Local group relationship analysis for group activity recognition
Dong-Gyu Lee, Pil-Soo Kim, Seong–Whan Lee · 2017
In this paper, we present an approach that exploits local group relationship to tackle the human group activity recognition problem. Specifically, rather than analyze every human motion, we first grouping individual human object into local groups to represent the relationship in the overall scene. The important movement information is maximized by modeling both each human motion and local group relationships. The gated recurrent unit model has been adopted to handle an arbitrary length of trajectory information with non-linear hidden units. In our experiment on public human group activity dataset, we compared the performance of proposed method with that of other competing methods and showed that the proposed method outperforms others.