An Improved Two-Stage Multi-person Pose Estimation Model

Abstract

Generally, multi-person pose estimation plays a crucial role in behavior recognition in images and videos. Previously, pose estimation of a single person is popular and achieves high prediction accuracy with the development of deep learning. However, pose estimation of multi-person remains to be a huge challenge and cannot achieve the same effect as that of a single person. It mainly results from the rare, missing or incorrect location detection and overlap of pose, which are usually caused by incomplete person identification. Therefore, we propose an improved two-stage multi-person pose estimation model (ITMPE) to further improve the performance of multi-person pose estimation. The first stage, Mask R-CNN is used for person identification. The second stage, processed images or videos with identified people only are fed into OpenPose model for multi-person pose estimation. The comparative experiments show that our proposed model achieves a significant improvement than original model. Our proposed model reduces the MSE, MAE by around 27.38%, 21.57% and increases R2, Mean values by 49.80% and 96.91% on average, respectively. The improvement in person identification and misclassification are shown in our comparison images. More people are captured and given the pose estimation, which directly affect the performance of behavior recognition.