摘要
针对长短时记忆网络(LSTM)不能充分提取视频前后关联信息导致识别精度偏低的问题,提出一种基于时空双流融合网络与Attention算法。以经典双流神经网络分别提取融合时空特征向量;构建Bi-LSTM提取时序特征;利用注意力(Attention)机制自适应地对相关性大的特征向量分配较大的权重;采用Softmax分类器对视频进行分类,实现人体行为识别。在数据集KTH上的实验结果非常出色,识别准确率可达98.4%。
Aiming at the problem that the LSTM cannot fully extract the relevant information before and after the video,resulting in low recognition accuracy,we propose a spatio-temporal two-stream fusion network and attention algorithm.The fusion spatio-temporal feature vector was extracted by classical dual-flow neural network;we constructed forward and backward Bi-LSTM networks to extract time series features;the attention mechanism was used to adaptively assign a large weight to the feature vectors with large correlation;we used the Softmax classifier to classify the video,so as to achieve human action recognition.The experimental results on the data set KTH are excellent,and the recognition accuracy is up to 98.4%.
作者
王毅
马翠红
毛志强
Wang Yi;Ma Cuihong;Mao Zhiqiang(College of Electrical Engineering,North China University of Science and Technology,Tangshan 063210,Hebei,China)
出处
《计算机应用与软件》
北大核心
2020年第8期156-159,193,共5页
Computer Applications and Software
基金
国家自然科学基金项目(61171058)。
关键词
双流卷积网络
时空特征
Bi-LSTM
注意力机制
Two-stream convolution network
Spatio-temporal features
Bi-LSTM
Attention mechanism
