Real-Time Fall Detection Based on Light-weight Human Pose Estimation and Graph Convolution Network
Wei He · Computer Science and Application · 2021
基于人体姿态估计的摔倒检测方法，因其人体姿态估计模型涉及十几个关节点的识别与处理，导致整体模型的检测速度较慢。为了摔倒检测达到实时性，提出了一种基于轻量级人体姿态估计模型和图卷积的摔倒实时检测方法。该方法首先采用优化后的基于目标检测的两阶段轻量级人体姿态估计模型进行关节点检测，使整体模型达到轻量级；然后使用只有6个特征提取模块的时空图卷积网络对人体关节点序列进行摔倒检测，提高整体模型摔倒检测的准确率。本文通过NTU-D-RGB-120和UR Fall Detection Dataset两个数据集进行实验，摔倒检测的正确率达到96.1%，整体模型在GTX1060Ti显卡中达到约33FPS。 The fall detection based on human pose estimation, because the human pose estimation involves the recognition and processing of more than a dozen joint points, the detection speed of the overall model is slow. In order to achieve real-time fall detection, a real-time fall detection method based on a lightweight human pose estimation and graph convolution network is proposed. The method first uses an optimized two-stage lightweight human pose estimation based on object detection to detect joint points, so that the overall model is lightweight; then uses the spatio-temporal graph convolutional network with only 6 feature extraction modules to perform fall detection on the human joint point sequence to improve the accuracy of the overall model fall detection. This article conducts experiments on two data sets, NTU-D-RGB-120 and UR Fall Detection Dataset, and the accuracy rate of fall detection reaches 96.1%, and the overall model reaches about 33FPS in the GTX1060Ti.