Abstract: Due to the diversity of video content and the complexity of video background,how to effectively extract spatio-temporal features is the main challenge of the video action recognition.In order to use deep networks to learn spatio-temporal features,researchers usually use two-stream networks and 3D convolution networks.Two-stream networks use the optical flow as its input to learn temporal features,but optical flow cannot express long-distance temporal relationships and the calculation of optical flow requires a lot of memory and time.On the other hand,3D convolution networks increase the computational cost by an order of magnitude compared with 2D convolution networks,which easily leads to over-fitting and slow convergence.To solve these problems,an attention-based multi-dimensional feature activation residual networks(MFARs) is proposed for video action recognition.A motion supplement excitation module is proposed to model temporal information and stimulate motion information.A united information excitation module is proposed to use temporal features to stimulate channels and spatial information in order to learn a better spatio-temporal features.Combing these two modules,MFARs is constructed for video action recognition.The proposed method obtainsan accuracy of 96.5% and 73.6% respectively on datasets UCF101 and HMDB51.Compared with the current mainstream action recognition models,the proposed multi-dimensional feature excitation method can effectively express spatial and temporal characteristics,and achieve a better balance of computation complexity and classification accuracy.

Key words: Action recognition, Deep learning, 2D convolution network, Attention mechanism, Video feature representation