Handbook of Video Databases: Design and Applications (Internet and Communications)

2. Contribution

There are few intelligent video exploration tools based on user profiles. The majority of state-of-the-art works concerned web applications. More particularly, they concerned user profiles, adaptive web sites [14], web log mining, OLAP [15], [23], [19], intelligent agents that detects user web topics [5], [8], [20], extraction of the most interesting pages [21], study of web of performance of various caching strategies [2] and continuous Markov models to influence caching priorities between primary, secondary and tertiary storages [18].

The particularity of our approach consists of:

• Applying probabilistic exploration using Markov models on video bases. Each state of the Markov model corresponds to a scene displayed by the user. In addition, each transition corresponds to the probability to explore a scene after another one.

• Avoiding the problem of Markov model high-dimensionality and sparsity. The Markov model matrix is typically very large (ℵ*ℵ for ℵ scenes). Moreover, the Markov matrix is usually very sparse for high-dimensional matrices. To avoid such problems, the approach consists of clustering video scenes, based on their content, before applying the Markov analysis. Therefore, we propose the grouping of similar states into clusters. The transition matrix lines and columns will concern the gravity centers of clusters. Newest dimension = log ℵ. ℵ is the number of scenes before clustering. The solution is clearly scalable for very large video bases.

To our knowledge, there are no solutions based on video exploration prediction. We believe that applying probabilistic approaches such as Markov models to video exploration inaugurates a new form of video exploitation. Furthermore, when we consider web applications, which are different than video explorations, not all the state-of-the-art works extract the sequence of hyper-link based on probabilistic information.