TT+GT at TRECVID 2010 Workshop.
Nakamasa Inoue, Toshiya Wada, Yusuke Kamishima, Koichi Shinoda, Ilseo Kim, Byungki Byun, Chin‐Hui Lee · 2010
We combine local features and global features and use them for detecting events. Local features are based on a person detector and they represent movements of individuals. Global features are based on optical flow features in video frames and they represent the flow of people. We tried to detect PersonRuns, PeopleMeet and PeopleSplitUp events. The results of three events are as follows; ANDCR of PersonRuns was 1.002, ANDCR of PeopleMeet was 1.003, and ANDCR of PeopleSplitUp was 1.008.