Human pose estimation using deep learning

The objective of human pose estimation (HPE) is to predict the positional coordinates of body keypoints in images. While there has been significant progress in HPE, certain challenges persist, causing the prediction of false-positive keypoints with considerable confidence. This problem may cause many troubles in the applications where high detection accuracy is required, such as in a traffic environment. As the majority of networks solely learn and predict the positional coordinates of each human joint, they do not effectively leverage the visibility information of such keypoint in the image. In this dissertation, We propose a modified model that uses visibility information to improve the accuracy of HPE. The visibility of keypoints indicates whether it is obstructed by other objects, visible or not captured in the photograph. Based on the original model, we add a visibility module to improve the accuracy of prediction. Moreover, compared with other similar methods, our network has a smaller number of parameters without significant loss the accuracy. We use the car cabin dataset DriPE to test the performance of our networks.Master of Science (Communications Engineering