针对长短时记忆网络(LSTM)不能有效地提取动作前后之间相互关联的信息导致行为识别率偏低的问题,提出了一种基于Bi-LSTM-Attention模型的人体行为识别算法。该算法首先从每个视频中提取20帧图像,通过Inceptionv3模型提取图像中的深层特征,然后构建向前和向后的Bi-LSTM神经网络学习特征向量中的时序信息,接着利用注意力机制自适应地感知对识别结果有较大影响的网络权重,使模型能够根据行为的前后关系实现更精确的识别,最后通过一层全连接层连接Softmax分类器并对视频进行分类。通过Action Youtobe和KTH人体行为数据集与现有的方法进行比较,实验结果表明,本文算法有效地提高了行为识别率。

This study proposed a human action recognition algorithm based on the Bi-LSTM-Attention model to solve the problem of low action recognition rate. This problem was caused by the inability of long short term memory (LSTM) networks to effectively extract correlative informations before and after actions. The proposed algorithm first extracted 20 image frames from each video and used the Inceptionv3 model to extract deep features from these frames. Then, forward and backward Bi-LSTM neural networks were constructed to learn the temporal information in the feature vectors. The influences of network weights on recognition results were adaptively perceived using the attention mechanism. This step was performed so that the model could achieve more accurate recognition based on the relationship between informations acquired before and after performing the given action. Finally, the videos were connected via a fully-connected layer to a Softmax classifier for classification. Comparison between the Action Youtobe and KTH human action datasets and existing methods revealed that the proposed algorithm effectively improved the action recognition rate.