Occulusion Aware 2D Human Pose Estimation Using Part Affinity Fields and Face Recognition

Video analytics provides us with useful information which will be helpful to in smart cities and connected community for crowd counting, event classification, traffic counting, etc. With the increasing concern over public safety, human activity monitoring in densely populated areas like the airport and shopping mall etc. has become a challenging problem with lots of different ways of solving them. One such application which can provide us with beneficial information is human pose estimation. The human pose estimation algorithm can provide us with useful information like the person's body keypoint location and limbs interconnection information which can be used to monitor the activity of a person, determine action performed at that particular point, etc. When subject to estimate pose of a specific person of interest in a crowded environment, we encounter a severe problem of occlusion which often causes the algorithm to lose the person of interest in the consecutive frame leading to either incomplete activity monitoring of the person of interest or incorrect results due to occlusion. In this thesis, three different deep learning algorithms, human face detection, human face recognition, and human pose estimation is utilized to provide a solution to address the problem of occlusion in 2d image human pose estimation, using the persons face as an identification tag.