一种基于双流卷积神经网络跌倒识别方法

发布时间：2018-02-03 07:28

本文关键词： 跌倒识别双流卷积神经网络视频帧光流图　出处：《河南师范大学学报(自然科学版)》2017年03期 　论文类型：期刊论文

【摘要】：针对跌倒行为的视觉特征难以提取的问题,提出一种由两路卷积神经网络和模型融合部分组成的双流卷积神经网络(Two-Stream CNN)的跌倒识别方法.该方法的一路对视频帧的运动人加框标记后,送三维卷积神经网络(3D-CNN)处理来消除视频背景的干扰;另一路从相邻视频帧获取光流图后,送VGGNet-16卷积神经网络处理;最后将3D-CNN和VGGNet-16的Softmax输出识别概率加权融合作为Two-Stream CNN输出结果.实验结果表明:标记运动人并经3D-CNN处理有效地消除了视频背景的干扰;Two-Stream CNN跌倒识别率为96%,比3D-CNN提高了4%,比VGGNet-16网络提高了3%.
[Abstract]:It is difficult to extract the visual features of fall behavior. A two-flow convolutional neural network, Two-Stream CNN, is proposed, which consists of two-channel convolution neural network and model fusion. This method marks the moving person of the video frame along the way after marking. The 3D convolution neural network 3D-CNN is sent to eliminate the interference of video background. On the other hand, the optical flow graph is obtained from the adjacent video frames, and the VGGNet-16 convolution neural network is sent to process it. Finally, the weighted fusion of Softmax output recognition probability of 3D-CNN and VGGNet-16 is used as the output result of Two-Stream CNN. The experimental results show that:. The moving person is labeled and processed by 3D-CNN, which effectively eliminates the interference of video background. The fall recognition rate of Two-Stream CNN is 96, 4% higher than that of 3D-CNN and 3% higher than that of VGGNet-16 network.
【作者单位】：华东交通大学信息工程学院;
【基金】：江西省自然科学基金(20142BAB207001)
【分类号】：TP391.4
【正文快照】： 中国已经步入老龄化社会,据统计每年大约有4000万老人发生跌倒,跌倒后没有即时获得援助会加重老人受伤的程度,严重的情况下甚至可能会导致死亡.因此,及时、准确的判断对老人意外跌倒行为具有重要的研究意义.随着图像处理技术的快速发展,基于计算机视觉的跌倒识别受到广泛的关

【参考文献】