Head Pose Estimation of Stroke Patients Based on Depth Residual Network

Abstract

The accuracy of the traditional head pose estimation method based on key feature points is easily affected by the accuracy of key feature points, serious occlusion or excessive angle deviation, resulting in bad deviation of the detection results. In order to improve the accuracy and stability of head pose estimation, a head pose estimation method using depth residual network ResNet101 as backbone network is proposed. The method AdaBound optimizer to optimize the training process gradient, use Softmax classifier and calculate the cross entropy loss function, and finally accurately predicts the head pose. We collected videos of stroke patients doing rehabilitation training, and established a new head posture data set after processing, which contains thousands of head posture RGB images of 40 stroke patients. We use the method proposed in this paper on this data set and the public dataset BIWI, and the results show that this method is very suitable for our dataset, and has good stability to different angles of the head posture, and has good robustness.