Hand Pose Estimation Via A Classification-guided Regression Method

With the popularization and development of smart home and smart devices,infor-mation interaction between human and equipment will become more and more frequent in daily life.Especially with the development of computer and artificial intelligence,the research field about desirable and non-contacting human computer interaction is getting more dynamic and sizeable.The above research fields include Eye-tracking,Speech Recognition,Facial Expression Recognition,Lip-reading Recognition,Face Recogni-tion,Hand Gesture Recognition and Human Pose Recognition,etc.Due to hand gesture has abundant information and is natural,comfortable and unrestricted,hand gesture in-teraction technology will become a hot and important research area in future.Since small volume of hand,fast motions and rapid changes in direction,degrees of freedom,strong similarity in the appearance of fingers and self-occlusions,efficient and accurate vision-based hand pose estimation methods are a challenging research subject.For the complex high dimension of hand pose space,large viewpoint variation and se-rious self-occlusions,this paper presents an "divide and conquer" based classification-guided regression algorithm which can effectively estimate hand pose joints' 3D lo-cations.To avoid a single model can't deal with the issue of all cases,the proposed method splits the complex whole task of regression to several relatively easy subtasks,each corresponding regressor is only trained to handle its exclusive subtask.At first,a classifier-GoogLeNet,which is a convolution neural network model and its input is a single depth image,is obtained by offline training method.Different from previous classifiers are trained according to viewpoint variation,our classifier is trained based on rigid aligned hand pose.For different hand gesture types,cascaded random forest regressors are offline trained on disjoint part of the whole training dataset.In the testing phase,based on the predicted class of classifier,a corresponding regressor is selected to estimate the final hand pose.Dense experiments demonstrate that the proposed classification-guided regression method is efficient and effective.From quantitative point of view,the proposed method is significantly superior to the entire regression algorithm without classification.Com-paring against the state-of-the-art works,the proposed method achieves superior accu-racy predictions on most error threshold interval.From qualitative point of view,the proposed method not only handles complex large viewpoint variation and serious self-occlusions,and also maintains a high frame rate.Therefore,the proposed classification-guided regression method completely satisfies the real-time accurate hand pose estima-tion application scenarios.