Recognizing Interactions Between People from Video Sequences

摘要

This research study proposes a new approach to group activity recognition which is fully automatic. The approach adopted is hierarchical, starting with tracking and modelling local movement leading to the segmentation of moving regions. Interactions between moving regions are modelled using Kullback-Leibler (KL) divergence. Then the statistics of such movement interactions or as relative positions of moving regions is represented using kernel density estimation (KDE). The dynamics of such movement interactions and relative locations is modelled as well in a development of the approach. Eventually, the KDE representations are subsampled and considered as inputs of a support vector machines (SVM) classifier. The proposed approach does not require any intervention by an operator.