融合宽残差和长短时记忆网络的动态手势识别研究

针对现有的动态手势识别方法对长时间序列的时空特征难以精确匹配的问题，提出了一种基于宽残差和双向长短时记忆网络的时空特征一致手势识别方法。首先使用已经训练好的3D卷积神经网络从视频的空间和时间维度同步提取出短时特征，再经双向空间长短时记忆网络同步解析后形成长时空特征连接单元，并作为残差网络的输入。为了验证算法的有效性，使用Kinect传感器构建了一个全新的多模式手势数据集，在三个手势识别公开数据集SLVM、Montalbano和SKIG上的实验表明，提出的方法有很好的性能表现，识别精度超越了目前已公开的最佳识别率。

Dynamic gesture recognition based on wide residual networks and long short-term memory networks

The current dynamic hand gesture recognition method is not able to capture long-term spatiotemporal features from image sequences accurately. In order to solve this problem, this paper proposed a new dynamic gesture recognition algorithm based on wide residual networks and long short-term memory networks that performed simultaneous detection and classification. Firstly, it extracted spatial and temporal features from the fine-tuned 3D convolutional neural networks. Next, it utilized a bidirectional convolutional long short-term memory networks to further take into account the temporal aspect of image sequences. Lastly, it sent these higher level features to the wide residual networks for final gesture recognition. In order to validate this method, if introduced a new challenging multimodal dynamic hand gesture dataset, which captured with Kinect sensors. Experimental results show that the proposed method achieves state-of -the-art performance on SLVM, Montalbano and SKIG.