Online Viterbi Optimisation for Simple Event Detection in Video
Håkan Ardö, Kalle Åström, Rikard Berthilsson · Lund University Publications (Lund University) · 2007
In this paper the problem of event detection in image sequences involving situations with multiple objects present is studied. A Hidden Markov Model describing the movements of a varying number of objects as well as their entries and exits is presented. Previously similar models have been used, but in online systems the standard dynamic programming Viterbi algorithm is typically not used to find the global optimum state sequence, as it requires that all past and future observations are available. In this paper we use an extension to the Viterbi algorithm that allows it to operate on infinite time sequences and produce the optimum state sequence with only a finite delay. This makes it possible to use the Viterbi algorithm in real time applications. Restrictions on where objects might enter or exit the scene is placed in the model, which makes the detection of some simple events trivial. We also show that extending the observation model to use use several overlapping or nonoverlapping cameras is straight forward. 1.