Human Pose Recognition Using Chamfer Distance in Reduced Background Edge for Human-Robot Interaction

Human pose estimation and recognition have recently attracted a lot of attention in the field of human-computer interface (HCI) and human-robot interface (HRI). This paper proposes human pose recognition method using a chamfer distance that computes similarities between an input image and pose templates stored in database. However, the chamfer distance has a disadvantage that it may produce false-positive in regions where similar structures in edge images as templates exist, even when no human pose is present. To tackle this problem, the proposed method tries to adaptively attenuate the edges in the background while preserving the edges across foreground/background boundaries and inside the foreground. The proposed algorithm builds on a key observation that edge information in the background is static when a human takes pose as the interface. Moreover, the algorithm additionally considers edge orientation to minimize loss of foreground edges, caused by edge attenuation. In the experiments, the proposed method is applied to the HRI. Edge information for the background is modeled when the robot stops in front of the human for interaction with gesture. The performance of the proposed method, time cost and accuracy, was better than the chamfer distance and pictorial structure method that estimates human pose.