부산대학교 도서관

Recognizing collective human activities has gained attention. Collective activities are such as queueing in a line, talking together and waiting by an intersection. It is often hard to differentiate between these activities only by the appearance of the individual. Hence, recent works exploit the contextual information of other people nearby. However, these works do not take enough care of the spacial and temporal consistency in a group (e.g. considering the consistency in only adjacent area). To solve the problem, this paper describes a method to integrate individual recognition result via fully connected CRFs, which assume the relationships among all the people. Unlike previous methods that determine the range of human relations by heuristics, our method describes the “multi-scale” relationships in position, size, movement and time sequence as flexible potentials, so as to handle various types, sizes and shapes of groups. Experimental results show that our method outperforms state-of-the art methods.