一种基于视频中人体姿态的跌倒检测方法

A method of fall detection based on human posture in video

为了及时、准确地对老年人跌倒行为进行检测,保障老年人的养老安全,提出一种基于人体姿态的跌倒检测方法。首先将视频图像送入到OpenPose算法中获取图像中人体的姿态信息,再利用三维卷积神经网络提取视频中人体姿态变化的时空特征。通过对局部特征的重新组合,得到抽象的全局特征进行跌倒检测。实验结果表明,所提出的跌倒检测方法计算复杂度低,对跌倒行为的平均正确检测率为98.32%,对其他日常行为的平均误检率为2.84%,兼顾了准确性和实时性的要求。

A method of fall detection based on human posture is proposed to detect the fall behavior of the elderly timely and accurately and ensure the security of the elderly. The video image is sent to the OpenPose algorithm to obtain the pose information of human body in the image,and then the three-dimensional convolutional neural network is used to extract the spatiotemporal features of human posture changes in the video. The abstract global features are obtained for fall detection by recombining the local features. The experimental results show that the proposed fall detection method has low computational complexity,its average accuracy rate of the fall behavior is 98.32%,and its average false-detection rate for other normal behavior is 2.84%,which gives consideration to the requirements of the accuracy and real-time capability.