Deep Learning Based Human Pose Estimation Recognition Algorithm for Elderly People Living Alone

In this paper, we propose an algorithm for such abnormal postures in the case of elderly people who cannot be detected in time when they are living alone at home. The algorithm is based on the real-time behavior recognition algorithm of lightweight OpenPose. By replacing the original VGG-19 backbone network with the MobileNetV3 architecture, and integrating the Depthwise Separable Convolution and Atrous Convolution optimization strategies to detect the skeleton of human posture, the number of model parameters and computational costs are significantly reduced at the expense of facial expression recognition and multi-person scenarios, so as to more accurately and timely recognize the human posture of the elderly living alone.