学术论文

      基于光流约束自编码器的动作识别

      Action recognition based on optical flow constrained auto-encoder

      摘要:
      为了改进特征学习在提取目标运动方向及运动幅度等方面的能力,提高动作识别精度,提出一种基于光流约束自编码器的动作特征学习算法.该算法是一种基于单层正则化自编码器的无监督特征学习算法,使用神经网络重构视频像素并将对应的运动光流作为正则化项.该神经网络在学习动作外观信息的同时能够编码物体的运动信息,生成联合编码动作特征.在多个标准动作数据集上的实验结果表明,光流约束自编码器能有效提取目标的运动部分,增加动作特征的判别能力,在相同的动作识别框架下该算法超越了经典的单层动作特征学习算法.
      Abstract:
      To improve the capability of feature learning in extracting motion information such as amplitudes and directions and to increase the recognition accuracy, an optical flow constrained auto-encoder is proposed to learn action features.The optical flow constrained auto-encoder is an unsupervised feature learning algorithm based on single layer regularized auto-encoder.The algorithm uses the neural network to reconstruct the video pixels and use the corresponding optical flows in video blocks as a revised regularization.The neural network learns the appearances of the action and encodes the motion information simultaneously.The associated codes are used as the final action features.The experimental results on several well-known benchmark datasets show that the optical flow constrained auto-encoder can detect the motion parts efficiently.On the same recognition framework, the proposed algorithm outperforms the state-of-the-art single layer action feature learning algorithms.
      作者: 李亚玮 [1] 金立左 [1] 孙长银 [1] 崔桐 [2]
      Author: Li Yawei [1] Jin Lizuo [1] Sun Changyin [1] Cui Tong [2]
      作者单位: 东南大学自动化学院,南京,210096 中国电科集团28所,南京,210007
      年,卷(期): 2017, 47(4)
      分类号: TP181
      在线出版日期: 2017年8月15日
      基金项目: 国家自然科学基金资助项目