For the diversity of group behavior characteristics in complex scenes and the problem of difficult interaction modeling

this paper proposes a new two-layered network architecture. The first layer of network combines a pseudo 3D residual network with a graph convolution network to capture the interaction characteristics. The second layer of network

uses the pseudo 3D residual network to capture the group global scene spatio-temporal characteristics. Based on the complementary role of the above features

their group behavior decisions are fused with a weight adaptive adjustment algorithm

which adaptively calculates importance weights for the group behavior categories predicted by the above two channels

and realizes decision fusion of the different prediction results. The method has achieved 91.4% and 97.9% average recognition accuracy on CAD and CAE respectively.