摘要
为解决基于限制玻尔兹曼机的时空深度置信网络的人体行为识别算法训练过程需要大量训练数据,在小样本训练前提下识别率低的问题,提出采用滑动窗口技术增加训练数据的方法。在视频帧序列上利用部分重叠的滑动窗口进行视频块截取,获得比将视频直接分块更多数量的视频块,在较小的视频数据中获取更大的训练数据用于神经网络的训练。实验结果表明,在测试视频较少的情况下,使用滑动窗口的时空深度置信网络识别率显著高于原始算法。
Human action recognition,which adopts the spatiotemporal deep belief networks（ST-DBN）based on restricted Boltzmann machines,needs a significant amount of training data.It has low recognition rate with a small training set.In view of this,a solution,which used sliding window to enlarge the training set,was proposed.Video clips were intercepted from the full video frames through partially overlapped sliding windows.From this step,more video clips were obtained than using the usual way which divided a video directly.The bigger training set was obtained.Experimental results show that the sliding window is superior to current original algorithm when testing video data set is small.
作者
高大鹏
朱建刚
GAO Da-peng, ZHU Jian-gang(College of Computer, Civil Aviation Flight University of China, Guanghan 618307, Chin)
出处
《计算机工程与设计》
北大核心
2018年第8期2654-2659,共6页
Computer Engineering and Design
基金
民航局科技基金项目(MHRDZ201004)
国家科技支撑计划基金项目(2011BAH24B06)
国家自然科学基金项目(60879022)
中国民航飞行学院科研基金面上基金项目(J2012-40)
关键词
行为识别
限制玻尔兹曼机
小样本训练集
滑动窗口
时空深度置信网络
action recognition
restricted Boltzmann machines
small training set
sliding window
spatiotemporal deep belief networks
