Research Of Human Head Pose Estimation Based On Binocular Vision

Transcranial magnetic stimulation is widely used in epilepsy,mental science,rehabilitation,pediatrics and other fields nowadays.While treating with transcranial magnetic stimulation,head of patients needed to be aimed at therapeutic equipment precisely which may bring discomfort feelings to patient.In this situation,medical staff has to make sure position of the head of patients in the right place all the time.And it not only cost more money and time,but also decreases accuracy.So a system tracking head position rapidly and efficiently is required to solve this situation.This paper presents a vision module of transcranial magnetic stimulation robot system designed with binocular vision measurement.The camera calibration is a key step in the spatial measurement of binocular vision,this paper first introduces the pinhole camera model,camera calibration and distortion correction is completed by the Zhang's calibration method.Corrected images by the epipolar geometry based on the binocular vision system.In order to overcome the difficulty of finding corresponding points in low texture area.A new stereo matching algorithm is proposed based on the research of the current research situations of binocular stereo matching,put forward a new stereo matching algorithm.The algorithm uses epipolar distance transform combined with cross-scale cost aggregation,this method solve the problem of fixed search window in epipolar distance transform by multiscale ideas,and improve the stereo matching points in face regions.To obtain the face image,This paper uses Adaboost algorithm to detection the face regions,and uses the active shape model algorithm to extract the feature points of the face region.This algorithm effectively decreases the cost of computing,and estimate the initial pose of human head through the 3D information of facial feature points.In the human head pose estimation section,a dense face disparity map obtained by stereo matching algorithm is used to create a face point cloud to estimate the human head pose.Then the face point cloud obtained from the initial posture of the patient is taken as template,and then the head position is estimated by the improved iterative closest point algorithm.Finally,an error analysis experimental platform is built through robot arm and head model to evaluate pose accuracy.Experiments show that our algorithm basically meets the needs of transcranial magnetic stimulation.