Collective activity localization by spatiality preservation search

摘要

In this paper, we propose a collective activity localization method based on the spatial contextual information among people. In contrast to previous works of collective activity recognition, the collective activity localization method recognizes groups of people participating in a consistent activity as well as their activity labels. The biggest challenge of collective activity localization is efficient computation, due to exponential growth of the number of possible groups. The combinatorial search methods cannot be naively applied for collective activity localization, because the participant's position dramatically changes by adding new participants to the groups. In this work, we introduce a novel search algorithm named the Spatiality Preservation Search (SPS), which provides exact solutions in polynomial time. By determining participants on the group boundaries first, the SPS enumerates groups without changing the participant's position in the groups. Also, using Haar-like contextual descriptors, the SPS extracts groups efficiently using integral images. Experiments on an existing data-set and challenging new data-sets show efficiency of the proposed method.