关键词: 光流, 冗余信息, 深度学习, 动作预测

Abstract: In recent years, the deep learning-based action prediction method which uses optical flow as input features has gradually become the mainstream. Due to the environmental factors, the optical flow can easily introduce irrele-vant redundant information to reduce the accuracy of action prediction. However, the existing methods do not take the redundant information in optical flow into consideration. In this paper, the redundant information in the optical flow is removed from three aspects: the redundant information is brought by the static part in the video; the reasonable motion area is selected to eliminate the redundant information introduced by irrelevant background factors; the motion of the camera is evaluated to remove the redundant information generated by the camera motion. Aiming at the optical flow after removing redundant information, an action prediction framework based on deep learning is proposed. The parameters of the model are reduced by using space convolution and time convolution. The voting mechanism based on time weight is used to predict the action. Experiments on UT-Interaction set1 and set2 demonstrate the effectiveness of this method.

Key words: optical flow, redundant information, deep learning, action prediction