Motion recognition of two-stream based on multi-attention
Bo He, Tianqing Zhang, Minghua Liu, Hongbo Shao · 2023
To address the problems that the traditional two-stream network has limitations in effectively fusing spatial-temporal information and the extracted temporal information hard to capture complex motion patterns, which resulting in low recognition accuracy, an improved two-stream convolutional neural network method for human motion recognition is proposed based on multi-attention mechanism. Firstly, in the temporal network, the C3D network is used to replace the original two-dimensional network to solve the problem that the temporal information cannot be extracted effectively; secondly, in the spatial network, multi-scale convolutional Transformer encoder is coded based on the contextual relative position, features are integrated adaptively at different scales under the action of the adaptive scale attention mechanism. Experimental findings conducted on the UCF101 dataset indicate the approach described in this paper performs better in terms of the number of parameters and recognition accuracy.