Human Action Recognition Based on Attention Mechanism and HRNet

Abstract A human action recognition network (AE-HRNet) based on high-resolution (HRNet) and attention mechanism is proposed for the problem that semantic location information of features are not sufficiently extracted by convolutional networks. Firstly, channel (ECA) module spatial (ESA) introduced; this basis, new base (EABasic) bottleneck (EANeck) modules constructed to reduce computational complexity while obtaining more accurate feature map. Experimental results MPII COCO validation sets in same environment configuration show AE-HRNet reduces improves accuracy compared network.