside the original data set.