基于深度神经网络和投影树的高效率动作识别算法

为了提高视频中动作识别的准确率和速度,提出一种基于深度神经网络和投影树的高效率动作识别算法。采用三维Harris角点检测时空域中发生显著变化的局部结构,划分动作识别的主要区域和次要区域;设计两种Siamese神经网络以及相应的损失函数,考虑连续帧间的局部一致性,学习视频的主要区域特征;为兴趣点的特征建立投影树,提高查询的匹配速度。基于公开数据集的仿真实验结果表明,该算法实现了较好的无监督学习效果,并且具有较高的效率。

HIGH EFFICIENT ACTION RECOGNITION ALGORITHM BASED ON DEEP NEURAL NETWORK AND PROJECTION TREE

In order to improve the accuracy and speed of action recognition in videos,this paper proposes a high efficient action recognition algorithm based on deep neural network and projection tree.We adopted three dimensional Harris corner to detect the significant local structure variations in both space domain and time domain,and distinguished the major regions and minor regions for action recognition.Then,two types of Siamese neural networks and two respective loss functions were designed.We considered the local consistency between consecutive frames and learned the main regional features of the video.We constructed a projection tree for features of interested points to improve the speed of query matching.Simulation experimental results based on the public datasets indicate that the proposed algorithm realizes a good unsupervised learning performance,and it has a higher efficiency.