Autonomous Learning Systems: From Data Streams to Knowledge in Real-time

13

Autonomous Novelty Detection and Object Tracking in Video Streams

13.1 Problem Definition

In machine learning the problem of autonomous novelty detection is not new and has been studied in relation to fault detection and video analytics extensively. It aims to identify a new or previously unknown data item in the data stream or object in a video. It plays a pivotal role in a range of applications such as computer vision and robotics, surveillance and security, machine health monitoring, medical imaging, human–computer interaction, and so on.

In this chapter this problem will be considered from the point of view of its application to autonomous video-analytics. Same principles are behind the pioneering results on autonomous real-time anomaly detection and flight data analysis (FDA) in aviation which resulted in an EU project SVETLANA (http://www.svetlanaproject.eu/). This has an increasing importance nowadays when there is a huge and growing amount of video streams produced (for example, in UK there are over 4 million CCTV cameras installed, which makes one camera for every 14 people in the country (Daily Mail, 2009)) that require real-time analysis. The main aim of such an autonomous system (Hampapur, 2005) would be to detect, identify/classify, and track anomalous movements behaviour and activities in the area being observed/monitored.

The data stream can take the form of a video (image frames) from a digital camera, electro-optical (EO), infrared (IR) or other source, for ...