Latent Regression Forest Based Human Pose Estimation And Gesture Recognition Application

Human-Computer Interaction(HCI)has always been a very popular research issue.With the development of intelligent wearable devices,the traditional external devices,such as mouse,keyboard and touch screen cannot meet the requirements of natural HCI,which replaced by people-centered interactive technology,such as speech recognition,gesture recognition and human pose estimate and so on.Among them,the vision-based interaction technology mainly includes human pose estimation and gesture recognition.Human pose estimation is the process in which the configuration of body parts,such as joint position,posture and movement,is estimated from sensor input;meanwhile,gesture recognition is judging the meaning of hand pose.With the development of hardware,it is possible to collect images containing depth information,which provides richer information for human behavior analysis.In this paper,we only research on depth map.Considering human body and hand are objects with complex joints and fixed structure,we introduces an algorithm called Latent Regression Forest,which is a combination of Random Forest and Latent Tree Model.The latent tree,which is generated by Chow-Liu Recursive Grouping algorithm in an unsupervised way,contains the geometric structure of the human body.And we use two metric: Euclidean distance and Geodesic distance to generate the latent tree to guiding the construction of regression forest.To evaluate the algorithm,we do experiments on public dataset and compared to the state-of-the-art.Results show that the proposed method can effectively estimate the human pose,especially in self-occlusions situation.In addition,to verify the performance of the Latent Regression Forest in practical,we also apply it to hand pose estimation,and construct a real-time gesture recognition application.