Feature Retrieving for Human Action Recognition by Mixed Scale Deep Feature Combined with Attention Model
Xiaolei Zhao, Yang Yi, Zemin Qiu, Qingqing Zeng · 2020 5th International Conference on Computer and Communication Systems (ICCCS) · 2020
In order to capture actions with different speed and range motions, a feature retrieving method with mixed scale deep feature and visual attention model is proposed. First, space and time network of a dual-stream deep neural network are taken to receive original video frames and optical flow images with different sampling steps respectively. Secondly, attention model is introduced into long-short-term memory networks to screen out visual attention areas in classification stage. Finally, the results of the space and time networks are fused together. Experimental results on datasets HMDB51 and UCF101 show that the proposed approach achieves higher recognition accuracy.