Abstract

Modeling interactions of multiple co-occurring objects  in a complex activity is becoming increasingly popular in  the video domain. The Dynamic Bayesian Network (DBN)  has been applied to this problem in the past due to its natural ability to statistically capture complex temporal  dependencies. However, standard DBN structure learning  algorithms are generatively learned, require manual  structure definitions, and/or are computationally complex  or restrictive. We propose a novel structure learning  solution that fuses the Granger Causality statistic, a direct  measure of temporal dependence, with the Adaboost  feature selection algorithm to automatically constrain the  temporal links of a DBN in a discriminative manner. This  approach enables us to completely define the DBN  structure prior to parameter learning, which reduces  computational complexity in addition to providing a more  descriptive structure. We refer to this modeling approach  as the Granger Constraints DBN (GCDBN). Our  experiments show how the GCDBN outperforms two of the  most relevant state-of-the-art graphical models in complex  activity classification on handball video data, surveillance  data, and synthetic data