Attention correlated appearance and motion feature followed temporal learning for activity recognition

<span lang="EN-US">Recent advances in deep neural networks have been successfully demonstrated with fairly good accuracy for multi-class activity identification. However, existing methods have limitations in achieving complex spatial-temporal dependencies. In this work, we design two stream fusion attention (2SFA) connected to a temporal bidirectional gated recurrent unit (GRU) one-layer model and classified by prediction voting classifier (PVC) to recognize the action in a video. Particularly in the proposed deep neural network (DNN), we present 2SFA for capturing appearance information from red green blue (RGB) and motion from optical flow, where both streams are correlated by proposed fusion attention (FA) as the input of a temporal network. On the other hand, the temporal network with a bi-directional temporal layer using a GRU single layer is preferred for temporal understanding because it yields practical merits against six topologies of temporal networks in the UCF101 dataset. Meanwhile, the new proposed classifier scheme called PVC employs multiple nearest class mean (NCM) and the SoftMax function to yield multiple features outputted from temporal networks, and then votes their properties for high-performance classifications. The experiments achieve the best average accuracy of 70.8% in HMDB51 and 91.9%, the second best in UCF101 in terms of 2DConvNet for action recognition.</span>